diff --git a/.gitattributes b/.gitattributes index a6344aac8c09253b3b630fb776ae94478aa0275b..2cbf38e2f7ca073ec8ba7e19601c721c2c736ca7 100644 --- a/.gitattributes +++ b/.gitattributes @@ -33,3 +33,28 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text *.zip filter=lfs diff=lfs merge=lfs -text *.zst filter=lfs diff=lfs merge=lfs -text *tfevents* filter=lfs diff=lfs merge=lfs -text +assets/stable-inpainting/merged-leopards.png filter=lfs diff=lfs merge=lfs -text +assets/stable-samples/depth2img/d2i.gif filter=lfs diff=lfs merge=lfs -text +assets/stable-samples/depth2img/depth2img01.png filter=lfs diff=lfs merge=lfs -text +assets/stable-samples/depth2img/depth2img02.png filter=lfs diff=lfs merge=lfs -text +assets/stable-samples/depth2img/merged-0000.png filter=lfs diff=lfs merge=lfs -text +assets/stable-samples/depth2img/merged-0004.png filter=lfs diff=lfs merge=lfs -text +assets/stable-samples/depth2img/merged-0005.png filter=lfs diff=lfs merge=lfs -text +assets/stable-samples/img2img/upscaling-in.png filter=lfs diff=lfs merge=lfs -text +assets/stable-samples/img2img/upscaling-out.png filter=lfs diff=lfs merge=lfs -text +assets/stable-samples/stable-unclip/unclip-variations_noise.png filter=lfs diff=lfs merge=lfs -text +assets/stable-samples/stable-unclip/unclip-variations.png filter=lfs diff=lfs merge=lfs -text +assets/stable-samples/txt2img/768/merged-0001.png filter=lfs diff=lfs merge=lfs -text +assets/stable-samples/txt2img/768/merged-0002.png filter=lfs diff=lfs merge=lfs -text +assets/stable-samples/txt2img/768/merged-0003.png filter=lfs diff=lfs merge=lfs -text +assets/stable-samples/txt2img/768/merged-0004.png filter=lfs diff=lfs merge=lfs -text +assets/stable-samples/txt2img/768/merged-0005.png filter=lfs diff=lfs merge=lfs -text +assets/stable-samples/txt2img/768/merged-0006.png filter=lfs diff=lfs merge=lfs -text +assets/stable-samples/txt2img/merged-0001.png filter=lfs diff=lfs merge=lfs -text +assets/stable-samples/txt2img/merged-0003.png filter=lfs diff=lfs merge=lfs -text +assets/stable-samples/txt2img/merged-0005.png filter=lfs diff=lfs merge=lfs -text +assets/stable-samples/txt2img/merged-0006.png filter=lfs diff=lfs merge=lfs -text +assets/stable-samples/txt2img/merged-0007.png filter=lfs diff=lfs merge=lfs -text +assets/stable-samples/upscaling/merged-dog.png filter=lfs diff=lfs merge=lfs -text +assets/stable-samples/upscaling/sampled-bear-x4.png filter=lfs diff=lfs merge=lfs -text +assets/stable-samples/upscaling/snow-leopard-x4.png filter=lfs diff=lfs merge=lfs -text diff --git a/LICENSE b/LICENSE new file mode 100644 index 0000000000000000000000000000000000000000..58a49c99b2b9151af5e1fee0dbd20307671f47ab --- /dev/null +++ b/LICENSE @@ -0,0 +1,21 @@ +MIT License + +Copyright (c) 2022 Stability AI + +Permission is hereby granted, free of charge, to any person obtaining a copy +of this software and associated documentation files (the "Software"), to deal +in the Software without restriction, including without limitation the rights +to use, copy, modify, merge, publish, distribute, sublicense, and/or sell +copies of the Software, and to permit persons to whom the Software is +furnished to do so, subject to the following conditions: + +The above copyright notice and this permission notice shall be included in all +copies or substantial portions of the Software. + +THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, +OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE +SOFTWARE. diff --git a/LICENSE-MODEL b/LICENSE-MODEL new file mode 100644 index 0000000000000000000000000000000000000000..9684533d88e7d853a55cabf6caa2f1e4a3e6fdc4 --- /dev/null +++ b/LICENSE-MODEL @@ -0,0 +1,84 @@ +Copyright (c) 2022 Stability AI and contributors + +CreativeML Open RAIL++-M License +dated November 24, 2022 + +Section I: PREAMBLE + +Multimodal generative models are being widely adopted and used, and have the potential to transform the way artists, among other individuals, conceive and benefit from AI or ML technologies as a tool for content creation. + +Notwithstanding the current and potential benefits that these artifacts can bring to society at large, there are also concerns about potential misuses of them, either due to their technical limitations or ethical considerations. + +In short, this license strives for both the open and responsible downstream use of the accompanying model. When it comes to the open character, we took inspiration from open source permissive licenses regarding the grant of IP rights. Referring to the downstream responsible use, we added use-based restrictions not permitting the use of the Model in very specific scenarios, in order for the licensor to be able to enforce the license in case potential misuses of the Model may occur. At the same time, we strive to promote open and responsible research on generative models for art and content generation. + +Even though downstream derivative versions of the model could be released under different licensing terms, the latter will always have to include - at minimum - the same use-based restrictions as the ones in the original license (this license). We believe in the intersection between open and responsible AI development; thus, this License aims to strike a balance between both in order to enable responsible open-science in the field of AI. + +This License governs the use of the model (and its derivatives) and is informed by the model card associated with the model. + +NOW THEREFORE, You and Licensor agree as follows: + +1. Definitions + +- "License" means the terms and conditions for use, reproduction, and Distribution as defined in this document. +- "Data" means a collection of information and/or content extracted from the dataset used with the Model, including to train, pretrain, or otherwise evaluate the Model. The Data is not licensed under this License. +- "Output" means the results of operating a Model as embodied in informational content resulting therefrom. +- "Model" means any accompanying machine-learning based assemblies (including checkpoints), consisting of learnt weights, parameters (including optimizer states), corresponding to the model architecture as embodied in the Complementary Material, that have been trained or tuned, in whole or in part on the Data, using the Complementary Material. +- "Derivatives of the Model" means all modifications to the Model, works based on the Model, or any other model which is created or initialized by transfer of patterns of the weights, parameters, activations or output of the Model, to the other model, in order to cause the other model to perform similarly to the Model, including - but not limited to - distillation methods entailing the use of intermediate data representations or methods based on the generation of synthetic data by the Model for training the other model. +- "Complementary Material" means the accompanying source code and scripts used to define, run, load, benchmark or evaluate the Model, and used to prepare data for training or evaluation, if any. This includes any accompanying documentation, tutorials, examples, etc, if any. +- "Distribution" means any transmission, reproduction, publication or other sharing of the Model or Derivatives of the Model to a third party, including providing the Model as a hosted service made available by electronic or other remote means - e.g. API-based or web access. +- "Licensor" means the copyright owner or entity authorized by the copyright owner that is granting the License, including the persons or entities that may have rights in the Model and/or distributing the Model. +- "You" (or "Your") means an individual or Legal Entity exercising permissions granted by this License and/or making use of the Model for whichever purpose and in any field of use, including usage of the Model in an end-use application - e.g. chatbot, translator, image generator. +- "Third Parties" means individuals or legal entities that are not under common control with Licensor or You. +- "Contribution" means any work of authorship, including the original version of the Model and any modifications or additions to that Model or Derivatives of the Model thereof, that is intentionally submitted to Licensor for inclusion in the Model by the copyright owner or by an individual or Legal Entity authorized to submit on behalf of the copyright owner. For the purposes of this definition, "submitted" means any form of electronic, verbal, or written communication sent to the Licensor or its representatives, including but not limited to communication on electronic mailing lists, source code control systems, and issue tracking systems that are managed by, or on behalf of, the Licensor for the purpose of discussing and improving the Model, but excluding communication that is conspicuously marked or otherwise designated in writing by the copyright owner as "Not a Contribution." +- "Contributor" means Licensor and any individual or Legal Entity on behalf of whom a Contribution has been received by Licensor and subsequently incorporated within the Model. + +Section II: INTELLECTUAL PROPERTY RIGHTS + +Both copyright and patent grants apply to the Model, Derivatives of the Model and Complementary Material. The Model and Derivatives of the Model are subject to additional terms as described in Section III. + +2. Grant of Copyright License. Subject to the terms and conditions of this License, each Contributor hereby grants to You a perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable copyright license to reproduce, prepare, publicly display, publicly perform, sublicense, and distribute the Complementary Material, the Model, and Derivatives of the Model. +3. Grant of Patent License. Subject to the terms and conditions of this License and where and as applicable, each Contributor hereby grants to You a perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable (except as stated in this paragraph) patent license to make, have made, use, offer to sell, sell, import, and otherwise transfer the Model and the Complementary Material, where such license applies only to those patent claims licensable by such Contributor that are necessarily infringed by their Contribution(s) alone or by combination of their Contribution(s) with the Model to which such Contribution(s) was submitted. If You institute patent litigation against any entity (including a cross-claim or counterclaim in a lawsuit) alleging that the Model and/or Complementary Material or a Contribution incorporated within the Model and/or Complementary Material constitutes direct or contributory patent infringement, then any patent licenses granted to You under this License for the Model and/or Work shall terminate as of the date such litigation is asserted or filed. + +Section III: CONDITIONS OF USAGE, DISTRIBUTION AND REDISTRIBUTION + +4. Distribution and Redistribution. You may host for Third Party remote access purposes (e.g. software-as-a-service), reproduce and distribute copies of the Model or Derivatives of the Model thereof in any medium, with or without modifications, provided that You meet the following conditions: +Use-based restrictions as referenced in paragraph 5 MUST be included as an enforceable provision by You in any type of legal agreement (e.g. a license) governing the use and/or distribution of the Model or Derivatives of the Model, and You shall give notice to subsequent users You Distribute to, that the Model or Derivatives of the Model are subject to paragraph 5. This provision does not apply to the use of Complementary Material. +You must give any Third Party recipients of the Model or Derivatives of the Model a copy of this License; +You must cause any modified files to carry prominent notices stating that You changed the files; +You must retain all copyright, patent, trademark, and attribution notices excluding those notices that do not pertain to any part of the Model, Derivatives of the Model. +You may add Your own copyright statement to Your modifications and may provide additional or different license terms and conditions - respecting paragraph 4.a. - for use, reproduction, or Distribution of Your modifications, or for any such Derivatives of the Model as a whole, provided Your use, reproduction, and Distribution of the Model otherwise complies with the conditions stated in this License. +5. Use-based restrictions. The restrictions set forth in Attachment A are considered Use-based restrictions. Therefore You cannot use the Model and the Derivatives of the Model for the specified restricted uses. You may use the Model subject to this License, including only for lawful purposes and in accordance with the License. Use may include creating any content with, finetuning, updating, running, training, evaluating and/or reparametrizing the Model. You shall require all of Your users who use the Model or a Derivative of the Model to comply with the terms of this paragraph (paragraph 5). +6. The Output You Generate. Except as set forth herein, Licensor claims no rights in the Output You generate using the Model. You are accountable for the Output you generate and its subsequent uses. No use of the output can contravene any provision as stated in the License. + +Section IV: OTHER PROVISIONS + +7. Updates and Runtime Restrictions. To the maximum extent permitted by law, Licensor reserves the right to restrict (remotely or otherwise) usage of the Model in violation of this License. +8. Trademarks and related. Nothing in this License permits You to make use of Licensors’ trademarks, trade names, logos or to otherwise suggest endorsement or misrepresent the relationship between the parties; and any rights not expressly granted herein are reserved by the Licensors. +9. Disclaimer of Warranty. Unless required by applicable law or agreed to in writing, Licensor provides the Model and the Complementary Material (and each Contributor provides its Contributions) on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied, including, without limitation, any warranties or conditions of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A PARTICULAR PURPOSE. You are solely responsible for determining the appropriateness of using or redistributing the Model, Derivatives of the Model, and the Complementary Material and assume any risks associated with Your exercise of permissions under this License. +10. Limitation of Liability. In no event and under no legal theory, whether in tort (including negligence), contract, or otherwise, unless required by applicable law (such as deliberate and grossly negligent acts) or agreed to in writing, shall any Contributor be liable to You for damages, including any direct, indirect, special, incidental, or consequential damages of any character arising as a result of this License or out of the use or inability to use the Model and the Complementary Material (including but not limited to damages for loss of goodwill, work stoppage, computer failure or malfunction, or any and all other commercial damages or losses), even if such Contributor has been advised of the possibility of such damages. +11. Accepting Warranty or Additional Liability. While redistributing the Model, Derivatives of the Model and the Complementary Material thereof, You may choose to offer, and charge a fee for, acceptance of support, warranty, indemnity, or other liability obligations and/or rights consistent with this License. However, in accepting such obligations, You may act only on Your own behalf and on Your sole responsibility, not on behalf of any other Contributor, and only if You agree to indemnify, defend, and hold each Contributor harmless for any liability incurred by, or claims asserted against, such Contributor by reason of your accepting any such warranty or additional liability. +12. If any provision of this License is held to be invalid, illegal or unenforceable, the remaining provisions shall be unaffected thereby and remain valid as if such provision had not been set forth herein. + +END OF TERMS AND CONDITIONS + + + + +Attachment A + +Use Restrictions + +You agree not to use the Model or Derivatives of the Model: + +- In any way that violates any applicable national, federal, state, local or international law or regulation; +- For the purpose of exploiting, harming or attempting to exploit or harm minors in any way; +- To generate or disseminate verifiably false information and/or content with the purpose of harming others; +- To generate or disseminate personal identifiable information that can be used to harm an individual; +- To defame, disparage or otherwise harass others; +- For fully automated decision making that adversely impacts an individual’s legal rights or otherwise creates or modifies a binding, enforceable obligation; +- For any use intended to or which has the effect of discriminating against or harming individuals or groups based on online or offline social behavior or known or predicted personal or personality characteristics; +- To exploit any of the vulnerabilities of a specific group of persons based on their age, social, physical or mental characteristics, in order to materially distort the behavior of a person pertaining to that group in a manner that causes or is likely to cause that person or another person physical or psychological harm; +- For any use intended to or which has the effect of discriminating against individuals or groups based on legally protected characteristics or categories; +- To provide medical advice and medical results interpretation; +- To generate or disseminate information for the purpose to be used for administration of justice, law enforcement, immigration or asylum processes, such as predicting an individual will commit fraud/crime commitment (e.g. by text profiling, drawing causal relationships between assertions made in documents, indiscriminate and arbitrarily-targeted use). + diff --git a/README.md b/README.md new file mode 100644 index 0000000000000000000000000000000000000000..2dfddca33cf71da343a937cf0fb5e32150f16fe3 --- /dev/null +++ b/README.md @@ -0,0 +1,302 @@ +# Stable Diffusion Version 2 +![t2i](assets/stable-samples/txt2img/768/merged-0006.png) +![t2i](assets/stable-samples/txt2img/768/merged-0002.png) +![t2i](assets/stable-samples/txt2img/768/merged-0005.png) + +This repository contains [Stable Diffusion](https://github.com/CompVis/stable-diffusion) models trained from scratch and will be continuously updated with +new checkpoints. The following list provides an overview of all currently available models. More coming soon. + +## News + + +**March 24, 2023** + +*Stable UnCLIP 2.1* + +- New stable diffusion finetune (_Stable unCLIP 2.1_, [Hugging Face](https://huggingface.co/stabilityai/)) at 768x768 resolution, based on SD2.1-768. This model allows for image variations and mixing operations as described in [*Hierarchical Text-Conditional Image Generation with CLIP Latents*](https://arxiv.org/abs/2204.06125), and, thanks to its modularity, can be combined with other models such as [KARLO](https://github.com/kakaobrain/karlo). Comes in two variants: [*Stable unCLIP-L*](https://huggingface.co/stabilityai/stable-diffusion-2-1-unclip/blob/main/sd21-unclip-l.ckpt) and [*Stable unCLIP-H*](https://huggingface.co/stabilityai/stable-diffusion-2-1-unclip/blob/main/sd21-unclip-h.ckpt), which are conditioned on CLIP ViT-L and ViT-H image embeddings, respectively. Instructions are available [here](doc/UNCLIP.MD). + +- A public demo of SD-unCLIP is already available at [clipdrop.co/stable-diffusion-reimagine](https://clipdrop.co/stable-diffusion-reimagine) + + +**December 7, 2022** + +*Version 2.1* + +- New stable diffusion model (_Stable Diffusion 2.1-v_, [Hugging Face](https://huggingface.co/stabilityai/stable-diffusion-2-1)) at 768x768 resolution and (_Stable Diffusion 2.1-base_, [HuggingFace](https://huggingface.co/stabilityai/stable-diffusion-2-1-base)) at 512x512 resolution, both based on the same number of parameters and architecture as 2.0 and fine-tuned on 2.0, on a less restrictive NSFW filtering of the [LAION-5B](https://laion.ai/blog/laion-5b/) dataset. +Per default, the attention operation of the model is evaluated at full precision when `xformers` is not installed. To enable fp16 (which can cause numerical instabilities with the vanilla attention module on the v2.1 model) , run your script with `ATTN_PRECISION=fp16 python ` + +**November 24, 2022** + +*Version 2.0* + +- New stable diffusion model (_Stable Diffusion 2.0-v_) at 768x768 resolution. Same number of parameters in the U-Net as 1.5, but uses [OpenCLIP-ViT/H](https://github.com/mlfoundations/open_clip) as the text encoder and is trained from scratch. _SD 2.0-v_ is a so-called [v-prediction](https://arxiv.org/abs/2202.00512) model. +- The above model is finetuned from _SD 2.0-base_, which was trained as a standard noise-prediction model on 512x512 images and is also made available. +- Added a [x4 upscaling latent text-guided diffusion model](#image-upscaling-with-stable-diffusion). +- New [depth-guided stable diffusion model](#depth-conditional-stable-diffusion), finetuned from _SD 2.0-base_. The model is conditioned on monocular depth estimates inferred via [MiDaS](https://github.com/isl-org/MiDaS) and can be used for structure-preserving img2img and shape-conditional synthesis. + + ![d2i](assets/stable-samples/depth2img/depth2img01.png) +- A [text-guided inpainting model](#image-inpainting-with-stable-diffusion), finetuned from SD _2.0-base_. + +We follow the [original repository](https://github.com/CompVis/stable-diffusion) and provide basic inference scripts to sample from the models. + +________________ +*The original Stable Diffusion model was created in a collaboration with [CompVis](https://arxiv.org/abs/2202.00512) and [RunwayML](https://runwayml.com/) and builds upon the work:* + +[**High-Resolution Image Synthesis with Latent Diffusion Models**](https://ommer-lab.com/research/latent-diffusion-models/)
+[Robin Rombach](https://github.com/rromb)\*, +[Andreas Blattmann](https://github.com/ablattmann)\*, +[Dominik Lorenz](https://github.com/qp-qp)\, +[Patrick Esser](https://github.com/pesser), +[Björn Ommer](https://hci.iwr.uni-heidelberg.de/Staff/bommer)
+_[CVPR '22 Oral](https://openaccess.thecvf.com/content/CVPR2022/html/Rombach_High-Resolution_Image_Synthesis_With_Latent_Diffusion_Models_CVPR_2022_paper.html) | +[GitHub](https://github.com/CompVis/latent-diffusion) | [arXiv](https://arxiv.org/abs/2112.10752) | [Project page](https://ommer-lab.com/research/latent-diffusion-models/)_ + +and [many others](#shout-outs). + +Stable Diffusion is a latent text-to-image diffusion model. +________________________________ + +## Requirements + +You can update an existing [latent diffusion](https://github.com/CompVis/latent-diffusion) environment by running + +``` +conda install pytorch==1.12.1 torchvision==0.13.1 -c pytorch +pip install transformers==4.19.2 diffusers invisible-watermark +pip install -e . +``` +#### xformers efficient attention +For more efficiency and speed on GPUs, +we highly recommended installing the [xformers](https://github.com/facebookresearch/xformers) +library. + +Tested on A100 with CUDA 11.4. +Installation needs a somewhat recent version of nvcc and gcc/g++, obtain those, e.g., via +```commandline +export CUDA_HOME=/usr/local/cuda-11.4 +conda install -c nvidia/label/cuda-11.4.0 cuda-nvcc +conda install -c conda-forge gcc +conda install -c conda-forge gxx_linux-64==9.5.0 +``` + +Then, run the following (compiling takes up to 30 min). + +```commandline +cd .. +git clone https://github.com/facebookresearch/xformers.git +cd xformers +git submodule update --init --recursive +pip install -r requirements.txt +pip install -e . +cd ../stablediffusion +``` +Upon successful installation, the code will automatically default to [memory efficient attention](https://github.com/facebookresearch/xformers) +for the self- and cross-attention layers in the U-Net and autoencoder. + +## General Disclaimer +Stable Diffusion models are general text-to-image diffusion models and therefore mirror biases and (mis-)conceptions that are present +in their training data. Although efforts were made to reduce the inclusion of explicit pornographic material, **we do not recommend using the provided weights for services or products without additional safety mechanisms and considerations. +The weights are research artifacts and should be treated as such.** +Details on the training procedure and data, as well as the intended use of the model can be found in the corresponding [model card](https://huggingface.co/stabilityai/stable-diffusion-2). +The weights are available via [the StabilityAI organization at Hugging Face](https://huggingface.co/StabilityAI) under the [CreativeML Open RAIL++-M License](LICENSE-MODEL). + + + +## Stable Diffusion v2 + +Stable Diffusion v2 refers to a specific configuration of the model +architecture that uses a downsampling-factor 8 autoencoder with an 865M UNet +and OpenCLIP ViT-H/14 text encoder for the diffusion model. The _SD 2-v_ model produces 768x768 px outputs. + +Evaluations with different classifier-free guidance scales (1.5, 2.0, 3.0, 4.0, +5.0, 6.0, 7.0, 8.0) and 50 DDIM sampling steps show the relative improvements of the checkpoints: + +![sd evaluation results](assets/model-variants.jpg) + + + +### Text-to-Image +![txt2img-stable2](assets/stable-samples/txt2img/merged-0003.png) +![txt2img-stable2](assets/stable-samples/txt2img/merged-0001.png) + +Stable Diffusion 2 is a latent diffusion model conditioned on the penultimate text embeddings of a CLIP ViT-H/14 text encoder. +We provide a [reference script for sampling](#reference-sampling-script). +#### Reference Sampling Script + +This script incorporates an [invisible watermarking](https://github.com/ShieldMnt/invisible-watermark) of the outputs, to help viewers [identify the images as machine-generated](scripts/tests/test_watermark.py). +We provide the configs for the _SD2-v_ (768px) and _SD2-base_ (512px) model. + +First, download the weights for [_SD2.1-v_](https://huggingface.co/stabilityai/stable-diffusion-2-1) and [_SD2.1-base_](https://huggingface.co/stabilityai/stable-diffusion-2-1-base). + +To sample from the _SD2.1-v_ model, run the following: + +``` +python scripts/txt2img.py --prompt "a professional photograph of an astronaut riding a horse" --ckpt --config configs/stable-diffusion/v2-inference-v.yaml --H 768 --W 768 +``` +or try out the Web Demo: [![Hugging Face Spaces](https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Spaces-blue)](https://huggingface.co/spaces/stabilityai/stable-diffusion). + +To sample from the base model, use +``` +python scripts/txt2img.py --prompt "a professional photograph of an astronaut riding a horse" --ckpt --config +``` + +By default, this uses the [DDIM sampler](https://arxiv.org/abs/2010.02502), and renders images of size 768x768 (which it was trained on) in 50 steps. +Empirically, the v-models can be sampled with higher guidance scales. + +Note: The inference config for all model versions is designed to be used with EMA-only checkpoints. +For this reason `use_ema=False` is set in the configuration, otherwise the code will try to switch from +non-EMA to EMA weights. + +#### Enable Intel® Extension for PyTorch* optimizations in Text-to-Image script + +If you're planning on running Text-to-Image on Intel® CPU, try to sample an image with TorchScript and Intel® Extension for PyTorch* optimizations. Intel® Extension for PyTorch* extends PyTorch by enabling up-to-date features optimizations for an extra performance boost on Intel® hardware. It can optimize memory layout of the operators to Channel Last memory format, which is generally beneficial for Intel CPUs, take advantage of the most advanced instruction set available on a machine, optimize operators and many more. + +**Prerequisites** + +Before running the script, make sure you have all needed libraries installed. (the optimization was checked on `Ubuntu 20.04`). Install [jemalloc](https://github.com/jemalloc/jemalloc), [numactl](https://linux.die.net/man/8/numactl), Intel® OpenMP and Intel® Extension for PyTorch*. + +```bash +apt-get install numactl libjemalloc-dev +pip install intel-openmp +pip install intel_extension_for_pytorch -f https://software.intel.com/ipex-whl-stable +``` + +To sample from the _SD2.1-v_ model with TorchScript+IPEX optimizations, run the following. Remember to specify desired number of instances you want to run the program on ([more](https://github.com/intel/intel-extension-for-pytorch/blob/master/intel_extension_for_pytorch/cpu/launch.py#L48)). + +``` +MALLOC_CONF=oversize_threshold:1,background_thread:true,metadata_thp:auto,dirty_decay_ms:9000000000,muzzy_decay_ms:9000000000 python -m intel_extension_for_pytorch.cpu.launch --ninstance --enable_jemalloc scripts/txt2img.py --prompt \"a corgi is playing guitar, oil on canvas\" --ckpt --config configs/stable-diffusion/intel/v2-inference-v-fp32.yaml --H 768 --W 768 --precision full --device cpu --torchscript --ipex +``` + +To sample from the base model with IPEX optimizations, use + +``` +MALLOC_CONF=oversize_threshold:1,background_thread:true,metadata_thp:auto,dirty_decay_ms:9000000000,muzzy_decay_ms:9000000000 python -m intel_extension_for_pytorch.cpu.launch --ninstance --enable_jemalloc scripts/txt2img.py --prompt \"a corgi is playing guitar, oil on canvas\" --ckpt --config configs/stable-diffusion/intel/v2-inference-fp32.yaml --n_samples 1 --n_iter 4 --precision full --device cpu --torchscript --ipex +``` + +If you're using a CPU that supports `bfloat16`, consider sample from the model with bfloat16 enabled for a performance boost, like so + +```bash +# SD2.1-v +MALLOC_CONF=oversize_threshold:1,background_thread:true,metadata_thp:auto,dirty_decay_ms:9000000000,muzzy_decay_ms:9000000000 python -m intel_extension_for_pytorch.cpu.launch --ninstance --enable_jemalloc scripts/txt2img.py --prompt \"a corgi is playing guitar, oil on canvas\" --ckpt --config configs/stable-diffusion/intel/v2-inference-v-bf16.yaml --H 768 --W 768 --precision full --device cpu --torchscript --ipex --bf16 +# SD2.1-base +MALLOC_CONF=oversize_threshold:1,background_thread:true,metadata_thp:auto,dirty_decay_ms:9000000000,muzzy_decay_ms:9000000000 python -m intel_extension_for_pytorch.cpu.launch --ninstance --enable_jemalloc scripts/txt2img.py --prompt \"a corgi is playing guitar, oil on canvas\" --ckpt --config configs/stable-diffusion/intel/v2-inference-bf16.yaml --precision full --device cpu --torchscript --ipex --bf16 +``` + +### Image Modification with Stable Diffusion + +![depth2img-stable2](assets/stable-samples/depth2img/merged-0000.png) +#### Depth-Conditional Stable Diffusion + +To augment the well-established [img2img](https://github.com/CompVis/stable-diffusion#image-modification-with-stable-diffusion) functionality of Stable Diffusion, we provide a _shape-preserving_ stable diffusion model. + + +Note that the original method for image modification introduces significant semantic changes w.r.t. the initial image. +If that is not desired, download our [depth-conditional stable diffusion](https://huggingface.co/stabilityai/stable-diffusion-2-depth) model and the `dpt_hybrid` MiDaS [model weights](https://github.com/intel-isl/DPT/releases/download/1_0/dpt_hybrid-midas-501f0c75.pt), place the latter in a folder `midas_models` and sample via +``` +python scripts/gradio/depth2img.py configs/stable-diffusion/v2-midas-inference.yaml +``` + +or + +``` +streamlit run scripts/streamlit/depth2img.py configs/stable-diffusion/v2-midas-inference.yaml +``` + +This method can be used on the samples of the base model itself. +For example, take [this sample](assets/stable-samples/depth2img/old_man.png) generated by an anonymous discord user. +Using the [gradio](https://gradio.app) or [streamlit](https://streamlit.io/) script `depth2img.py`, the MiDaS model first infers a monocular depth estimate given this input, +and the diffusion model is then conditioned on the (relative) depth output. + +

+ depth2image
+ +

+ +This model is particularly useful for a photorealistic style; see the [examples](assets/stable-samples/depth2img). +For a maximum strength of 1.0, the model removes all pixel-based information and only relies on the text prompt and the inferred monocular depth estimate. + +![depth2img-stable3](assets/stable-samples/depth2img/merged-0005.png) + +#### Classic Img2Img + +For running the "classic" img2img, use +``` +python scripts/img2img.py --prompt "A fantasy landscape, trending on artstation" --init-img --strength 0.8 --ckpt +``` +and adapt the checkpoint and config paths accordingly. + +### Image Upscaling with Stable Diffusion +![upscaling-x4](assets/stable-samples/upscaling/merged-dog.png) +After [downloading the weights](https://huggingface.co/stabilityai/stable-diffusion-x4-upscaler), run +``` +python scripts/gradio/superresolution.py configs/stable-diffusion/x4-upscaling.yaml +``` + +or + +``` +streamlit run scripts/streamlit/superresolution.py -- configs/stable-diffusion/x4-upscaling.yaml +``` + +for a Gradio or Streamlit demo of the text-guided x4 superresolution model. +This model can be used both on real inputs and on synthesized examples. For the latter, we recommend setting a higher +`noise_level`, e.g. `noise_level=100`. + +### Image Inpainting with Stable Diffusion + +![inpainting-stable2](assets/stable-inpainting/merged-leopards.png) + +[Download the SD 2.0-inpainting checkpoint](https://huggingface.co/stabilityai/stable-diffusion-2-inpainting) and run + +``` +python scripts/gradio/inpainting.py configs/stable-diffusion/v2-inpainting-inference.yaml +``` + +or + +``` +streamlit run scripts/streamlit/inpainting.py -- configs/stable-diffusion/v2-inpainting-inference.yaml +``` + +for a Gradio or Streamlit demo of the inpainting model. +This scripts adds invisible watermarking to the demo in the [RunwayML](https://github.com/runwayml/stable-diffusion/blob/main/scripts/inpaint_st.py) repository, but both should work interchangeably with the checkpoints/configs. + + + +## Shout-Outs +- Thanks to [Hugging Face](https://huggingface.co/) and in particular [Apolinário](https://github.com/apolinario) for support with our model releases! +- Stable Diffusion would not be possible without [LAION](https://laion.ai/) and their efforts to create open, large-scale datasets. +- The [DeepFloyd team](https://twitter.com/deepfloydai) at Stability AI, for creating the subset of [LAION-5B](https://laion.ai/blog/laion-5b/) dataset used to train the model. +- Stable Diffusion 2.0 uses [OpenCLIP](https://laion.ai/blog/large-openclip/), trained by [Romain Beaumont](https://github.com/rom1504). +- Our codebase for the diffusion models builds heavily on [OpenAI's ADM codebase](https://github.com/openai/guided-diffusion) +and [https://github.com/lucidrains/denoising-diffusion-pytorch](https://github.com/lucidrains/denoising-diffusion-pytorch). +Thanks for open-sourcing! +- [CompVis](https://github.com/CompVis/stable-diffusion) initial stable diffusion release +- [Patrick](https://github.com/pesser)'s [implementation](https://github.com/runwayml/stable-diffusion/blob/main/scripts/inpaint_st.py) of the streamlit demo for inpainting. +- `img2img` is an application of [SDEdit](https://arxiv.org/abs/2108.01073) by [Chenlin Meng](https://cs.stanford.edu/~chenlin/) from the [Stanford AI Lab](https://cs.stanford.edu/~ermon/website/). +- [Kat's implementation]((https://github.com/CompVis/latent-diffusion/pull/51)) of the [PLMS](https://arxiv.org/abs/2202.09778) sampler, and [more](https://github.com/crowsonkb/k-diffusion). +- [DPMSolver](https://arxiv.org/abs/2206.00927) [integration](https://github.com/CompVis/stable-diffusion/pull/440) by [Cheng Lu](https://github.com/LuChengTHU). +- Facebook's [xformers](https://github.com/facebookresearch/xformers) for efficient attention computation. +- [MiDaS](https://github.com/isl-org/MiDaS) for monocular depth estimation. + + +## License + +The code in this repository is released under the MIT License. + +The weights are available via [the StabilityAI organization at Hugging Face](https://huggingface.co/StabilityAI), and released under the [CreativeML Open RAIL++-M License](LICENSE-MODEL) License. + +## BibTeX + +``` +@misc{rombach2021highresolution, + title={High-Resolution Image Synthesis with Latent Diffusion Models}, + author={Robin Rombach and Andreas Blattmann and Dominik Lorenz and Patrick Esser and Björn Ommer}, + year={2021}, + eprint={2112.10752}, + archivePrefix={arXiv}, + primaryClass={cs.CV} +} +``` + + diff --git a/assets/model-variants.jpg b/assets/model-variants.jpg new file mode 100644 index 0000000000000000000000000000000000000000..de5bb3fc8af5d60d0c0fac8c0c866e9f8d857ba3 Binary files /dev/null and b/assets/model-variants.jpg differ diff --git a/assets/modelfigure.png b/assets/modelfigure.png new file mode 100644 index 0000000000000000000000000000000000000000..6b1d3e6b9d59fd8d38468e7bce47c903a4e1c932 Binary files /dev/null and b/assets/modelfigure.png differ diff --git a/assets/rick.jpeg b/assets/rick.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..995486061ba50bd0ae2e213c72de87a27326632f Binary files /dev/null and b/assets/rick.jpeg differ diff --git a/assets/stable-inpainting/inpainting.gif b/assets/stable-inpainting/inpainting.gif new file mode 100644 index 0000000000000000000000000000000000000000..4f511347bc057f858db95015725488560ec1f676 Binary files /dev/null and b/assets/stable-inpainting/inpainting.gif differ diff --git a/assets/stable-inpainting/merged-leopards.png b/assets/stable-inpainting/merged-leopards.png new file mode 100644 index 0000000000000000000000000000000000000000..b877158fff2a15d4016135ce2307e19ece877b59 --- /dev/null +++ b/assets/stable-inpainting/merged-leopards.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:94a05d717a340d7b240283e72e91984e82093750ba066aa05ab0759188467e69 +size 4958974 diff --git a/assets/stable-samples/depth2img/d2i.gif b/assets/stable-samples/depth2img/d2i.gif new file mode 100644 index 0000000000000000000000000000000000000000..265c1d1ae56feb60c5aabbf88526a6ed47fda383 --- /dev/null +++ b/assets/stable-samples/depth2img/d2i.gif @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7236fc7f4349740c537ef9c8730590c15d198aaf42925a46755ded26bc436bc4 +size 1140400 diff --git a/assets/stable-samples/depth2img/depth2fantasy.jpeg b/assets/stable-samples/depth2img/depth2fantasy.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..8c0c94ee7354cc19e4c86134900397c1ed13eb01 Binary files /dev/null and b/assets/stable-samples/depth2img/depth2fantasy.jpeg differ diff --git a/assets/stable-samples/depth2img/depth2img01.png b/assets/stable-samples/depth2img/depth2img01.png new file mode 100644 index 0000000000000000000000000000000000000000..f062b45a32bf71ce830b4ab2401475d421417d03 --- /dev/null +++ b/assets/stable-samples/depth2img/depth2img01.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:60cb68c273602eae8e2fb769a2848e55844d812196260ada112a9aecc604f735 +size 3324111 diff --git a/assets/stable-samples/depth2img/depth2img02.png b/assets/stable-samples/depth2img/depth2img02.png new file mode 100644 index 0000000000000000000000000000000000000000..f91cad5b401955221d8875529578504762ca6bfc --- /dev/null +++ b/assets/stable-samples/depth2img/depth2img02.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:adabff92594b17a2554408257d2ab61eb9b89270d5917eafd44a9b75740aab04 +size 1775470 diff --git a/assets/stable-samples/depth2img/merged-0000.png b/assets/stable-samples/depth2img/merged-0000.png new file mode 100644 index 0000000000000000000000000000000000000000..fce6c7a0b76bd34a2028a20efa5e5a497953cc7e --- /dev/null +++ b/assets/stable-samples/depth2img/merged-0000.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b19da6052f01a3b115ac3315ef5db1b7dcdb58091879c0dfe3895a7765a491ac +size 2129264 diff --git a/assets/stable-samples/depth2img/merged-0004.png b/assets/stable-samples/depth2img/merged-0004.png new file mode 100644 index 0000000000000000000000000000000000000000..41641d3e2d07575fb5fa0e87609770ab1bae8126 --- /dev/null +++ b/assets/stable-samples/depth2img/merged-0004.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0d3ae25218f696375aa953e55d12d034da3bd7abce68616a3af916bdae01cc86 +size 1448945 diff --git a/assets/stable-samples/depth2img/merged-0005.png b/assets/stable-samples/depth2img/merged-0005.png new file mode 100644 index 0000000000000000000000000000000000000000..46634965fdf77e2134e2ac574fe02c12cd34d96e --- /dev/null +++ b/assets/stable-samples/depth2img/merged-0005.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:935252e202a3a6cafa476443f3a0ae3ac95cf85c37c0133f4f32af2aafb8f9ab +size 4546641 diff --git a/assets/stable-samples/depth2img/midas.jpeg b/assets/stable-samples/depth2img/midas.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..4890ab7c6ca378f27dcf1eef6012848798713c94 Binary files /dev/null and b/assets/stable-samples/depth2img/midas.jpeg differ diff --git a/assets/stable-samples/depth2img/old_man.png b/assets/stable-samples/depth2img/old_man.png new file mode 100644 index 0000000000000000000000000000000000000000..44bbc0051b5f12743b1c7ef92f23ed5758eed4e2 Binary files /dev/null and b/assets/stable-samples/depth2img/old_man.png differ diff --git a/assets/stable-samples/img2img/mountains-1.png b/assets/stable-samples/img2img/mountains-1.png new file mode 100644 index 0000000000000000000000000000000000000000..d01b8350743e3bd4fdf653d1563ee7d5c2153323 Binary files /dev/null and b/assets/stable-samples/img2img/mountains-1.png differ diff --git a/assets/stable-samples/img2img/mountains-2.png b/assets/stable-samples/img2img/mountains-2.png new file mode 100644 index 0000000000000000000000000000000000000000..e9f4e708535f0e5b53372a3a39e6aa31dce383fd Binary files /dev/null and b/assets/stable-samples/img2img/mountains-2.png differ diff --git a/assets/stable-samples/img2img/mountains-3.png b/assets/stable-samples/img2img/mountains-3.png new file mode 100644 index 0000000000000000000000000000000000000000..017de3012c2f03e4f87cce21b4d3342713b9ae95 Binary files /dev/null and b/assets/stable-samples/img2img/mountains-3.png differ diff --git a/assets/stable-samples/img2img/sketch-mountains-input.jpg b/assets/stable-samples/img2img/sketch-mountains-input.jpg new file mode 100644 index 0000000000000000000000000000000000000000..79d652b8003bbcd1d0c0ba2d984dbbe299ac5916 Binary files /dev/null and b/assets/stable-samples/img2img/sketch-mountains-input.jpg differ diff --git a/assets/stable-samples/img2img/upscaling-in.png b/assets/stable-samples/img2img/upscaling-in.png new file mode 100644 index 0000000000000000000000000000000000000000..6a16bf53a95850a4eb7730105cea61af7c435c5e --- /dev/null +++ b/assets/stable-samples/img2img/upscaling-in.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:16e043b62bdfcc5be7d0eca5c747878b78e4e6ffaeb3cd1257568cbc2b5e6f7a +size 1167237 diff --git a/assets/stable-samples/img2img/upscaling-out.png b/assets/stable-samples/img2img/upscaling-out.png new file mode 100644 index 0000000000000000000000000000000000000000..b7926bc81099736f7c8df32cadc4481c07eddbd6 --- /dev/null +++ b/assets/stable-samples/img2img/upscaling-out.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c171218814d141f87884672cb00ae07c3ed0e14ce7f7023f2041678e01d93f59 +size 1317941 diff --git a/assets/stable-samples/stable-unclip/houses_out.jpeg b/assets/stable-samples/stable-unclip/houses_out.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..b4b8fb9e9cb3087d02b4ebac9591dbc9d44fcd50 Binary files /dev/null and b/assets/stable-samples/stable-unclip/houses_out.jpeg differ diff --git a/assets/stable-samples/stable-unclip/oldcar000.jpeg b/assets/stable-samples/stable-unclip/oldcar000.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..d61c85e0d832a2fa9592d375bf5a1ca1d9bcab09 Binary files /dev/null and b/assets/stable-samples/stable-unclip/oldcar000.jpeg differ diff --git a/assets/stable-samples/stable-unclip/oldcar500.jpeg b/assets/stable-samples/stable-unclip/oldcar500.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..175d2825101fd64f5a11456c2f86e15f1b281a83 Binary files /dev/null and b/assets/stable-samples/stable-unclip/oldcar500.jpeg differ diff --git a/assets/stable-samples/stable-unclip/oldcar800.jpeg b/assets/stable-samples/stable-unclip/oldcar800.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..d4d91130c0fb4a97389b27cdc8824c0df4d9c485 Binary files /dev/null and b/assets/stable-samples/stable-unclip/oldcar800.jpeg differ diff --git a/assets/stable-samples/stable-unclip/panda.jpg b/assets/stable-samples/stable-unclip/panda.jpg new file mode 100644 index 0000000000000000000000000000000000000000..49aa1ba4d5a818df96988f3dc2ade49200c767ab Binary files /dev/null and b/assets/stable-samples/stable-unclip/panda.jpg differ diff --git a/assets/stable-samples/stable-unclip/plates_out.jpeg b/assets/stable-samples/stable-unclip/plates_out.jpeg new file mode 100644 index 0000000000000000000000000000000000000000..ecbc11c3c281ad14a98457d0b531c1df19f6f9a3 Binary files /dev/null and b/assets/stable-samples/stable-unclip/plates_out.jpeg differ diff --git a/assets/stable-samples/stable-unclip/unclip-variations.png b/assets/stable-samples/stable-unclip/unclip-variations.png new file mode 100644 index 0000000000000000000000000000000000000000..80c656c49da2ffca9c2394423dc600d48b55588a --- /dev/null +++ b/assets/stable-samples/stable-unclip/unclip-variations.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e9bc45418c5c4ded4fe8ef054c6fc85fa23efe9bab4cdbc42d3ec55f2a57bc39 +size 1774803 diff --git a/assets/stable-samples/stable-unclip/unclip-variations_noise.png b/assets/stable-samples/stable-unclip/unclip-variations_noise.png new file mode 100644 index 0000000000000000000000000000000000000000..9f155786eb36901255c663ec843dafac09d89d67 --- /dev/null +++ b/assets/stable-samples/stable-unclip/unclip-variations_noise.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:83e2cdb18e95cb074db4d6d78dc7c2333936ab641e27945c358b2c4160eeb6da +size 1541201 diff --git a/assets/stable-samples/txt2img/000002025.png b/assets/stable-samples/txt2img/000002025.png new file mode 100644 index 0000000000000000000000000000000000000000..66891c142e9cfe3e0b0193d16b5a45d7b72a7d8a Binary files /dev/null and b/assets/stable-samples/txt2img/000002025.png differ diff --git a/assets/stable-samples/txt2img/000002035.png b/assets/stable-samples/txt2img/000002035.png new file mode 100644 index 0000000000000000000000000000000000000000..c707c130932a13fae76584324368fedc33faeba4 Binary files /dev/null and b/assets/stable-samples/txt2img/000002035.png differ diff --git a/assets/stable-samples/txt2img/768/merged-0001.png b/assets/stable-samples/txt2img/768/merged-0001.png new file mode 100644 index 0000000000000000000000000000000000000000..a3260bf98645b5f78e7eaf6d08b417a53c310fc2 --- /dev/null +++ b/assets/stable-samples/txt2img/768/merged-0001.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ed10e1df0f4c0f83794310e59a77098b4836d96a2b12cc809ddf39e77b1b6c94 +size 4627059 diff --git a/assets/stable-samples/txt2img/768/merged-0002.png b/assets/stable-samples/txt2img/768/merged-0002.png new file mode 100644 index 0000000000000000000000000000000000000000..297dbfd110876b8e9783e72bb6cf650cf1c760fd --- /dev/null +++ b/assets/stable-samples/txt2img/768/merged-0002.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b4a009d112633ac788fbbe7a7176d4002e95407b64672afa8104755534bb4641 +size 3458875 diff --git a/assets/stable-samples/txt2img/768/merged-0003.png b/assets/stable-samples/txt2img/768/merged-0003.png new file mode 100644 index 0000000000000000000000000000000000000000..987e2a6dd7caff691af2656e115d9d5723b418dd --- /dev/null +++ b/assets/stable-samples/txt2img/768/merged-0003.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:84adec03ab2e2d54990950a113af97750eab90135596461689efe5ebfa1ebf92 +size 3834227 diff --git a/assets/stable-samples/txt2img/768/merged-0004.png b/assets/stable-samples/txt2img/768/merged-0004.png new file mode 100644 index 0000000000000000000000000000000000000000..f70da62c63f24b9d87347f8c8613dd1cf2993886 --- /dev/null +++ b/assets/stable-samples/txt2img/768/merged-0004.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8ac938c03b4b554e1c475a4b3c5df50b72a890eec21542fa7911a8ff01bf13f4 +size 4097005 diff --git a/assets/stable-samples/txt2img/768/merged-0005.png b/assets/stable-samples/txt2img/768/merged-0005.png new file mode 100644 index 0000000000000000000000000000000000000000..067259de9ef2422adac2793d3417245fe2671b24 --- /dev/null +++ b/assets/stable-samples/txt2img/768/merged-0005.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:849b877a80752b4578afc2ece4ea0726768809298ee9a38284ba4e159d0a817c +size 2166197 diff --git a/assets/stable-samples/txt2img/768/merged-0006.png b/assets/stable-samples/txt2img/768/merged-0006.png new file mode 100644 index 0000000000000000000000000000000000000000..0002baa798eaae545f6f2fc205bf6cb965a2b97c --- /dev/null +++ b/assets/stable-samples/txt2img/768/merged-0006.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2d31dbbf76633677be3b8eba933e9eec82825925535ef9c557a3003daf16ad42 +size 4371148 diff --git a/assets/stable-samples/txt2img/merged-0001.png b/assets/stable-samples/txt2img/merged-0001.png new file mode 100644 index 0000000000000000000000000000000000000000..c76fd405d99ba9a83ea93c974d6fa12bd9555a8f --- /dev/null +++ b/assets/stable-samples/txt2img/merged-0001.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:71ca5f77befffa10a2ef6d4b69f8bb721e7ebd7ea03538e2c359dc44f526b0e8 +size 2409358 diff --git a/assets/stable-samples/txt2img/merged-0003.png b/assets/stable-samples/txt2img/merged-0003.png new file mode 100644 index 0000000000000000000000000000000000000000..cdc0e618cf42d225d417901b64b1203732eeb879 --- /dev/null +++ b/assets/stable-samples/txt2img/merged-0003.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9fde5a40c512d61e2390e70d9f14b0d33f0af84cbde2dcd9d86e1f9b38072266 +size 2273528 diff --git a/assets/stable-samples/txt2img/merged-0005.png b/assets/stable-samples/txt2img/merged-0005.png new file mode 100644 index 0000000000000000000000000000000000000000..88fcd04df9f576070d97f5655733b7a185352676 --- /dev/null +++ b/assets/stable-samples/txt2img/merged-0005.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a417aadc1d91b91531ca6bbf89840a36f432d8e9382aaa953610bedce22ff76f +size 2576264 diff --git a/assets/stable-samples/txt2img/merged-0006.png b/assets/stable-samples/txt2img/merged-0006.png new file mode 100644 index 0000000000000000000000000000000000000000..831c2096d8a99cb50323e48e17cc3f4dbba60f8c --- /dev/null +++ b/assets/stable-samples/txt2img/merged-0006.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1d55ba7d103da275b4612976e93f405fcb593f7e6a6fda31f2e180b41c8e4f59 +size 2638534 diff --git a/assets/stable-samples/txt2img/merged-0007.png b/assets/stable-samples/txt2img/merged-0007.png new file mode 100644 index 0000000000000000000000000000000000000000..05e0f3fd7c8b6db82f15026053687558906e9e30 --- /dev/null +++ b/assets/stable-samples/txt2img/merged-0007.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:920ccf908b7fa5073a7c5cd3f4e109b5e66f7e29517ef5462ca55e931d0b5689 +size 2406799 diff --git a/assets/stable-samples/upscaling/merged-dog.png b/assets/stable-samples/upscaling/merged-dog.png new file mode 100644 index 0000000000000000000000000000000000000000..5af3eff5d6e42150b29a2c5a6cad5b3b025212d6 --- /dev/null +++ b/assets/stable-samples/upscaling/merged-dog.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d85d15bd51b3fa162f2b020ccac5a64b10ced728c6f22dcba183dc65ab6e8b5a +size 1823141 diff --git a/assets/stable-samples/upscaling/sampled-bear-x4.png b/assets/stable-samples/upscaling/sampled-bear-x4.png new file mode 100644 index 0000000000000000000000000000000000000000..00dc7730d2504b40de6958d41411252852a246cd --- /dev/null +++ b/assets/stable-samples/upscaling/sampled-bear-x4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d4f2aaa8eb3054cda0a6e8577170d09c1494809cceca21973497602e17a22f1e +size 3155724 diff --git a/assets/stable-samples/upscaling/snow-leopard-x4.png b/assets/stable-samples/upscaling/snow-leopard-x4.png new file mode 100644 index 0000000000000000000000000000000000000000..30a30a0929d898a5ea71b0b5e6403f54cb992da8 --- /dev/null +++ b/assets/stable-samples/upscaling/snow-leopard-x4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fe8231dddcf77ada4b46f6949b4ea7757ff2d006253e1807ba6e1168077aad19 +size 3887284 diff --git a/checkpoints/checkpoints.txt b/checkpoints/checkpoints.txt new file mode 100644 index 0000000000000000000000000000000000000000..d92df3108fcbd2608ee4f9901d6f2a0eb081674f --- /dev/null +++ b/checkpoints/checkpoints.txt @@ -0,0 +1 @@ +Put unCLIP checkpoints here. \ No newline at end of file diff --git a/configs/karlo/decoder_900M_vit_l.yaml b/configs/karlo/decoder_900M_vit_l.yaml new file mode 100644 index 0000000000000000000000000000000000000000..02a35303aaa3c3556c649768643814cbd7c8fe66 --- /dev/null +++ b/configs/karlo/decoder_900M_vit_l.yaml @@ -0,0 +1,37 @@ +model: + type: t2i-decoder + diffusion_sampler: uniform + hparams: + image_size: 64 + num_channels: 320 + num_res_blocks: 3 + channel_mult: '' + attention_resolutions: 32,16,8 + num_heads: -1 + num_head_channels: 64 + num_heads_upsample: -1 + use_scale_shift_norm: true + dropout: 0.1 + clip_dim: 768 + clip_emb_mult: 4 + text_ctx: 77 + xf_width: 1536 + xf_layers: 0 + xf_heads: 0 + xf_final_ln: false + resblock_updown: true + learn_sigma: true + text_drop: 0.3 + clip_emb_type: image + clip_emb_drop: 0.1 + use_plm: true + +diffusion: + steps: 1000 + learn_sigma: true + sigma_small: false + noise_schedule: squaredcos_cap_v2 + use_kl: false + predict_xstart: false + rescale_learned_sigmas: true + timestep_respacing: '' diff --git a/configs/karlo/improved_sr_64_256_1.4B.yaml b/configs/karlo/improved_sr_64_256_1.4B.yaml new file mode 100644 index 0000000000000000000000000000000000000000..282d3cb0db6a51e076e9848f8d26e55e7ae406bb --- /dev/null +++ b/configs/karlo/improved_sr_64_256_1.4B.yaml @@ -0,0 +1,27 @@ +model: + type: improved_sr_64_256 + diffusion_sampler: uniform + hparams: + channels: 320 + depth: 3 + channels_multiple: + - 1 + - 2 + - 3 + - 4 + dropout: 0.0 + +diffusion: + steps: 1000 + learn_sigma: false + sigma_small: true + noise_schedule: squaredcos_cap_v2 + use_kl: false + predict_xstart: false + rescale_learned_sigmas: true + timestep_respacing: '7' + + +sampling: + timestep_respacing: '7' # fix + clip_denoise: true diff --git a/configs/karlo/prior_1B_vit_l.yaml b/configs/karlo/prior_1B_vit_l.yaml new file mode 100644 index 0000000000000000000000000000000000000000..159330d3085527ae142f122b182cfec7a560a3db --- /dev/null +++ b/configs/karlo/prior_1B_vit_l.yaml @@ -0,0 +1,21 @@ +model: + type: prior + diffusion_sampler: uniform + hparams: + text_ctx: 77 + xf_width: 2048 + xf_layers: 20 + xf_heads: 32 + xf_final_ln: true + text_drop: 0.2 + clip_dim: 768 + +diffusion: + steps: 1000 + learn_sigma: false + sigma_small: true + noise_schedule: squaredcos_cap_v2 + use_kl: false + predict_xstart: true + rescale_learned_sigmas: false + timestep_respacing: '' diff --git a/configs/stable-diffusion/intel/v2-inference-bf16.yaml b/configs/stable-diffusion/intel/v2-inference-bf16.yaml new file mode 100644 index 0000000000000000000000000000000000000000..66f0dbd8331d84a78a3448fdb2a69dcd7020eb59 --- /dev/null +++ b/configs/stable-diffusion/intel/v2-inference-bf16.yaml @@ -0,0 +1,71 @@ +# Copyright (C) 2022 Intel Corporation +# SPDX-License-Identifier: MIT + +model: + base_learning_rate: 1.0e-4 + target: ldm.models.diffusion.ddpm.LatentDiffusion + params: + linear_start: 0.00085 + linear_end: 0.0120 + num_timesteps_cond: 1 + log_every_t: 200 + timesteps: 1000 + first_stage_key: "jpg" + cond_stage_key: "txt" + image_size: 64 + channels: 4 + cond_stage_trainable: false + conditioning_key: crossattn + monitor: val/loss_simple_ema + scale_factor: 0.18215 + use_ema: False # we set this to false because this is an inference only config + + unet_config: + target: ldm.modules.diffusionmodules.openaimodel.UNetModel + params: + use_checkpoint: False + use_fp16: False + use_bf16: True + image_size: 32 # unused + in_channels: 4 + out_channels: 4 + model_channels: 320 + attention_resolutions: [ 4, 2, 1 ] + num_res_blocks: 2 + channel_mult: [ 1, 2, 4, 4 ] + num_head_channels: 64 # need to fix for flash-attn + use_spatial_transformer: True + use_linear_in_transformer: True + transformer_depth: 1 + context_dim: 1024 + legacy: False + + first_stage_config: + target: ldm.models.autoencoder.AutoencoderKL + params: + embed_dim: 4 + monitor: val/rec_loss + ddconfig: + #attn_type: "vanilla-xformers" + double_z: true + z_channels: 4 + resolution: 256 + in_channels: 3 + out_ch: 3 + ch: 128 + ch_mult: + - 1 + - 2 + - 4 + - 4 + num_res_blocks: 2 + attn_resolutions: [] + dropout: 0.0 + lossconfig: + target: torch.nn.Identity + + cond_stage_config: + target: ldm.modules.encoders.modules.FrozenOpenCLIPEmbedder + params: + freeze: True + layer: "penultimate" diff --git a/configs/stable-diffusion/intel/v2-inference-fp32.yaml b/configs/stable-diffusion/intel/v2-inference-fp32.yaml new file mode 100644 index 0000000000000000000000000000000000000000..7b66ac85830e8884b26cb9e70e094172c9ee5e62 --- /dev/null +++ b/configs/stable-diffusion/intel/v2-inference-fp32.yaml @@ -0,0 +1,70 @@ +# Copyright (C) 2022 Intel Corporation +# SPDX-License-Identifier: MIT + +model: + base_learning_rate: 1.0e-4 + target: ldm.models.diffusion.ddpm.LatentDiffusion + params: + linear_start: 0.00085 + linear_end: 0.0120 + num_timesteps_cond: 1 + log_every_t: 200 + timesteps: 1000 + first_stage_key: "jpg" + cond_stage_key: "txt" + image_size: 64 + channels: 4 + cond_stage_trainable: false + conditioning_key: crossattn + monitor: val/loss_simple_ema + scale_factor: 0.18215 + use_ema: False # we set this to false because this is an inference only config + + unet_config: + target: ldm.modules.diffusionmodules.openaimodel.UNetModel + params: + use_checkpoint: False + use_fp16: False + image_size: 32 # unused + in_channels: 4 + out_channels: 4 + model_channels: 320 + attention_resolutions: [ 4, 2, 1 ] + num_res_blocks: 2 + channel_mult: [ 1, 2, 4, 4 ] + num_head_channels: 64 # need to fix for flash-attn + use_spatial_transformer: True + use_linear_in_transformer: True + transformer_depth: 1 + context_dim: 1024 + legacy: False + + first_stage_config: + target: ldm.models.autoencoder.AutoencoderKL + params: + embed_dim: 4 + monitor: val/rec_loss + ddconfig: + #attn_type: "vanilla-xformers" + double_z: true + z_channels: 4 + resolution: 256 + in_channels: 3 + out_ch: 3 + ch: 128 + ch_mult: + - 1 + - 2 + - 4 + - 4 + num_res_blocks: 2 + attn_resolutions: [] + dropout: 0.0 + lossconfig: + target: torch.nn.Identity + + cond_stage_config: + target: ldm.modules.encoders.modules.FrozenOpenCLIPEmbedder + params: + freeze: True + layer: "penultimate" diff --git a/configs/stable-diffusion/intel/v2-inference-v-bf16.yaml b/configs/stable-diffusion/intel/v2-inference-v-bf16.yaml new file mode 100644 index 0000000000000000000000000000000000000000..2b4b0e6fabce186a0e5cd29d4efa29eebee152f7 --- /dev/null +++ b/configs/stable-diffusion/intel/v2-inference-v-bf16.yaml @@ -0,0 +1,72 @@ +# Copyright (C) 2022 Intel Corporation +# SPDX-License-Identifier: MIT + +model: + base_learning_rate: 1.0e-4 + target: ldm.models.diffusion.ddpm.LatentDiffusion + params: + parameterization: "v" + linear_start: 0.00085 + linear_end: 0.0120 + num_timesteps_cond: 1 + log_every_t: 200 + timesteps: 1000 + first_stage_key: "jpg" + cond_stage_key: "txt" + image_size: 64 + channels: 4 + cond_stage_trainable: false + conditioning_key: crossattn + monitor: val/loss_simple_ema + scale_factor: 0.18215 + use_ema: False # we set this to false because this is an inference only config + + unet_config: + target: ldm.modules.diffusionmodules.openaimodel.UNetModel + params: + use_checkpoint: False + use_fp16: False + use_bf16: True + image_size: 32 # unused + in_channels: 4 + out_channels: 4 + model_channels: 320 + attention_resolutions: [ 4, 2, 1 ] + num_res_blocks: 2 + channel_mult: [ 1, 2, 4, 4 ] + num_head_channels: 64 # need to fix for flash-attn + use_spatial_transformer: True + use_linear_in_transformer: True + transformer_depth: 1 + context_dim: 1024 + legacy: False + + first_stage_config: + target: ldm.models.autoencoder.AutoencoderKL + params: + embed_dim: 4 + monitor: val/rec_loss + ddconfig: + #attn_type: "vanilla-xformers" + double_z: true + z_channels: 4 + resolution: 256 + in_channels: 3 + out_ch: 3 + ch: 128 + ch_mult: + - 1 + - 2 + - 4 + - 4 + num_res_blocks: 2 + attn_resolutions: [] + dropout: 0.0 + lossconfig: + target: torch.nn.Identity + + cond_stage_config: + target: ldm.modules.encoders.modules.FrozenOpenCLIPEmbedder + params: + freeze: True + layer: "penultimate" diff --git a/configs/stable-diffusion/intel/v2-inference-v-fp32.yaml b/configs/stable-diffusion/intel/v2-inference-v-fp32.yaml new file mode 100644 index 0000000000000000000000000000000000000000..8ccd92e295d71681750f749bb7972a62cc44b3cb --- /dev/null +++ b/configs/stable-diffusion/intel/v2-inference-v-fp32.yaml @@ -0,0 +1,71 @@ +# Copyright (C) 2022 Intel Corporation +# SPDX-License-Identifier: MIT + +model: + base_learning_rate: 1.0e-4 + target: ldm.models.diffusion.ddpm.LatentDiffusion + params: + parameterization: "v" + linear_start: 0.00085 + linear_end: 0.0120 + num_timesteps_cond: 1 + log_every_t: 200 + timesteps: 1000 + first_stage_key: "jpg" + cond_stage_key: "txt" + image_size: 64 + channels: 4 + cond_stage_trainable: false + conditioning_key: crossattn + monitor: val/loss_simple_ema + scale_factor: 0.18215 + use_ema: False # we set this to false because this is an inference only config + + unet_config: + target: ldm.modules.diffusionmodules.openaimodel.UNetModel + params: + use_checkpoint: False + use_fp16: False + image_size: 32 # unused + in_channels: 4 + out_channels: 4 + model_channels: 320 + attention_resolutions: [ 4, 2, 1 ] + num_res_blocks: 2 + channel_mult: [ 1, 2, 4, 4 ] + num_head_channels: 64 # need to fix for flash-attn + use_spatial_transformer: True + use_linear_in_transformer: True + transformer_depth: 1 + context_dim: 1024 + legacy: False + + first_stage_config: + target: ldm.models.autoencoder.AutoencoderKL + params: + embed_dim: 4 + monitor: val/rec_loss + ddconfig: + #attn_type: "vanilla-xformers" + double_z: true + z_channels: 4 + resolution: 256 + in_channels: 3 + out_ch: 3 + ch: 128 + ch_mult: + - 1 + - 2 + - 4 + - 4 + num_res_blocks: 2 + attn_resolutions: [] + dropout: 0.0 + lossconfig: + target: torch.nn.Identity + + cond_stage_config: + target: ldm.modules.encoders.modules.FrozenOpenCLIPEmbedder + params: + freeze: True + layer: "penultimate" diff --git a/configs/stable-diffusion/v2-1-stable-unclip-h-inference.yaml b/configs/stable-diffusion/v2-1-stable-unclip-h-inference.yaml new file mode 100644 index 0000000000000000000000000000000000000000..1bd0c64d3cdf2e3027fcd68b986b3236d4b4101c --- /dev/null +++ b/configs/stable-diffusion/v2-1-stable-unclip-h-inference.yaml @@ -0,0 +1,80 @@ +model: + base_learning_rate: 1.0e-04 + target: ldm.models.diffusion.ddpm.ImageEmbeddingConditionedLatentDiffusion + params: + embedding_dropout: 0.25 + parameterization: "v" + linear_start: 0.00085 + linear_end: 0.0120 + log_every_t: 200 + timesteps: 1000 + first_stage_key: "jpg" + cond_stage_key: "txt" + image_size: 96 + channels: 4 + cond_stage_trainable: false + conditioning_key: crossattn-adm + scale_factor: 0.18215 + monitor: val/loss_simple_ema + use_ema: False + + embedder_config: + target: ldm.modules.encoders.modules.FrozenOpenCLIPImageEmbedder + + noise_aug_config: + target: ldm.modules.encoders.modules.CLIPEmbeddingNoiseAugmentation + params: + timestep_dim: 1024 + noise_schedule_config: + timesteps: 1000 + beta_schedule: squaredcos_cap_v2 + + unet_config: + target: ldm.modules.diffusionmodules.openaimodel.UNetModel + params: + num_classes: "sequential" + adm_in_channels: 2048 + use_checkpoint: True + image_size: 32 # unused + in_channels: 4 + out_channels: 4 + model_channels: 320 + attention_resolutions: [ 4, 2, 1 ] + num_res_blocks: 2 + channel_mult: [ 1, 2, 4, 4 ] + num_head_channels: 64 # need to fix for flash-attn + use_spatial_transformer: True + use_linear_in_transformer: True + transformer_depth: 1 + context_dim: 1024 + legacy: False + + first_stage_config: + target: ldm.models.autoencoder.AutoencoderKL + params: + embed_dim: 4 + monitor: val/rec_loss + ddconfig: + attn_type: "vanilla-xformers" + double_z: true + z_channels: 4 + resolution: 256 + in_channels: 3 + out_ch: 3 + ch: 128 + ch_mult: + - 1 + - 2 + - 4 + - 4 + num_res_blocks: 2 + attn_resolutions: [ ] + dropout: 0.0 + lossconfig: + target: torch.nn.Identity + + cond_stage_config: + target: ldm.modules.encoders.modules.FrozenOpenCLIPEmbedder + params: + freeze: True + layer: "penultimate" diff --git a/configs/stable-diffusion/v2-1-stable-unclip-l-inference.yaml b/configs/stable-diffusion/v2-1-stable-unclip-l-inference.yaml new file mode 100644 index 0000000000000000000000000000000000000000..335fd61f3e559ba093320b67c950dec5494bf489 --- /dev/null +++ b/configs/stable-diffusion/v2-1-stable-unclip-l-inference.yaml @@ -0,0 +1,83 @@ +model: + base_learning_rate: 1.0e-04 + target: ldm.models.diffusion.ddpm.ImageEmbeddingConditionedLatentDiffusion + params: + embedding_dropout: 0.25 + parameterization: "v" + linear_start: 0.00085 + linear_end: 0.0120 + log_every_t: 200 + timesteps: 1000 + first_stage_key: "jpg" + cond_stage_key: "txt" + image_size: 96 + channels: 4 + cond_stage_trainable: false + conditioning_key: crossattn-adm + scale_factor: 0.18215 + monitor: val/loss_simple_ema + use_ema: False + + embedder_config: + target: ldm.modules.encoders.modules.ClipImageEmbedder + params: + model: "ViT-L/14" + + noise_aug_config: + target: ldm.modules.encoders.modules.CLIPEmbeddingNoiseAugmentation + params: + clip_stats_path: "checkpoints/karlo_models/ViT-L-14_stats.th" + timestep_dim: 768 + noise_schedule_config: + timesteps: 1000 + beta_schedule: squaredcos_cap_v2 + + unet_config: + target: ldm.modules.diffusionmodules.openaimodel.UNetModel + params: + num_classes: "sequential" + adm_in_channels: 1536 + use_checkpoint: True + image_size: 32 # unused + in_channels: 4 + out_channels: 4 + model_channels: 320 + attention_resolutions: [ 4, 2, 1 ] + num_res_blocks: 2 + channel_mult: [ 1, 2, 4, 4 ] + num_head_channels: 64 # need to fix for flash-attn + use_spatial_transformer: True + use_linear_in_transformer: True + transformer_depth: 1 + context_dim: 1024 + legacy: False + + first_stage_config: + target: ldm.models.autoencoder.AutoencoderKL + params: + embed_dim: 4 + monitor: val/rec_loss + ddconfig: + attn_type: "vanilla-xformers" + double_z: true + z_channels: 4 + resolution: 256 + in_channels: 3 + out_ch: 3 + ch: 128 + ch_mult: + - 1 + - 2 + - 4 + - 4 + num_res_blocks: 2 + attn_resolutions: [ ] + dropout: 0.0 + lossconfig: + target: torch.nn.Identity + + cond_stage_config: + target: ldm.modules.encoders.modules.FrozenOpenCLIPEmbedder + params: + freeze: True + layer: "penultimate" \ No newline at end of file diff --git a/configs/stable-diffusion/v2-inference-v.yaml b/configs/stable-diffusion/v2-inference-v.yaml new file mode 100644 index 0000000000000000000000000000000000000000..8ec8dfbfefe94ae8522c93017668fea78d580acf --- /dev/null +++ b/configs/stable-diffusion/v2-inference-v.yaml @@ -0,0 +1,68 @@ +model: + base_learning_rate: 1.0e-4 + target: ldm.models.diffusion.ddpm.LatentDiffusion + params: + parameterization: "v" + linear_start: 0.00085 + linear_end: 0.0120 + num_timesteps_cond: 1 + log_every_t: 200 + timesteps: 1000 + first_stage_key: "jpg" + cond_stage_key: "txt" + image_size: 64 + channels: 4 + cond_stage_trainable: false + conditioning_key: crossattn + monitor: val/loss_simple_ema + scale_factor: 0.18215 + use_ema: False # we set this to false because this is an inference only config + + unet_config: + target: ldm.modules.diffusionmodules.openaimodel.UNetModel + params: + use_checkpoint: True + use_fp16: True + image_size: 32 # unused + in_channels: 4 + out_channels: 4 + model_channels: 320 + attention_resolutions: [ 4, 2, 1 ] + num_res_blocks: 2 + channel_mult: [ 1, 2, 4, 4 ] + num_head_channels: 64 # need to fix for flash-attn + use_spatial_transformer: True + use_linear_in_transformer: True + transformer_depth: 1 + context_dim: 1024 + legacy: False + + first_stage_config: + target: ldm.models.autoencoder.AutoencoderKL + params: + embed_dim: 4 + monitor: val/rec_loss + ddconfig: + #attn_type: "vanilla-xformers" + double_z: true + z_channels: 4 + resolution: 256 + in_channels: 3 + out_ch: 3 + ch: 128 + ch_mult: + - 1 + - 2 + - 4 + - 4 + num_res_blocks: 2 + attn_resolutions: [] + dropout: 0.0 + lossconfig: + target: torch.nn.Identity + + cond_stage_config: + target: ldm.modules.encoders.modules.FrozenOpenCLIPEmbedder + params: + freeze: True + layer: "penultimate" diff --git a/configs/stable-diffusion/v2-inference.yaml b/configs/stable-diffusion/v2-inference.yaml new file mode 100644 index 0000000000000000000000000000000000000000..152c4f3c2b36c3b246a9cb10eb8166134b0d2e1c --- /dev/null +++ b/configs/stable-diffusion/v2-inference.yaml @@ -0,0 +1,67 @@ +model: + base_learning_rate: 1.0e-4 + target: ldm.models.diffusion.ddpm.LatentDiffusion + params: + linear_start: 0.00085 + linear_end: 0.0120 + num_timesteps_cond: 1 + log_every_t: 200 + timesteps: 1000 + first_stage_key: "jpg" + cond_stage_key: "txt" + image_size: 64 + channels: 4 + cond_stage_trainable: false + conditioning_key: crossattn + monitor: val/loss_simple_ema + scale_factor: 0.18215 + use_ema: False # we set this to false because this is an inference only config + + unet_config: + target: ldm.modules.diffusionmodules.openaimodel.UNetModel + params: + use_checkpoint: True + use_fp16: True + image_size: 32 # unused + in_channels: 4 + out_channels: 4 + model_channels: 320 + attention_resolutions: [ 4, 2, 1 ] + num_res_blocks: 2 + channel_mult: [ 1, 2, 4, 4 ] + num_head_channels: 64 # need to fix for flash-attn + use_spatial_transformer: True + use_linear_in_transformer: True + transformer_depth: 1 + context_dim: 1024 + legacy: False + + first_stage_config: + target: ldm.models.autoencoder.AutoencoderKL + params: + embed_dim: 4 + monitor: val/rec_loss + ddconfig: + #attn_type: "vanilla-xformers" + double_z: true + z_channels: 4 + resolution: 256 + in_channels: 3 + out_ch: 3 + ch: 128 + ch_mult: + - 1 + - 2 + - 4 + - 4 + num_res_blocks: 2 + attn_resolutions: [] + dropout: 0.0 + lossconfig: + target: torch.nn.Identity + + cond_stage_config: + target: ldm.modules.encoders.modules.FrozenOpenCLIPEmbedder + params: + freeze: True + layer: "penultimate" diff --git a/configs/stable-diffusion/v2-inpainting-inference.yaml b/configs/stable-diffusion/v2-inpainting-inference.yaml new file mode 100644 index 0000000000000000000000000000000000000000..32a9471d71b828c51bcbbabfe34c5f6c8282c803 --- /dev/null +++ b/configs/stable-diffusion/v2-inpainting-inference.yaml @@ -0,0 +1,158 @@ +model: + base_learning_rate: 5.0e-05 + target: ldm.models.diffusion.ddpm.LatentInpaintDiffusion + params: + linear_start: 0.00085 + linear_end: 0.0120 + num_timesteps_cond: 1 + log_every_t: 200 + timesteps: 1000 + first_stage_key: "jpg" + cond_stage_key: "txt" + image_size: 64 + channels: 4 + cond_stage_trainable: false + conditioning_key: hybrid + scale_factor: 0.18215 + monitor: val/loss_simple_ema + finetune_keys: null + use_ema: False + + unet_config: + target: ldm.modules.diffusionmodules.openaimodel.UNetModel + params: + use_checkpoint: True + image_size: 32 # unused + in_channels: 9 + out_channels: 4 + model_channels: 320 + attention_resolutions: [ 4, 2, 1 ] + num_res_blocks: 2 + channel_mult: [ 1, 2, 4, 4 ] + num_head_channels: 64 # need to fix for flash-attn + use_spatial_transformer: True + use_linear_in_transformer: True + transformer_depth: 1 + context_dim: 1024 + legacy: False + + first_stage_config: + target: ldm.models.autoencoder.AutoencoderKL + params: + embed_dim: 4 + monitor: val/rec_loss + ddconfig: + #attn_type: "vanilla-xformers" + double_z: true + z_channels: 4 + resolution: 256 + in_channels: 3 + out_ch: 3 + ch: 128 + ch_mult: + - 1 + - 2 + - 4 + - 4 + num_res_blocks: 2 + attn_resolutions: [ ] + dropout: 0.0 + lossconfig: + target: torch.nn.Identity + + cond_stage_config: + target: ldm.modules.encoders.modules.FrozenOpenCLIPEmbedder + params: + freeze: True + layer: "penultimate" + + +data: + target: ldm.data.laion.WebDataModuleFromConfig + params: + tar_base: null # for concat as in LAION-A + p_unsafe_threshold: 0.1 + filter_word_list: "data/filters.yaml" + max_pwatermark: 0.45 + batch_size: 8 + num_workers: 6 + multinode: True + min_size: 512 + train: + shards: + - "pipe:aws s3 cp s3://stability-aws/laion-a-native/part-0/{00000..18699}.tar -" + - "pipe:aws s3 cp s3://stability-aws/laion-a-native/part-1/{00000..18699}.tar -" + - "pipe:aws s3 cp s3://stability-aws/laion-a-native/part-2/{00000..18699}.tar -" + - "pipe:aws s3 cp s3://stability-aws/laion-a-native/part-3/{00000..18699}.tar -" + - "pipe:aws s3 cp s3://stability-aws/laion-a-native/part-4/{00000..18699}.tar -" #{00000-94333}.tar" + shuffle: 10000 + image_key: jpg + image_transforms: + - target: torchvision.transforms.Resize + params: + size: 512 + interpolation: 3 + - target: torchvision.transforms.RandomCrop + params: + size: 512 + postprocess: + target: ldm.data.laion.AddMask + params: + mode: "512train-large" + p_drop: 0.25 + # NOTE use enough shards to avoid empty validation loops in workers + validation: + shards: + - "pipe:aws s3 cp s3://deep-floyd-s3/datasets/laion_cleaned-part5/{93001..94333}.tar - " + shuffle: 0 + image_key: jpg + image_transforms: + - target: torchvision.transforms.Resize + params: + size: 512 + interpolation: 3 + - target: torchvision.transforms.CenterCrop + params: + size: 512 + postprocess: + target: ldm.data.laion.AddMask + params: + mode: "512train-large" + p_drop: 0.25 + +lightning: + find_unused_parameters: True + modelcheckpoint: + params: + every_n_train_steps: 5000 + + callbacks: + metrics_over_trainsteps_checkpoint: + params: + every_n_train_steps: 10000 + + image_logger: + target: main.ImageLogger + params: + enable_autocast: False + disabled: False + batch_frequency: 1000 + max_images: 4 + increase_log_steps: False + log_first_step: False + log_images_kwargs: + use_ema_scope: False + inpaint: False + plot_progressive_rows: False + plot_diffusion_rows: False + N: 4 + unconditional_guidance_scale: 5.0 + unconditional_guidance_label: [""] + ddim_steps: 50 # todo check these out for depth2img, + ddim_eta: 0.0 # todo check these out for depth2img, + + trainer: + benchmark: True + val_check_interval: 5000000 + num_sanity_val_steps: 0 + accumulate_grad_batches: 1 diff --git a/configs/stable-diffusion/v2-midas-inference.yaml b/configs/stable-diffusion/v2-midas-inference.yaml new file mode 100644 index 0000000000000000000000000000000000000000..f20c30f618b81091e31c2c4cf15325fa38638af4 --- /dev/null +++ b/configs/stable-diffusion/v2-midas-inference.yaml @@ -0,0 +1,74 @@ +model: + base_learning_rate: 5.0e-07 + target: ldm.models.diffusion.ddpm.LatentDepth2ImageDiffusion + params: + linear_start: 0.00085 + linear_end: 0.0120 + num_timesteps_cond: 1 + log_every_t: 200 + timesteps: 1000 + first_stage_key: "jpg" + cond_stage_key: "txt" + image_size: 64 + channels: 4 + cond_stage_trainable: false + conditioning_key: hybrid + scale_factor: 0.18215 + monitor: val/loss_simple_ema + finetune_keys: null + use_ema: False + + depth_stage_config: + target: ldm.modules.midas.api.MiDaSInference + params: + model_type: "dpt_hybrid" + + unet_config: + target: ldm.modules.diffusionmodules.openaimodel.UNetModel + params: + use_checkpoint: True + image_size: 32 # unused + in_channels: 5 + out_channels: 4 + model_channels: 320 + attention_resolutions: [ 4, 2, 1 ] + num_res_blocks: 2 + channel_mult: [ 1, 2, 4, 4 ] + num_head_channels: 64 # need to fix for flash-attn + use_spatial_transformer: True + use_linear_in_transformer: True + transformer_depth: 1 + context_dim: 1024 + legacy: False + + first_stage_config: + target: ldm.models.autoencoder.AutoencoderKL + params: + embed_dim: 4 + monitor: val/rec_loss + ddconfig: + #attn_type: "vanilla-xformers" + double_z: true + z_channels: 4 + resolution: 256 + in_channels: 3 + out_ch: 3 + ch: 128 + ch_mult: + - 1 + - 2 + - 4 + - 4 + num_res_blocks: 2 + attn_resolutions: [ ] + dropout: 0.0 + lossconfig: + target: torch.nn.Identity + + cond_stage_config: + target: ldm.modules.encoders.modules.FrozenOpenCLIPEmbedder + params: + freeze: True + layer: "penultimate" + + diff --git a/configs/stable-diffusion/x4-upscaling.yaml b/configs/stable-diffusion/x4-upscaling.yaml new file mode 100644 index 0000000000000000000000000000000000000000..2db0964af699f86d1891c761710a2d53f59b842c --- /dev/null +++ b/configs/stable-diffusion/x4-upscaling.yaml @@ -0,0 +1,76 @@ +model: + base_learning_rate: 1.0e-04 + target: ldm.models.diffusion.ddpm.LatentUpscaleDiffusion + params: + parameterization: "v" + low_scale_key: "lr" + linear_start: 0.0001 + linear_end: 0.02 + num_timesteps_cond: 1 + log_every_t: 200 + timesteps: 1000 + first_stage_key: "jpg" + cond_stage_key: "txt" + image_size: 128 + channels: 4 + cond_stage_trainable: false + conditioning_key: "hybrid-adm" + monitor: val/loss_simple_ema + scale_factor: 0.08333 + use_ema: False + + low_scale_config: + target: ldm.modules.diffusionmodules.upscaling.ImageConcatWithNoiseAugmentation + params: + noise_schedule_config: # image space + linear_start: 0.0001 + linear_end: 0.02 + max_noise_level: 350 + + unet_config: + target: ldm.modules.diffusionmodules.openaimodel.UNetModel + params: + use_checkpoint: True + num_classes: 1000 # timesteps for noise conditioning (here constant, just need one) + image_size: 128 + in_channels: 7 + out_channels: 4 + model_channels: 256 + attention_resolutions: [ 2,4,8] + num_res_blocks: 2 + channel_mult: [ 1, 2, 2, 4] + disable_self_attentions: [True, True, True, False] + disable_middle_self_attn: False + num_heads: 8 + use_spatial_transformer: True + transformer_depth: 1 + context_dim: 1024 + legacy: False + use_linear_in_transformer: True + + first_stage_config: + target: ldm.models.autoencoder.AutoencoderKL + params: + embed_dim: 4 + ddconfig: + # attn_type: "vanilla-xformers" this model needs efficient attention to be feasible on HR data, also the decoder seems to break in half precision (UNet is fine though) + double_z: True + z_channels: 4 + resolution: 256 + in_channels: 3 + out_ch: 3 + ch: 128 + ch_mult: [ 1,2,4 ] # num_down = len(ch_mult)-1 + num_res_blocks: 2 + attn_resolutions: [ ] + dropout: 0.0 + + lossconfig: + target: torch.nn.Identity + + cond_stage_config: + target: ldm.modules.encoders.modules.FrozenOpenCLIPEmbedder + params: + freeze: True + layer: "penultimate" + diff --git a/doc/UNCLIP.MD b/doc/UNCLIP.MD new file mode 100644 index 0000000000000000000000000000000000000000..989533b99599725cc9d61ae6055fa97f9a885f23 --- /dev/null +++ b/doc/UNCLIP.MD @@ -0,0 +1,88 @@ +### Stable unCLIP + +[unCLIP](https://openai.com/dall-e-2/) is the approach behind OpenAI's [DALL·E 2](https://openai.com/dall-e-2/), +trained to invert CLIP image embeddings. +We finetuned SD 2.1 to accept a CLIP ViT-L/14 image embedding in addition to the text encodings. +This means that the model can be used to produce image variations, but can also be combined with a text-to-image +embedding prior to yield a full text-to-image model at 768x768 resolution. + +If you would like to try a demo of this model on the web, please visit https://clipdrop.co/stable-diffusion-reimagine + +We provide two models, trained on OpenAI CLIP-L and OpenCLIP-H image embeddings, respectively, +available from [https://huggingface.co/stabilityai/stable-diffusion-2-1-unclip](https://huggingface.co/stabilityai/stable-diffusion-2-1-unclip/tree/main). +To use them, download from Hugging Face, and put and the weights into the `checkpoints` folder. + +#### Image Variations +![image-variations-l-1](../assets/stable-samples/stable-unclip/unclip-variations.png) + +Diffusers integration +Stable UnCLIP Image Variations is integrated with the [🧨 diffusers](https://github.com/huggingface/diffusers) library +```python +#pip install git+https://github.com/huggingface/diffusers.git transformers accelerate +import requests +import torch +from PIL import Image +from io import BytesIO + +from diffusers import StableUnCLIPImg2ImgPipeline + +#Start the StableUnCLIP Image variations pipeline +pipe = StableUnCLIPImg2ImgPipeline.from_pretrained( + "stabilityai/stable-diffusion-2-1-unclip", torch_dtype=torch.float16, variation="fp16" +) +pipe = pipe.to("cuda") + +#Get image from URL +url = "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/stable_unclip/tarsila_do_amaral.png" +response = requests.get(url) +init_image = Image.open(BytesIO(response.content)).convert("RGB") + +#Pipe to make the variation +images = pipe(init_image).images +images[0].save("tarsila_variation.png") +``` +Check out the [Stable UnCLIP pipeline docs here](https://huggingface.co/docs/diffusers/main/en/api/pipelines/stable_unclip) + +Streamlit UI demo + +``` +streamlit run scripts/streamlit/stableunclip.py +``` +to launch a streamlit script than can be used to make image variations with both models (CLIP-L and OpenCLIP-H). +These models can process a `noise_level`, which specifies an amount of Gaussian noise added to the CLIP embeddings. +This can be used to increase output variance as in the following examples. + +![image-variations-noise](../assets/stable-samples/stable-unclip/unclip-variations_noise.png) + + +### Stable Diffusion Meets Karlo +![panda](../assets/stable-samples/stable-unclip/panda.jpg) + +Recently, [KakaoBrain](https://kakaobrain.com/) openly released [Karlo](https://github.com/kakaobrain/karlo), a pretrained, large-scale replication of [unCLIP](https://arxiv.org/abs/2204.06125). +We introduce _Stable Karlo_, a combination of the Karlo CLIP image embedding prior, and Stable Diffusion v2.1-768. + +To run the model, first download the KARLO checkpoints +```shell +mkdir -p checkpoints/karlo_models +cd checkpoints/karlo_models +wget https://arena.kakaocdn.net/brainrepo/models/karlo-public/v1.0.0.alpha/096db1af569b284eb76b3881534822d9/ViT-L-14.pt +wget https://arena.kakaocdn.net/brainrepo/models/karlo-public/v1.0.0.alpha/0b62380a75e56f073e2844ab5199153d/ViT-L-14_stats.th +wget https://arena.kakaocdn.net/brainrepo/models/karlo-public/v1.0.0.alpha/85626483eaca9f581e2a78d31ff905ca/prior-ckpt-step%3D01000000-of-01000000.ckpt +cd ../../ +``` +and the finetuned SD2.1 unCLIP-L checkpoint from [here](https://huggingface.co/stabilityai/stable-diffusion-2-1-unclip/blob/main/sd21-unclip-l.ckpt), and put the ckpt into the `checkpoints folder` + +Then, run + +``` +streamlit run scripts/streamlit/stableunclip.py +``` +and pick the `use_karlo` option in the GUI. +The script optionally supports sampling from the full Karlo model. To use it, download the 64x64 decoder and 64->256 upscaler +via +```shell +cd checkpoints/karlo_models +wget https://arena.kakaocdn.net/brainrepo/models/karlo-public/v1.0.0.alpha/efdf6206d8ed593961593dc029a8affa/decoder-ckpt-step%3D01000000-of-01000000.ckpt +wget https://arena.kakaocdn.net/brainrepo/models/karlo-public/v1.0.0.alpha/4226b831ae0279020d134281f3c31590/improved-sr-ckpt-step%3D1.2M.ckpt +cd ../../ +``` diff --git a/environment.yaml b/environment.yaml new file mode 100644 index 0000000000000000000000000000000000000000..4687b309b60ae2d6040fcb3f90a380cf6fb11b21 --- /dev/null +++ b/environment.yaml @@ -0,0 +1,29 @@ +name: ldm +channels: + - pytorch + - defaults +dependencies: + - python=3.8.5 + - pip=20.3 + - cudatoolkit=11.3 + - pytorch=1.12.1 + - torchvision=0.13.1 + - numpy=1.23.1 + - pip: + - albumentations==1.3.0 + - opencv-python==4.6.0.66 + - imageio==2.9.0 + - imageio-ffmpeg==0.4.2 + - pytorch-lightning==1.4.2 + - omegaconf==2.1.1 + - test-tube>=0.7.5 + - streamlit==1.12.1 + - einops==0.3.0 + - transformers==4.19.2 + - webdataset==0.2.5 + - kornia==0.6 + - open_clip_torch==2.0.2 + - invisible-watermark>=0.1.5 + - streamlit-drawable-canvas==0.8.0 + - torchmetrics==0.6.0 + - -e . diff --git a/ldm/data/__init__.py b/ldm/data/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/ldm/data/util.py b/ldm/data/util.py new file mode 100644 index 0000000000000000000000000000000000000000..5b60ceb2349e3bd7900ff325740e2022d2903b1c --- /dev/null +++ b/ldm/data/util.py @@ -0,0 +1,24 @@ +import torch + +from ldm.modules.midas.api import load_midas_transform + + +class AddMiDaS(object): + def __init__(self, model_type): + super().__init__() + self.transform = load_midas_transform(model_type) + + def pt2np(self, x): + x = ((x + 1.0) * .5).detach().cpu().numpy() + return x + + def np2pt(self, x): + x = torch.from_numpy(x) * 2 - 1. + return x + + def __call__(self, sample): + # sample['jpg'] is tensor hwc in [-1, 1] at this point + x = self.pt2np(sample['jpg']) + x = self.transform({"image": x})["image"] + sample['midas_in'] = x + return sample \ No newline at end of file diff --git a/ldm/models/autoencoder.py b/ldm/models/autoencoder.py new file mode 100644 index 0000000000000000000000000000000000000000..d122549995ce2cd64092c81a58419ed4a15a02fd --- /dev/null +++ b/ldm/models/autoencoder.py @@ -0,0 +1,219 @@ +import torch +import pytorch_lightning as pl +import torch.nn.functional as F +from contextlib import contextmanager + +from ldm.modules.diffusionmodules.model import Encoder, Decoder +from ldm.modules.distributions.distributions import DiagonalGaussianDistribution + +from ldm.util import instantiate_from_config +from ldm.modules.ema import LitEma + + +class AutoencoderKL(pl.LightningModule): + def __init__(self, + ddconfig, + lossconfig, + embed_dim, + ckpt_path=None, + ignore_keys=[], + image_key="image", + colorize_nlabels=None, + monitor=None, + ema_decay=None, + learn_logvar=False + ): + super().__init__() + self.learn_logvar = learn_logvar + self.image_key = image_key + self.encoder = Encoder(**ddconfig) + self.decoder = Decoder(**ddconfig) + self.loss = instantiate_from_config(lossconfig) + assert ddconfig["double_z"] + self.quant_conv = torch.nn.Conv2d(2*ddconfig["z_channels"], 2*embed_dim, 1) + self.post_quant_conv = torch.nn.Conv2d(embed_dim, ddconfig["z_channels"], 1) + self.embed_dim = embed_dim + if colorize_nlabels is not None: + assert type(colorize_nlabels)==int + self.register_buffer("colorize", torch.randn(3, colorize_nlabels, 1, 1)) + if monitor is not None: + self.monitor = monitor + + self.use_ema = ema_decay is not None + if self.use_ema: + self.ema_decay = ema_decay + assert 0. < ema_decay < 1. + self.model_ema = LitEma(self, decay=ema_decay) + print(f"Keeping EMAs of {len(list(self.model_ema.buffers()))}.") + + if ckpt_path is not None: + self.init_from_ckpt(ckpt_path, ignore_keys=ignore_keys) + + def init_from_ckpt(self, path, ignore_keys=list()): + sd = torch.load(path, map_location="cpu")["state_dict"] + keys = list(sd.keys()) + for k in keys: + for ik in ignore_keys: + if k.startswith(ik): + print("Deleting key {} from state_dict.".format(k)) + del sd[k] + self.load_state_dict(sd, strict=False) + print(f"Restored from {path}") + + @contextmanager + def ema_scope(self, context=None): + if self.use_ema: + self.model_ema.store(self.parameters()) + self.model_ema.copy_to(self) + if context is not None: + print(f"{context}: Switched to EMA weights") + try: + yield None + finally: + if self.use_ema: + self.model_ema.restore(self.parameters()) + if context is not None: + print(f"{context}: Restored training weights") + + def on_train_batch_end(self, *args, **kwargs): + if self.use_ema: + self.model_ema(self) + + def encode(self, x): + h = self.encoder(x) + moments = self.quant_conv(h) + posterior = DiagonalGaussianDistribution(moments) + return posterior + + def decode(self, z): + z = self.post_quant_conv(z) + dec = self.decoder(z) + return dec + + def forward(self, input, sample_posterior=True): + posterior = self.encode(input) + if sample_posterior: + z = posterior.sample() + else: + z = posterior.mode() + dec = self.decode(z) + return dec, posterior + + def get_input(self, batch, k): + x = batch[k] + if len(x.shape) == 3: + x = x[..., None] + x = x.permute(0, 3, 1, 2).to(memory_format=torch.contiguous_format).float() + return x + + def training_step(self, batch, batch_idx, optimizer_idx): + inputs = self.get_input(batch, self.image_key) + reconstructions, posterior = self(inputs) + + if optimizer_idx == 0: + # train encoder+decoder+logvar + aeloss, log_dict_ae = self.loss(inputs, reconstructions, posterior, optimizer_idx, self.global_step, + last_layer=self.get_last_layer(), split="train") + self.log("aeloss", aeloss, prog_bar=True, logger=True, on_step=True, on_epoch=True) + self.log_dict(log_dict_ae, prog_bar=False, logger=True, on_step=True, on_epoch=False) + return aeloss + + if optimizer_idx == 1: + # train the discriminator + discloss, log_dict_disc = self.loss(inputs, reconstructions, posterior, optimizer_idx, self.global_step, + last_layer=self.get_last_layer(), split="train") + + self.log("discloss", discloss, prog_bar=True, logger=True, on_step=True, on_epoch=True) + self.log_dict(log_dict_disc, prog_bar=False, logger=True, on_step=True, on_epoch=False) + return discloss + + def validation_step(self, batch, batch_idx): + log_dict = self._validation_step(batch, batch_idx) + with self.ema_scope(): + log_dict_ema = self._validation_step(batch, batch_idx, postfix="_ema") + return log_dict + + def _validation_step(self, batch, batch_idx, postfix=""): + inputs = self.get_input(batch, self.image_key) + reconstructions, posterior = self(inputs) + aeloss, log_dict_ae = self.loss(inputs, reconstructions, posterior, 0, self.global_step, + last_layer=self.get_last_layer(), split="val"+postfix) + + discloss, log_dict_disc = self.loss(inputs, reconstructions, posterior, 1, self.global_step, + last_layer=self.get_last_layer(), split="val"+postfix) + + self.log(f"val{postfix}/rec_loss", log_dict_ae[f"val{postfix}/rec_loss"]) + self.log_dict(log_dict_ae) + self.log_dict(log_dict_disc) + return self.log_dict + + def configure_optimizers(self): + lr = self.learning_rate + ae_params_list = list(self.encoder.parameters()) + list(self.decoder.parameters()) + list( + self.quant_conv.parameters()) + list(self.post_quant_conv.parameters()) + if self.learn_logvar: + print(f"{self.__class__.__name__}: Learning logvar") + ae_params_list.append(self.loss.logvar) + opt_ae = torch.optim.Adam(ae_params_list, + lr=lr, betas=(0.5, 0.9)) + opt_disc = torch.optim.Adam(self.loss.discriminator.parameters(), + lr=lr, betas=(0.5, 0.9)) + return [opt_ae, opt_disc], [] + + def get_last_layer(self): + return self.decoder.conv_out.weight + + @torch.no_grad() + def log_images(self, batch, only_inputs=False, log_ema=False, **kwargs): + log = dict() + x = self.get_input(batch, self.image_key) + x = x.to(self.device) + if not only_inputs: + xrec, posterior = self(x) + if x.shape[1] > 3: + # colorize with random projection + assert xrec.shape[1] > 3 + x = self.to_rgb(x) + xrec = self.to_rgb(xrec) + log["samples"] = self.decode(torch.randn_like(posterior.sample())) + log["reconstructions"] = xrec + if log_ema or self.use_ema: + with self.ema_scope(): + xrec_ema, posterior_ema = self(x) + if x.shape[1] > 3: + # colorize with random projection + assert xrec_ema.shape[1] > 3 + xrec_ema = self.to_rgb(xrec_ema) + log["samples_ema"] = self.decode(torch.randn_like(posterior_ema.sample())) + log["reconstructions_ema"] = xrec_ema + log["inputs"] = x + return log + + def to_rgb(self, x): + assert self.image_key == "segmentation" + if not hasattr(self, "colorize"): + self.register_buffer("colorize", torch.randn(3, x.shape[1], 1, 1).to(x)) + x = F.conv2d(x, weight=self.colorize) + x = 2.*(x-x.min())/(x.max()-x.min()) - 1. + return x + + +class IdentityFirstStage(torch.nn.Module): + def __init__(self, *args, vq_interface=False, **kwargs): + self.vq_interface = vq_interface + super().__init__() + + def encode(self, x, *args, **kwargs): + return x + + def decode(self, x, *args, **kwargs): + return x + + def quantize(self, x, *args, **kwargs): + if self.vq_interface: + return x, None, [None, None, None] + return x + + def forward(self, x, *args, **kwargs): + return x + diff --git a/ldm/models/diffusion/__init__.py b/ldm/models/diffusion/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/ldm/models/diffusion/ddim.py b/ldm/models/diffusion/ddim.py new file mode 100644 index 0000000000000000000000000000000000000000..c6cfd5712281a72c086e592ed57fc0730c5a5a3b --- /dev/null +++ b/ldm/models/diffusion/ddim.py @@ -0,0 +1,337 @@ +"""SAMPLING ONLY.""" + +import torch +import numpy as np +from tqdm import tqdm + +from ldm.modules.diffusionmodules.util import make_ddim_sampling_parameters, make_ddim_timesteps, noise_like, extract_into_tensor + + +class DDIMSampler(object): + def __init__(self, model, schedule="linear", device=torch.device("cuda"), **kwargs): + super().__init__() + self.model = model + self.ddpm_num_timesteps = model.num_timesteps + self.schedule = schedule + self.device = device + + def register_buffer(self, name, attr): + if type(attr) == torch.Tensor: + if attr.device != self.device: + attr = attr.to(self.device) + setattr(self, name, attr) + + def make_schedule(self, ddim_num_steps, ddim_discretize="uniform", ddim_eta=0., verbose=True): + self.ddim_timesteps = make_ddim_timesteps(ddim_discr_method=ddim_discretize, num_ddim_timesteps=ddim_num_steps, + num_ddpm_timesteps=self.ddpm_num_timesteps,verbose=verbose) + alphas_cumprod = self.model.alphas_cumprod + assert alphas_cumprod.shape[0] == self.ddpm_num_timesteps, 'alphas have to be defined for each timestep' + to_torch = lambda x: x.clone().detach().to(torch.float32).to(self.model.device) + + self.register_buffer('betas', to_torch(self.model.betas)) + self.register_buffer('alphas_cumprod', to_torch(alphas_cumprod)) + self.register_buffer('alphas_cumprod_prev', to_torch(self.model.alphas_cumprod_prev)) + + # calculations for diffusion q(x_t | x_{t-1}) and others + self.register_buffer('sqrt_alphas_cumprod', to_torch(np.sqrt(alphas_cumprod.cpu()))) + self.register_buffer('sqrt_one_minus_alphas_cumprod', to_torch(np.sqrt(1. - alphas_cumprod.cpu()))) + self.register_buffer('log_one_minus_alphas_cumprod', to_torch(np.log(1. - alphas_cumprod.cpu()))) + self.register_buffer('sqrt_recip_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod.cpu()))) + self.register_buffer('sqrt_recipm1_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod.cpu() - 1))) + + # ddim sampling parameters + ddim_sigmas, ddim_alphas, ddim_alphas_prev = make_ddim_sampling_parameters(alphacums=alphas_cumprod.cpu(), + ddim_timesteps=self.ddim_timesteps, + eta=ddim_eta,verbose=verbose) + self.register_buffer('ddim_sigmas', ddim_sigmas) + self.register_buffer('ddim_alphas', ddim_alphas) + self.register_buffer('ddim_alphas_prev', ddim_alphas_prev) + self.register_buffer('ddim_sqrt_one_minus_alphas', np.sqrt(1. - ddim_alphas)) + sigmas_for_original_sampling_steps = ddim_eta * torch.sqrt( + (1 - self.alphas_cumprod_prev) / (1 - self.alphas_cumprod) * ( + 1 - self.alphas_cumprod / self.alphas_cumprod_prev)) + self.register_buffer('ddim_sigmas_for_original_num_steps', sigmas_for_original_sampling_steps) + + @torch.no_grad() + def sample(self, + S, + batch_size, + shape, + conditioning=None, + callback=None, + normals_sequence=None, + img_callback=None, + quantize_x0=False, + eta=0., + mask=None, + x0=None, + temperature=1., + noise_dropout=0., + score_corrector=None, + corrector_kwargs=None, + verbose=True, + x_T=None, + log_every_t=100, + unconditional_guidance_scale=1., + unconditional_conditioning=None, # this has to come in the same format as the conditioning, # e.g. as encoded tokens, ... + dynamic_threshold=None, + ucg_schedule=None, + **kwargs + ): + if conditioning is not None: + if isinstance(conditioning, dict): + ctmp = conditioning[list(conditioning.keys())[0]] + while isinstance(ctmp, list): ctmp = ctmp[0] + cbs = ctmp.shape[0] + if cbs != batch_size: + print(f"Warning: Got {cbs} conditionings but batch-size is {batch_size}") + + elif isinstance(conditioning, list): + for ctmp in conditioning: + if ctmp.shape[0] != batch_size: + print(f"Warning: Got {cbs} conditionings but batch-size is {batch_size}") + + else: + if conditioning.shape[0] != batch_size: + print(f"Warning: Got {conditioning.shape[0]} conditionings but batch-size is {batch_size}") + + self.make_schedule(ddim_num_steps=S, ddim_eta=eta, verbose=verbose) + # sampling + C, H, W = shape + size = (batch_size, C, H, W) + print(f'Data shape for DDIM sampling is {size}, eta {eta}') + + samples, intermediates = self.ddim_sampling(conditioning, size, + callback=callback, + img_callback=img_callback, + quantize_denoised=quantize_x0, + mask=mask, x0=x0, + ddim_use_original_steps=False, + noise_dropout=noise_dropout, + temperature=temperature, + score_corrector=score_corrector, + corrector_kwargs=corrector_kwargs, + x_T=x_T, + log_every_t=log_every_t, + unconditional_guidance_scale=unconditional_guidance_scale, + unconditional_conditioning=unconditional_conditioning, + dynamic_threshold=dynamic_threshold, + ucg_schedule=ucg_schedule + ) + return samples, intermediates + + @torch.no_grad() + def ddim_sampling(self, cond, shape, + x_T=None, ddim_use_original_steps=False, + callback=None, timesteps=None, quantize_denoised=False, + mask=None, x0=None, img_callback=None, log_every_t=100, + temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None, + unconditional_guidance_scale=1., unconditional_conditioning=None, dynamic_threshold=None, + ucg_schedule=None): + device = self.model.betas.device + b = shape[0] + if x_T is None: + img = torch.randn(shape, device=device) + else: + img = x_T + + if timesteps is None: + timesteps = self.ddpm_num_timesteps if ddim_use_original_steps else self.ddim_timesteps + elif timesteps is not None and not ddim_use_original_steps: + subset_end = int(min(timesteps / self.ddim_timesteps.shape[0], 1) * self.ddim_timesteps.shape[0]) - 1 + timesteps = self.ddim_timesteps[:subset_end] + + intermediates = {'x_inter': [img], 'pred_x0': [img]} + time_range = reversed(range(0,timesteps)) if ddim_use_original_steps else np.flip(timesteps) + total_steps = timesteps if ddim_use_original_steps else timesteps.shape[0] + print(f"Running DDIM Sampling with {total_steps} timesteps") + + iterator = tqdm(time_range, desc='DDIM Sampler', total=total_steps) + + for i, step in enumerate(iterator): + index = total_steps - i - 1 + ts = torch.full((b,), step, device=device, dtype=torch.long) + + if mask is not None: + assert x0 is not None + img_orig = self.model.q_sample(x0, ts) # TODO: deterministic forward pass? + img = img_orig * mask + (1. - mask) * img + + if ucg_schedule is not None: + assert len(ucg_schedule) == len(time_range) + unconditional_guidance_scale = ucg_schedule[i] + + outs = self.p_sample_ddim(img, cond, ts, index=index, use_original_steps=ddim_use_original_steps, + quantize_denoised=quantize_denoised, temperature=temperature, + noise_dropout=noise_dropout, score_corrector=score_corrector, + corrector_kwargs=corrector_kwargs, + unconditional_guidance_scale=unconditional_guidance_scale, + unconditional_conditioning=unconditional_conditioning, + dynamic_threshold=dynamic_threshold) + img, pred_x0 = outs + if callback: callback(i) + if img_callback: img_callback(pred_x0, i) + + if index % log_every_t == 0 or index == total_steps - 1: + intermediates['x_inter'].append(img) + intermediates['pred_x0'].append(pred_x0) + + return img, intermediates + + @torch.no_grad() + def p_sample_ddim(self, x, c, t, index, repeat_noise=False, use_original_steps=False, quantize_denoised=False, + temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None, + unconditional_guidance_scale=1., unconditional_conditioning=None, + dynamic_threshold=None): + b, *_, device = *x.shape, x.device + + if unconditional_conditioning is None or unconditional_guidance_scale == 1.: + model_output = self.model.apply_model(x, t, c) + else: + x_in = torch.cat([x] * 2) + t_in = torch.cat([t] * 2) + if isinstance(c, dict): + assert isinstance(unconditional_conditioning, dict) + c_in = dict() + for k in c: + if isinstance(c[k], list): + c_in[k] = [torch.cat([ + unconditional_conditioning[k][i], + c[k][i]]) for i in range(len(c[k]))] + else: + c_in[k] = torch.cat([ + unconditional_conditioning[k], + c[k]]) + elif isinstance(c, list): + c_in = list() + assert isinstance(unconditional_conditioning, list) + for i in range(len(c)): + c_in.append(torch.cat([unconditional_conditioning[i], c[i]])) + else: + c_in = torch.cat([unconditional_conditioning, c]) + model_uncond, model_t = self.model.apply_model(x_in, t_in, c_in).chunk(2) + model_output = model_uncond + unconditional_guidance_scale * (model_t - model_uncond) + + if self.model.parameterization == "v": + e_t = self.model.predict_eps_from_z_and_v(x, t, model_output) + else: + e_t = model_output + + if score_corrector is not None: + assert self.model.parameterization == "eps", 'not implemented' + e_t = score_corrector.modify_score(self.model, e_t, x, t, c, **corrector_kwargs) + + alphas = self.model.alphas_cumprod if use_original_steps else self.ddim_alphas + alphas_prev = self.model.alphas_cumprod_prev if use_original_steps else self.ddim_alphas_prev + sqrt_one_minus_alphas = self.model.sqrt_one_minus_alphas_cumprod if use_original_steps else self.ddim_sqrt_one_minus_alphas + sigmas = self.model.ddim_sigmas_for_original_num_steps if use_original_steps else self.ddim_sigmas + # select parameters corresponding to the currently considered timestep + a_t = torch.full((b, 1, 1, 1), alphas[index], device=device) + a_prev = torch.full((b, 1, 1, 1), alphas_prev[index], device=device) + sigma_t = torch.full((b, 1, 1, 1), sigmas[index], device=device) + sqrt_one_minus_at = torch.full((b, 1, 1, 1), sqrt_one_minus_alphas[index],device=device) + + # current prediction for x_0 + if self.model.parameterization != "v": + pred_x0 = (x - sqrt_one_minus_at * e_t) / a_t.sqrt() + else: + pred_x0 = self.model.predict_start_from_z_and_v(x, t, model_output) + + if quantize_denoised: + pred_x0, _, *_ = self.model.first_stage_model.quantize(pred_x0) + + if dynamic_threshold is not None: + raise NotImplementedError() + + # direction pointing to x_t + dir_xt = (1. - a_prev - sigma_t**2).sqrt() * e_t + noise = sigma_t * noise_like(x.shape, device, repeat_noise) * temperature + if noise_dropout > 0.: + noise = torch.nn.functional.dropout(noise, p=noise_dropout) + x_prev = a_prev.sqrt() * pred_x0 + dir_xt + noise + return x_prev, pred_x0 + + @torch.no_grad() + def encode(self, x0, c, t_enc, use_original_steps=False, return_intermediates=None, + unconditional_guidance_scale=1.0, unconditional_conditioning=None, callback=None): + num_reference_steps = self.ddpm_num_timesteps if use_original_steps else self.ddim_timesteps.shape[0] + + assert t_enc <= num_reference_steps + num_steps = t_enc + + if use_original_steps: + alphas_next = self.alphas_cumprod[:num_steps] + alphas = self.alphas_cumprod_prev[:num_steps] + else: + alphas_next = self.ddim_alphas[:num_steps] + alphas = torch.tensor(self.ddim_alphas_prev[:num_steps]) + + x_next = x0 + intermediates = [] + inter_steps = [] + for i in tqdm(range(num_steps), desc='Encoding Image'): + t = torch.full((x0.shape[0],), i, device=self.model.device, dtype=torch.long) + if unconditional_guidance_scale == 1.: + noise_pred = self.model.apply_model(x_next, t, c) + else: + assert unconditional_conditioning is not None + e_t_uncond, noise_pred = torch.chunk( + self.model.apply_model(torch.cat((x_next, x_next)), torch.cat((t, t)), + torch.cat((unconditional_conditioning, c))), 2) + noise_pred = e_t_uncond + unconditional_guidance_scale * (noise_pred - e_t_uncond) + + xt_weighted = (alphas_next[i] / alphas[i]).sqrt() * x_next + weighted_noise_pred = alphas_next[i].sqrt() * ( + (1 / alphas_next[i] - 1).sqrt() - (1 / alphas[i] - 1).sqrt()) * noise_pred + x_next = xt_weighted + weighted_noise_pred + if return_intermediates and i % ( + num_steps // return_intermediates) == 0 and i < num_steps - 1: + intermediates.append(x_next) + inter_steps.append(i) + elif return_intermediates and i >= num_steps - 2: + intermediates.append(x_next) + inter_steps.append(i) + if callback: callback(i) + + out = {'x_encoded': x_next, 'intermediate_steps': inter_steps} + if return_intermediates: + out.update({'intermediates': intermediates}) + return x_next, out + + @torch.no_grad() + def stochastic_encode(self, x0, t, use_original_steps=False, noise=None): + # fast, but does not allow for exact reconstruction + # t serves as an index to gather the correct alphas + if use_original_steps: + sqrt_alphas_cumprod = self.sqrt_alphas_cumprod + sqrt_one_minus_alphas_cumprod = self.sqrt_one_minus_alphas_cumprod + else: + sqrt_alphas_cumprod = torch.sqrt(self.ddim_alphas) + sqrt_one_minus_alphas_cumprod = self.ddim_sqrt_one_minus_alphas + + if noise is None: + noise = torch.randn_like(x0) + return (extract_into_tensor(sqrt_alphas_cumprod, t, x0.shape) * x0 + + extract_into_tensor(sqrt_one_minus_alphas_cumprod, t, x0.shape) * noise) + + @torch.no_grad() + def decode(self, x_latent, cond, t_start, unconditional_guidance_scale=1.0, unconditional_conditioning=None, + use_original_steps=False, callback=None): + + timesteps = np.arange(self.ddpm_num_timesteps) if use_original_steps else self.ddim_timesteps + timesteps = timesteps[:t_start] + + time_range = np.flip(timesteps) + total_steps = timesteps.shape[0] + print(f"Running DDIM Sampling with {total_steps} timesteps") + + iterator = tqdm(time_range, desc='Decoding image', total=total_steps) + x_dec = x_latent + for i, step in enumerate(iterator): + index = total_steps - i - 1 + ts = torch.full((x_latent.shape[0],), step, device=x_latent.device, dtype=torch.long) + x_dec, _ = self.p_sample_ddim(x_dec, cond, ts, index=index, use_original_steps=use_original_steps, + unconditional_guidance_scale=unconditional_guidance_scale, + unconditional_conditioning=unconditional_conditioning) + if callback: callback(i) + return x_dec \ No newline at end of file diff --git a/ldm/models/diffusion/ddpm.py b/ldm/models/diffusion/ddpm.py new file mode 100644 index 0000000000000000000000000000000000000000..3350c032f40b9b04e0b652aa296ad4ae4b2bd9ff --- /dev/null +++ b/ldm/models/diffusion/ddpm.py @@ -0,0 +1,1873 @@ +""" +wild mixture of +https://github.com/lucidrains/denoising-diffusion-pytorch/blob/7706bdfc6f527f58d33f84b7b522e61e6e3164b3/denoising_diffusion_pytorch/denoising_diffusion_pytorch.py +https://github.com/openai/improved-diffusion/blob/e94489283bb876ac1477d5dd7709bbbd2d9902ce/improved_diffusion/gaussian_diffusion.py +https://github.com/CompVis/taming-transformers +-- merci +""" + +import torch +import torch.nn as nn +import numpy as np +import pytorch_lightning as pl +from torch.optim.lr_scheduler import LambdaLR +from einops import rearrange, repeat +from contextlib import contextmanager, nullcontext +from functools import partial +import itertools +from tqdm import tqdm +from torchvision.utils import make_grid +from pytorch_lightning.utilities.distributed import rank_zero_only +from omegaconf import ListConfig + +from ldm.util import log_txt_as_img, exists, default, ismap, isimage, mean_flat, count_params, instantiate_from_config +from ldm.modules.ema import LitEma +from ldm.modules.distributions.distributions import normal_kl, DiagonalGaussianDistribution +from ldm.models.autoencoder import IdentityFirstStage, AutoencoderKL +from ldm.modules.diffusionmodules.util import make_beta_schedule, extract_into_tensor, noise_like +from ldm.models.diffusion.ddim import DDIMSampler + + +__conditioning_keys__ = {'concat': 'c_concat', + 'crossattn': 'c_crossattn', + 'adm': 'y'} + + +def disabled_train(self, mode=True): + """Overwrite model.train with this function to make sure train/eval mode + does not change anymore.""" + return self + + +def uniform_on_device(r1, r2, shape, device): + return (r1 - r2) * torch.rand(*shape, device=device) + r2 + + +class DDPM(pl.LightningModule): + # classic DDPM with Gaussian diffusion, in image space + def __init__(self, + unet_config, + timesteps=1000, + beta_schedule="linear", + loss_type="l2", + ckpt_path=None, + ignore_keys=[], + load_only_unet=False, + monitor="val/loss", + use_ema=True, + first_stage_key="image", + image_size=256, + channels=3, + log_every_t=100, + clip_denoised=True, + linear_start=1e-4, + linear_end=2e-2, + cosine_s=8e-3, + given_betas=None, + original_elbo_weight=0., + v_posterior=0., # weight for choosing posterior variance as sigma = (1-v) * beta_tilde + v * beta + l_simple_weight=1., + conditioning_key=None, + parameterization="eps", # all assuming fixed variance schedules + scheduler_config=None, + use_positional_encodings=False, + learn_logvar=False, + logvar_init=0., + make_it_fit=False, + ucg_training=None, + reset_ema=False, + reset_num_ema_updates=False, + ): + super().__init__() + assert parameterization in ["eps", "x0", "v"], 'currently only supporting "eps" and "x0" and "v"' + self.parameterization = parameterization + print(f"{self.__class__.__name__}: Running in {self.parameterization}-prediction mode") + self.cond_stage_model = None + self.clip_denoised = clip_denoised + self.log_every_t = log_every_t + self.first_stage_key = first_stage_key + self.image_size = image_size # try conv? + self.channels = channels + self.use_positional_encodings = use_positional_encodings + self.model = DiffusionWrapper(unet_config, conditioning_key) + count_params(self.model, verbose=True) + self.use_ema = use_ema + if self.use_ema: + self.model_ema = LitEma(self.model) + print(f"Keeping EMAs of {len(list(self.model_ema.buffers()))}.") + + self.use_scheduler = scheduler_config is not None + if self.use_scheduler: + self.scheduler_config = scheduler_config + + self.v_posterior = v_posterior + self.original_elbo_weight = original_elbo_weight + self.l_simple_weight = l_simple_weight + + if monitor is not None: + self.monitor = monitor + self.make_it_fit = make_it_fit + if reset_ema: assert exists(ckpt_path) + if ckpt_path is not None: + self.init_from_ckpt(ckpt_path, ignore_keys=ignore_keys, only_model=load_only_unet) + if reset_ema: + assert self.use_ema + print(f"Resetting ema to pure model weights. This is useful when restoring from an ema-only checkpoint.") + self.model_ema = LitEma(self.model) + if reset_num_ema_updates: + print(" +++++++++++ WARNING: RESETTING NUM_EMA UPDATES TO ZERO +++++++++++ ") + assert self.use_ema + self.model_ema.reset_num_updates() + + self.register_schedule(given_betas=given_betas, beta_schedule=beta_schedule, timesteps=timesteps, + linear_start=linear_start, linear_end=linear_end, cosine_s=cosine_s) + + self.loss_type = loss_type + + self.learn_logvar = learn_logvar + self.logvar = torch.full(fill_value=logvar_init, size=(self.num_timesteps,)) + if self.learn_logvar: + self.logvar = nn.Parameter(self.logvar, requires_grad=True) + + self.ucg_training = ucg_training or dict() + if self.ucg_training: + self.ucg_prng = np.random.RandomState() + + def register_schedule(self, given_betas=None, beta_schedule="linear", timesteps=1000, + linear_start=1e-4, linear_end=2e-2, cosine_s=8e-3): + if exists(given_betas): + betas = given_betas + else: + betas = make_beta_schedule(beta_schedule, timesteps, linear_start=linear_start, linear_end=linear_end, + cosine_s=cosine_s) + alphas = 1. - betas + alphas_cumprod = np.cumprod(alphas, axis=0) + alphas_cumprod_prev = np.append(1., alphas_cumprod[:-1]) + + timesteps, = betas.shape + self.num_timesteps = int(timesteps) + self.linear_start = linear_start + self.linear_end = linear_end + assert alphas_cumprod.shape[0] == self.num_timesteps, 'alphas have to be defined for each timestep' + + to_torch = partial(torch.tensor, dtype=torch.float32) + + self.register_buffer('betas', to_torch(betas)) + self.register_buffer('alphas_cumprod', to_torch(alphas_cumprod)) + self.register_buffer('alphas_cumprod_prev', to_torch(alphas_cumprod_prev)) + + # calculations for diffusion q(x_t | x_{t-1}) and others + self.register_buffer('sqrt_alphas_cumprod', to_torch(np.sqrt(alphas_cumprod))) + self.register_buffer('sqrt_one_minus_alphas_cumprod', to_torch(np.sqrt(1. - alphas_cumprod))) + self.register_buffer('log_one_minus_alphas_cumprod', to_torch(np.log(1. - alphas_cumprod))) + self.register_buffer('sqrt_recip_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod))) + self.register_buffer('sqrt_recipm1_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod - 1))) + + # calculations for posterior q(x_{t-1} | x_t, x_0) + posterior_variance = (1 - self.v_posterior) * betas * (1. - alphas_cumprod_prev) / ( + 1. - alphas_cumprod) + self.v_posterior * betas + # above: equal to 1. / (1. / (1. - alpha_cumprod_tm1) + alpha_t / beta_t) + self.register_buffer('posterior_variance', to_torch(posterior_variance)) + # below: log calculation clipped because the posterior variance is 0 at the beginning of the diffusion chain + self.register_buffer('posterior_log_variance_clipped', to_torch(np.log(np.maximum(posterior_variance, 1e-20)))) + self.register_buffer('posterior_mean_coef1', to_torch( + betas * np.sqrt(alphas_cumprod_prev) / (1. - alphas_cumprod))) + self.register_buffer('posterior_mean_coef2', to_torch( + (1. - alphas_cumprod_prev) * np.sqrt(alphas) / (1. - alphas_cumprod))) + + if self.parameterization == "eps": + lvlb_weights = self.betas ** 2 / ( + 2 * self.posterior_variance * to_torch(alphas) * (1 - self.alphas_cumprod)) + elif self.parameterization == "x0": + lvlb_weights = 0.5 * np.sqrt(torch.Tensor(alphas_cumprod)) / (2. * 1 - torch.Tensor(alphas_cumprod)) + elif self.parameterization == "v": + lvlb_weights = torch.ones_like(self.betas ** 2 / ( + 2 * self.posterior_variance * to_torch(alphas) * (1 - self.alphas_cumprod))) + else: + raise NotImplementedError("mu not supported") + lvlb_weights[0] = lvlb_weights[1] + self.register_buffer('lvlb_weights', lvlb_weights, persistent=False) + assert not torch.isnan(self.lvlb_weights).all() + + @contextmanager + def ema_scope(self, context=None): + if self.use_ema: + self.model_ema.store(self.model.parameters()) + self.model_ema.copy_to(self.model) + if context is not None: + print(f"{context}: Switched to EMA weights") + try: + yield None + finally: + if self.use_ema: + self.model_ema.restore(self.model.parameters()) + if context is not None: + print(f"{context}: Restored training weights") + + @torch.no_grad() + def init_from_ckpt(self, path, ignore_keys=list(), only_model=False): + sd = torch.load(path, map_location="cpu") + if "state_dict" in list(sd.keys()): + sd = sd["state_dict"] + keys = list(sd.keys()) + for k in keys: + for ik in ignore_keys: + if k.startswith(ik): + print("Deleting key {} from state_dict.".format(k)) + del sd[k] + if self.make_it_fit: + n_params = len([name for name, _ in + itertools.chain(self.named_parameters(), + self.named_buffers())]) + for name, param in tqdm( + itertools.chain(self.named_parameters(), + self.named_buffers()), + desc="Fitting old weights to new weights", + total=n_params + ): + if not name in sd: + continue + old_shape = sd[name].shape + new_shape = param.shape + assert len(old_shape) == len(new_shape) + if len(new_shape) > 2: + # we only modify first two axes + assert new_shape[2:] == old_shape[2:] + # assumes first axis corresponds to output dim + if not new_shape == old_shape: + new_param = param.clone() + old_param = sd[name] + if len(new_shape) == 1: + for i in range(new_param.shape[0]): + new_param[i] = old_param[i % old_shape[0]] + elif len(new_shape) >= 2: + for i in range(new_param.shape[0]): + for j in range(new_param.shape[1]): + new_param[i, j] = old_param[i % old_shape[0], j % old_shape[1]] + + n_used_old = torch.ones(old_shape[1]) + for j in range(new_param.shape[1]): + n_used_old[j % old_shape[1]] += 1 + n_used_new = torch.zeros(new_shape[1]) + for j in range(new_param.shape[1]): + n_used_new[j] = n_used_old[j % old_shape[1]] + + n_used_new = n_used_new[None, :] + while len(n_used_new.shape) < len(new_shape): + n_used_new = n_used_new.unsqueeze(-1) + new_param /= n_used_new + + sd[name] = new_param + + missing, unexpected = self.load_state_dict(sd, strict=False) if not only_model else self.model.load_state_dict( + sd, strict=False) + print(f"Restored from {path} with {len(missing)} missing and {len(unexpected)} unexpected keys") + if len(missing) > 0: + print(f"Missing Keys:\n {missing}") + if len(unexpected) > 0: + print(f"\nUnexpected Keys:\n {unexpected}") + + def q_mean_variance(self, x_start, t): + """ + Get the distribution q(x_t | x_0). + :param x_start: the [N x C x ...] tensor of noiseless inputs. + :param t: the number of diffusion steps (minus 1). Here, 0 means one step. + :return: A tuple (mean, variance, log_variance), all of x_start's shape. + """ + mean = (extract_into_tensor(self.sqrt_alphas_cumprod, t, x_start.shape) * x_start) + variance = extract_into_tensor(1.0 - self.alphas_cumprod, t, x_start.shape) + log_variance = extract_into_tensor(self.log_one_minus_alphas_cumprod, t, x_start.shape) + return mean, variance, log_variance + + def predict_start_from_noise(self, x_t, t, noise): + return ( + extract_into_tensor(self.sqrt_recip_alphas_cumprod, t, x_t.shape) * x_t - + extract_into_tensor(self.sqrt_recipm1_alphas_cumprod, t, x_t.shape) * noise + ) + + def predict_start_from_z_and_v(self, x_t, t, v): + # self.register_buffer('sqrt_alphas_cumprod', to_torch(np.sqrt(alphas_cumprod))) + # self.register_buffer('sqrt_one_minus_alphas_cumprod', to_torch(np.sqrt(1. - alphas_cumprod))) + return ( + extract_into_tensor(self.sqrt_alphas_cumprod, t, x_t.shape) * x_t - + extract_into_tensor(self.sqrt_one_minus_alphas_cumprod, t, x_t.shape) * v + ) + + def predict_eps_from_z_and_v(self, x_t, t, v): + return ( + extract_into_tensor(self.sqrt_alphas_cumprod, t, x_t.shape) * v + + extract_into_tensor(self.sqrt_one_minus_alphas_cumprod, t, x_t.shape) * x_t + ) + + def q_posterior(self, x_start, x_t, t): + posterior_mean = ( + extract_into_tensor(self.posterior_mean_coef1, t, x_t.shape) * x_start + + extract_into_tensor(self.posterior_mean_coef2, t, x_t.shape) * x_t + ) + posterior_variance = extract_into_tensor(self.posterior_variance, t, x_t.shape) + posterior_log_variance_clipped = extract_into_tensor(self.posterior_log_variance_clipped, t, x_t.shape) + return posterior_mean, posterior_variance, posterior_log_variance_clipped + + def p_mean_variance(self, x, t, clip_denoised: bool): + model_out = self.model(x, t) + if self.parameterization == "eps": + x_recon = self.predict_start_from_noise(x, t=t, noise=model_out) + elif self.parameterization == "x0": + x_recon = model_out + if clip_denoised: + x_recon.clamp_(-1., 1.) + + model_mean, posterior_variance, posterior_log_variance = self.q_posterior(x_start=x_recon, x_t=x, t=t) + return model_mean, posterior_variance, posterior_log_variance + + @torch.no_grad() + def p_sample(self, x, t, clip_denoised=True, repeat_noise=False): + b, *_, device = *x.shape, x.device + model_mean, _, model_log_variance = self.p_mean_variance(x=x, t=t, clip_denoised=clip_denoised) + noise = noise_like(x.shape, device, repeat_noise) + # no noise when t == 0 + nonzero_mask = (1 - (t == 0).float()).reshape(b, *((1,) * (len(x.shape) - 1))) + return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise + + @torch.no_grad() + def p_sample_loop(self, shape, return_intermediates=False): + device = self.betas.device + b = shape[0] + img = torch.randn(shape, device=device) + intermediates = [img] + for i in tqdm(reversed(range(0, self.num_timesteps)), desc='Sampling t', total=self.num_timesteps): + img = self.p_sample(img, torch.full((b,), i, device=device, dtype=torch.long), + clip_denoised=self.clip_denoised) + if i % self.log_every_t == 0 or i == self.num_timesteps - 1: + intermediates.append(img) + if return_intermediates: + return img, intermediates + return img + + @torch.no_grad() + def sample(self, batch_size=16, return_intermediates=False): + image_size = self.image_size + channels = self.channels + return self.p_sample_loop((batch_size, channels, image_size, image_size), + return_intermediates=return_intermediates) + + def q_sample(self, x_start, t, noise=None): + noise = default(noise, lambda: torch.randn_like(x_start)) + return (extract_into_tensor(self.sqrt_alphas_cumprod, t, x_start.shape) * x_start + + extract_into_tensor(self.sqrt_one_minus_alphas_cumprod, t, x_start.shape) * noise) + + def get_v(self, x, noise, t): + return ( + extract_into_tensor(self.sqrt_alphas_cumprod, t, x.shape) * noise - + extract_into_tensor(self.sqrt_one_minus_alphas_cumprod, t, x.shape) * x + ) + + def get_loss(self, pred, target, mean=True): + if self.loss_type == 'l1': + loss = (target - pred).abs() + if mean: + loss = loss.mean() + elif self.loss_type == 'l2': + if mean: + loss = torch.nn.functional.mse_loss(target, pred) + else: + loss = torch.nn.functional.mse_loss(target, pred, reduction='none') + else: + raise NotImplementedError("unknown loss type '{loss_type}'") + + return loss + + def p_losses(self, x_start, t, noise=None): + noise = default(noise, lambda: torch.randn_like(x_start)) + x_noisy = self.q_sample(x_start=x_start, t=t, noise=noise) + model_out = self.model(x_noisy, t) + + loss_dict = {} + if self.parameterization == "eps": + target = noise + elif self.parameterization == "x0": + target = x_start + elif self.parameterization == "v": + target = self.get_v(x_start, noise, t) + else: + raise NotImplementedError(f"Parameterization {self.parameterization} not yet supported") + + loss = self.get_loss(model_out, target, mean=False).mean(dim=[1, 2, 3]) + + log_prefix = 'train' if self.training else 'val' + + loss_dict.update({f'{log_prefix}/loss_simple': loss.mean()}) + loss_simple = loss.mean() * self.l_simple_weight + + loss_vlb = (self.lvlb_weights[t] * loss).mean() + loss_dict.update({f'{log_prefix}/loss_vlb': loss_vlb}) + + loss = loss_simple + self.original_elbo_weight * loss_vlb + + loss_dict.update({f'{log_prefix}/loss': loss}) + + return loss, loss_dict + + def forward(self, x, *args, **kwargs): + # b, c, h, w, device, img_size, = *x.shape, x.device, self.image_size + # assert h == img_size and w == img_size, f'height and width of image must be {img_size}' + t = torch.randint(0, self.num_timesteps, (x.shape[0],), device=self.device).long() + return self.p_losses(x, t, *args, **kwargs) + + def get_input(self, batch, k): + x = batch[k] + if len(x.shape) == 3: + x = x[..., None] + x = rearrange(x, 'b h w c -> b c h w') + x = x.to(memory_format=torch.contiguous_format).float() + return x + + def shared_step(self, batch): + x = self.get_input(batch, self.first_stage_key) + loss, loss_dict = self(x) + return loss, loss_dict + + def training_step(self, batch, batch_idx): + for k in self.ucg_training: + p = self.ucg_training[k]["p"] + val = self.ucg_training[k]["val"] + if val is None: + val = "" + for i in range(len(batch[k])): + if self.ucg_prng.choice(2, p=[1 - p, p]): + batch[k][i] = val + + loss, loss_dict = self.shared_step(batch) + + self.log_dict(loss_dict, prog_bar=True, + logger=True, on_step=True, on_epoch=True) + + self.log("global_step", self.global_step, + prog_bar=True, logger=True, on_step=True, on_epoch=False) + + if self.use_scheduler: + lr = self.optimizers().param_groups[0]['lr'] + self.log('lr_abs', lr, prog_bar=True, logger=True, on_step=True, on_epoch=False) + + return loss + + @torch.no_grad() + def validation_step(self, batch, batch_idx): + _, loss_dict_no_ema = self.shared_step(batch) + with self.ema_scope(): + _, loss_dict_ema = self.shared_step(batch) + loss_dict_ema = {key + '_ema': loss_dict_ema[key] for key in loss_dict_ema} + self.log_dict(loss_dict_no_ema, prog_bar=False, logger=True, on_step=False, on_epoch=True) + self.log_dict(loss_dict_ema, prog_bar=False, logger=True, on_step=False, on_epoch=True) + + def on_train_batch_end(self, *args, **kwargs): + if self.use_ema: + self.model_ema(self.model) + + def _get_rows_from_list(self, samples): + n_imgs_per_row = len(samples) + denoise_grid = rearrange(samples, 'n b c h w -> b n c h w') + denoise_grid = rearrange(denoise_grid, 'b n c h w -> (b n) c h w') + denoise_grid = make_grid(denoise_grid, nrow=n_imgs_per_row) + return denoise_grid + + @torch.no_grad() + def log_images(self, batch, N=8, n_row=2, sample=True, return_keys=None, **kwargs): + log = dict() + x = self.get_input(batch, self.first_stage_key) + N = min(x.shape[0], N) + n_row = min(x.shape[0], n_row) + x = x.to(self.device)[:N] + log["inputs"] = x + + # get diffusion row + diffusion_row = list() + x_start = x[:n_row] + + for t in range(self.num_timesteps): + if t % self.log_every_t == 0 or t == self.num_timesteps - 1: + t = repeat(torch.tensor([t]), '1 -> b', b=n_row) + t = t.to(self.device).long() + noise = torch.randn_like(x_start) + x_noisy = self.q_sample(x_start=x_start, t=t, noise=noise) + diffusion_row.append(x_noisy) + + log["diffusion_row"] = self._get_rows_from_list(diffusion_row) + + if sample: + # get denoise row + with self.ema_scope("Plotting"): + samples, denoise_row = self.sample(batch_size=N, return_intermediates=True) + + log["samples"] = samples + log["denoise_row"] = self._get_rows_from_list(denoise_row) + + if return_keys: + if np.intersect1d(list(log.keys()), return_keys).shape[0] == 0: + return log + else: + return {key: log[key] for key in return_keys} + return log + + def configure_optimizers(self): + lr = self.learning_rate + params = list(self.model.parameters()) + if self.learn_logvar: + params = params + [self.logvar] + opt = torch.optim.AdamW(params, lr=lr) + return opt + + +class LatentDiffusion(DDPM): + """main class""" + + def __init__(self, + first_stage_config, + cond_stage_config, + num_timesteps_cond=None, + cond_stage_key="image", + cond_stage_trainable=False, + concat_mode=True, + cond_stage_forward=None, + conditioning_key=None, + scale_factor=1.0, + scale_by_std=False, + force_null_conditioning=False, + *args, **kwargs): + self.force_null_conditioning = force_null_conditioning + self.num_timesteps_cond = default(num_timesteps_cond, 1) + self.scale_by_std = scale_by_std + assert self.num_timesteps_cond <= kwargs['timesteps'] + # for backwards compatibility after implementation of DiffusionWrapper + if conditioning_key is None: + conditioning_key = 'concat' if concat_mode else 'crossattn' + if cond_stage_config == '__is_unconditional__' and not self.force_null_conditioning: + conditioning_key = None + ckpt_path = kwargs.pop("ckpt_path", None) + reset_ema = kwargs.pop("reset_ema", False) + reset_num_ema_updates = kwargs.pop("reset_num_ema_updates", False) + ignore_keys = kwargs.pop("ignore_keys", []) + super().__init__(conditioning_key=conditioning_key, *args, **kwargs) + self.concat_mode = concat_mode + self.cond_stage_trainable = cond_stage_trainable + self.cond_stage_key = cond_stage_key + try: + self.num_downs = len(first_stage_config.params.ddconfig.ch_mult) - 1 + except: + self.num_downs = 0 + if not scale_by_std: + self.scale_factor = scale_factor + else: + self.register_buffer('scale_factor', torch.tensor(scale_factor)) + self.instantiate_first_stage(first_stage_config) + self.instantiate_cond_stage(cond_stage_config) + self.cond_stage_forward = cond_stage_forward + self.clip_denoised = False + self.bbox_tokenizer = None + + self.restarted_from_ckpt = False + if ckpt_path is not None: + self.init_from_ckpt(ckpt_path, ignore_keys) + self.restarted_from_ckpt = True + if reset_ema: + assert self.use_ema + print( + f"Resetting ema to pure model weights. This is useful when restoring from an ema-only checkpoint.") + self.model_ema = LitEma(self.model) + if reset_num_ema_updates: + print(" +++++++++++ WARNING: RESETTING NUM_EMA UPDATES TO ZERO +++++++++++ ") + assert self.use_ema + self.model_ema.reset_num_updates() + + def make_cond_schedule(self, ): + self.cond_ids = torch.full(size=(self.num_timesteps,), fill_value=self.num_timesteps - 1, dtype=torch.long) + ids = torch.round(torch.linspace(0, self.num_timesteps - 1, self.num_timesteps_cond)).long() + self.cond_ids[:self.num_timesteps_cond] = ids + + @rank_zero_only + @torch.no_grad() + def on_train_batch_start(self, batch, batch_idx, dataloader_idx): + # only for very first batch + if self.scale_by_std and self.current_epoch == 0 and self.global_step == 0 and batch_idx == 0 and not self.restarted_from_ckpt: + assert self.scale_factor == 1., 'rather not use custom rescaling and std-rescaling simultaneously' + # set rescale weight to 1./std of encodings + print("### USING STD-RESCALING ###") + x = super().get_input(batch, self.first_stage_key) + x = x.to(self.device) + encoder_posterior = self.encode_first_stage(x) + z = self.get_first_stage_encoding(encoder_posterior).detach() + del self.scale_factor + self.register_buffer('scale_factor', 1. / z.flatten().std()) + print(f"setting self.scale_factor to {self.scale_factor}") + print("### USING STD-RESCALING ###") + + def register_schedule(self, + given_betas=None, beta_schedule="linear", timesteps=1000, + linear_start=1e-4, linear_end=2e-2, cosine_s=8e-3): + super().register_schedule(given_betas, beta_schedule, timesteps, linear_start, linear_end, cosine_s) + + self.shorten_cond_schedule = self.num_timesteps_cond > 1 + if self.shorten_cond_schedule: + self.make_cond_schedule() + + def instantiate_first_stage(self, config): + model = instantiate_from_config(config) + self.first_stage_model = model.eval() + self.first_stage_model.train = disabled_train + for param in self.first_stage_model.parameters(): + param.requires_grad = False + + def instantiate_cond_stage(self, config): + if not self.cond_stage_trainable: + if config == "__is_first_stage__": + print("Using first stage also as cond stage.") + self.cond_stage_model = self.first_stage_model + elif config == "__is_unconditional__": + print(f"Training {self.__class__.__name__} as an unconditional model.") + self.cond_stage_model = None + # self.be_unconditional = True + else: + model = instantiate_from_config(config) + self.cond_stage_model = model.eval() + self.cond_stage_model.train = disabled_train + for param in self.cond_stage_model.parameters(): + param.requires_grad = False + else: + assert config != '__is_first_stage__' + assert config != '__is_unconditional__' + model = instantiate_from_config(config) + self.cond_stage_model = model + + def _get_denoise_row_from_list(self, samples, desc='', force_no_decoder_quantization=False): + denoise_row = [] + for zd in tqdm(samples, desc=desc): + denoise_row.append(self.decode_first_stage(zd.to(self.device), + force_not_quantize=force_no_decoder_quantization)) + n_imgs_per_row = len(denoise_row) + denoise_row = torch.stack(denoise_row) # n_log_step, n_row, C, H, W + denoise_grid = rearrange(denoise_row, 'n b c h w -> b n c h w') + denoise_grid = rearrange(denoise_grid, 'b n c h w -> (b n) c h w') + denoise_grid = make_grid(denoise_grid, nrow=n_imgs_per_row) + return denoise_grid + + def get_first_stage_encoding(self, encoder_posterior): + if isinstance(encoder_posterior, DiagonalGaussianDistribution): + z = encoder_posterior.sample() + elif isinstance(encoder_posterior, torch.Tensor): + z = encoder_posterior + else: + raise NotImplementedError(f"encoder_posterior of type '{type(encoder_posterior)}' not yet implemented") + return self.scale_factor * z + + def get_learned_conditioning(self, c): + if self.cond_stage_forward is None: + if hasattr(self.cond_stage_model, 'encode') and callable(self.cond_stage_model.encode): + c = self.cond_stage_model.encode(c) + if isinstance(c, DiagonalGaussianDistribution): + c = c.mode() + else: + c = self.cond_stage_model(c) + else: + assert hasattr(self.cond_stage_model, self.cond_stage_forward) + c = getattr(self.cond_stage_model, self.cond_stage_forward)(c) + return c + + def meshgrid(self, h, w): + y = torch.arange(0, h).view(h, 1, 1).repeat(1, w, 1) + x = torch.arange(0, w).view(1, w, 1).repeat(h, 1, 1) + + arr = torch.cat([y, x], dim=-1) + return arr + + def delta_border(self, h, w): + """ + :param h: height + :param w: width + :return: normalized distance to image border, + wtith min distance = 0 at border and max dist = 0.5 at image center + """ + lower_right_corner = torch.tensor([h - 1, w - 1]).view(1, 1, 2) + arr = self.meshgrid(h, w) / lower_right_corner + dist_left_up = torch.min(arr, dim=-1, keepdims=True)[0] + dist_right_down = torch.min(1 - arr, dim=-1, keepdims=True)[0] + edge_dist = torch.min(torch.cat([dist_left_up, dist_right_down], dim=-1), dim=-1)[0] + return edge_dist + + def get_weighting(self, h, w, Ly, Lx, device): + weighting = self.delta_border(h, w) + weighting = torch.clip(weighting, self.split_input_params["clip_min_weight"], + self.split_input_params["clip_max_weight"], ) + weighting = weighting.view(1, h * w, 1).repeat(1, 1, Ly * Lx).to(device) + + if self.split_input_params["tie_braker"]: + L_weighting = self.delta_border(Ly, Lx) + L_weighting = torch.clip(L_weighting, + self.split_input_params["clip_min_tie_weight"], + self.split_input_params["clip_max_tie_weight"]) + + L_weighting = L_weighting.view(1, 1, Ly * Lx).to(device) + weighting = weighting * L_weighting + return weighting + + def get_fold_unfold(self, x, kernel_size, stride, uf=1, df=1): # todo load once not every time, shorten code + """ + :param x: img of size (bs, c, h, w) + :return: n img crops of size (n, bs, c, kernel_size[0], kernel_size[1]) + """ + bs, nc, h, w = x.shape + + # number of crops in image + Ly = (h - kernel_size[0]) // stride[0] + 1 + Lx = (w - kernel_size[1]) // stride[1] + 1 + + if uf == 1 and df == 1: + fold_params = dict(kernel_size=kernel_size, dilation=1, padding=0, stride=stride) + unfold = torch.nn.Unfold(**fold_params) + + fold = torch.nn.Fold(output_size=x.shape[2:], **fold_params) + + weighting = self.get_weighting(kernel_size[0], kernel_size[1], Ly, Lx, x.device).to(x.dtype) + normalization = fold(weighting).view(1, 1, h, w) # normalizes the overlap + weighting = weighting.view((1, 1, kernel_size[0], kernel_size[1], Ly * Lx)) + + elif uf > 1 and df == 1: + fold_params = dict(kernel_size=kernel_size, dilation=1, padding=0, stride=stride) + unfold = torch.nn.Unfold(**fold_params) + + fold_params2 = dict(kernel_size=(kernel_size[0] * uf, kernel_size[0] * uf), + dilation=1, padding=0, + stride=(stride[0] * uf, stride[1] * uf)) + fold = torch.nn.Fold(output_size=(x.shape[2] * uf, x.shape[3] * uf), **fold_params2) + + weighting = self.get_weighting(kernel_size[0] * uf, kernel_size[1] * uf, Ly, Lx, x.device).to(x.dtype) + normalization = fold(weighting).view(1, 1, h * uf, w * uf) # normalizes the overlap + weighting = weighting.view((1, 1, kernel_size[0] * uf, kernel_size[1] * uf, Ly * Lx)) + + elif df > 1 and uf == 1: + fold_params = dict(kernel_size=kernel_size, dilation=1, padding=0, stride=stride) + unfold = torch.nn.Unfold(**fold_params) + + fold_params2 = dict(kernel_size=(kernel_size[0] // df, kernel_size[0] // df), + dilation=1, padding=0, + stride=(stride[0] // df, stride[1] // df)) + fold = torch.nn.Fold(output_size=(x.shape[2] // df, x.shape[3] // df), **fold_params2) + + weighting = self.get_weighting(kernel_size[0] // df, kernel_size[1] // df, Ly, Lx, x.device).to(x.dtype) + normalization = fold(weighting).view(1, 1, h // df, w // df) # normalizes the overlap + weighting = weighting.view((1, 1, kernel_size[0] // df, kernel_size[1] // df, Ly * Lx)) + + else: + raise NotImplementedError + + return fold, unfold, normalization, weighting + + @torch.no_grad() + def get_input(self, batch, k, return_first_stage_outputs=False, force_c_encode=False, + cond_key=None, return_original_cond=False, bs=None, return_x=False): + x = super().get_input(batch, k) + if bs is not None: + x = x[:bs] + x = x.to(self.device) + encoder_posterior = self.encode_first_stage(x) + z = self.get_first_stage_encoding(encoder_posterior).detach() + + if self.model.conditioning_key is not None and not self.force_null_conditioning: + if cond_key is None: + cond_key = self.cond_stage_key + if cond_key != self.first_stage_key: + if cond_key in ['caption', 'coordinates_bbox', "txt"]: + xc = batch[cond_key] + elif cond_key in ['class_label', 'cls']: + xc = batch + else: + xc = super().get_input(batch, cond_key).to(self.device) + else: + xc = x + if not self.cond_stage_trainable or force_c_encode: + if isinstance(xc, dict) or isinstance(xc, list): + c = self.get_learned_conditioning(xc) + else: + c = self.get_learned_conditioning(xc.to(self.device)) + else: + c = xc + if bs is not None: + c = c[:bs] + + if self.use_positional_encodings: + pos_x, pos_y = self.compute_latent_shifts(batch) + ckey = __conditioning_keys__[self.model.conditioning_key] + c = {ckey: c, 'pos_x': pos_x, 'pos_y': pos_y} + + else: + c = None + xc = None + if self.use_positional_encodings: + pos_x, pos_y = self.compute_latent_shifts(batch) + c = {'pos_x': pos_x, 'pos_y': pos_y} + out = [z, c] + if return_first_stage_outputs: + xrec = self.decode_first_stage(z) + out.extend([x, xrec]) + if return_x: + out.extend([x]) + if return_original_cond: + out.append(xc) + return out + + @torch.no_grad() + def decode_first_stage(self, z, predict_cids=False, force_not_quantize=False): + if predict_cids: + if z.dim() == 4: + z = torch.argmax(z.exp(), dim=1).long() + z = self.first_stage_model.quantize.get_codebook_entry(z, shape=None) + z = rearrange(z, 'b h w c -> b c h w').contiguous() + + z = 1. / self.scale_factor * z + return self.first_stage_model.decode(z) + + @torch.no_grad() + def encode_first_stage(self, x): + return self.first_stage_model.encode(x) + + def shared_step(self, batch, **kwargs): + x, c = self.get_input(batch, self.first_stage_key) + loss = self(x, c) + return loss + + def forward(self, x, c, *args, **kwargs): + t = torch.randint(0, self.num_timesteps, (x.shape[0],), device=self.device).long() + if self.model.conditioning_key is not None: + assert c is not None + if self.cond_stage_trainable: + c = self.get_learned_conditioning(c) + if self.shorten_cond_schedule: # TODO: drop this option + tc = self.cond_ids[t].to(self.device) + c = self.q_sample(x_start=c, t=tc, noise=torch.randn_like(c.float())) + return self.p_losses(x, c, t, *args, **kwargs) + + def apply_model(self, x_noisy, t, cond, return_ids=False): + if isinstance(cond, dict): + # hybrid case, cond is expected to be a dict + pass + else: + if not isinstance(cond, list): + cond = [cond] + key = 'c_concat' if self.model.conditioning_key == 'concat' else 'c_crossattn' + cond = {key: cond} + + x_recon = self.model(x_noisy, t, **cond) + + if isinstance(x_recon, tuple) and not return_ids: + return x_recon[0] + else: + return x_recon + + def _predict_eps_from_xstart(self, x_t, t, pred_xstart): + return (extract_into_tensor(self.sqrt_recip_alphas_cumprod, t, x_t.shape) * x_t - pred_xstart) / \ + extract_into_tensor(self.sqrt_recipm1_alphas_cumprod, t, x_t.shape) + + def _prior_bpd(self, x_start): + """ + Get the prior KL term for the variational lower-bound, measured in + bits-per-dim. + This term can't be optimized, as it only depends on the encoder. + :param x_start: the [N x C x ...] tensor of inputs. + :return: a batch of [N] KL values (in bits), one per batch element. + """ + batch_size = x_start.shape[0] + t = torch.tensor([self.num_timesteps - 1] * batch_size, device=x_start.device) + qt_mean, _, qt_log_variance = self.q_mean_variance(x_start, t) + kl_prior = normal_kl(mean1=qt_mean, logvar1=qt_log_variance, mean2=0.0, logvar2=0.0) + return mean_flat(kl_prior) / np.log(2.0) + + def p_losses(self, x_start, cond, t, noise=None): + noise = default(noise, lambda: torch.randn_like(x_start)) + x_noisy = self.q_sample(x_start=x_start, t=t, noise=noise) + model_output = self.apply_model(x_noisy, t, cond) + + loss_dict = {} + prefix = 'train' if self.training else 'val' + + if self.parameterization == "x0": + target = x_start + elif self.parameterization == "eps": + target = noise + elif self.parameterization == "v": + target = self.get_v(x_start, noise, t) + else: + raise NotImplementedError() + + loss_simple = self.get_loss(model_output, target, mean=False).mean([1, 2, 3]) + loss_dict.update({f'{prefix}/loss_simple': loss_simple.mean()}) + + logvar_t = self.logvar[t].to(self.device) + loss = loss_simple / torch.exp(logvar_t) + logvar_t + # loss = loss_simple / torch.exp(self.logvar) + self.logvar + if self.learn_logvar: + loss_dict.update({f'{prefix}/loss_gamma': loss.mean()}) + loss_dict.update({'logvar': self.logvar.data.mean()}) + + loss = self.l_simple_weight * loss.mean() + + loss_vlb = self.get_loss(model_output, target, mean=False).mean(dim=(1, 2, 3)) + loss_vlb = (self.lvlb_weights[t] * loss_vlb).mean() + loss_dict.update({f'{prefix}/loss_vlb': loss_vlb}) + loss += (self.original_elbo_weight * loss_vlb) + loss_dict.update({f'{prefix}/loss': loss}) + + return loss, loss_dict + + def p_mean_variance(self, x, c, t, clip_denoised: bool, return_codebook_ids=False, quantize_denoised=False, + return_x0=False, score_corrector=None, corrector_kwargs=None): + t_in = t + model_out = self.apply_model(x, t_in, c, return_ids=return_codebook_ids) + + if score_corrector is not None: + assert self.parameterization == "eps" + model_out = score_corrector.modify_score(self, model_out, x, t, c, **corrector_kwargs) + + if return_codebook_ids: + model_out, logits = model_out + + if self.parameterization == "eps": + x_recon = self.predict_start_from_noise(x, t=t, noise=model_out) + elif self.parameterization == "x0": + x_recon = model_out + else: + raise NotImplementedError() + + if clip_denoised: + x_recon.clamp_(-1., 1.) + if quantize_denoised: + x_recon, _, [_, _, indices] = self.first_stage_model.quantize(x_recon) + model_mean, posterior_variance, posterior_log_variance = self.q_posterior(x_start=x_recon, x_t=x, t=t) + if return_codebook_ids: + return model_mean, posterior_variance, posterior_log_variance, logits + elif return_x0: + return model_mean, posterior_variance, posterior_log_variance, x_recon + else: + return model_mean, posterior_variance, posterior_log_variance + + @torch.no_grad() + def p_sample(self, x, c, t, clip_denoised=False, repeat_noise=False, + return_codebook_ids=False, quantize_denoised=False, return_x0=False, + temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None): + b, *_, device = *x.shape, x.device + outputs = self.p_mean_variance(x=x, c=c, t=t, clip_denoised=clip_denoised, + return_codebook_ids=return_codebook_ids, + quantize_denoised=quantize_denoised, + return_x0=return_x0, + score_corrector=score_corrector, corrector_kwargs=corrector_kwargs) + if return_codebook_ids: + raise DeprecationWarning("Support dropped.") + model_mean, _, model_log_variance, logits = outputs + elif return_x0: + model_mean, _, model_log_variance, x0 = outputs + else: + model_mean, _, model_log_variance = outputs + + noise = noise_like(x.shape, device, repeat_noise) * temperature + if noise_dropout > 0.: + noise = torch.nn.functional.dropout(noise, p=noise_dropout) + # no noise when t == 0 + nonzero_mask = (1 - (t == 0).float()).reshape(b, *((1,) * (len(x.shape) - 1))) + + if return_codebook_ids: + return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise, logits.argmax(dim=1) + if return_x0: + return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise, x0 + else: + return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise + + @torch.no_grad() + def progressive_denoising(self, cond, shape, verbose=True, callback=None, quantize_denoised=False, + img_callback=None, mask=None, x0=None, temperature=1., noise_dropout=0., + score_corrector=None, corrector_kwargs=None, batch_size=None, x_T=None, start_T=None, + log_every_t=None): + if not log_every_t: + log_every_t = self.log_every_t + timesteps = self.num_timesteps + if batch_size is not None: + b = batch_size if batch_size is not None else shape[0] + shape = [batch_size] + list(shape) + else: + b = batch_size = shape[0] + if x_T is None: + img = torch.randn(shape, device=self.device) + else: + img = x_T + intermediates = [] + if cond is not None: + if isinstance(cond, dict): + cond = {key: cond[key][:batch_size] if not isinstance(cond[key], list) else + list(map(lambda x: x[:batch_size], cond[key])) for key in cond} + else: + cond = [c[:batch_size] for c in cond] if isinstance(cond, list) else cond[:batch_size] + + if start_T is not None: + timesteps = min(timesteps, start_T) + iterator = tqdm(reversed(range(0, timesteps)), desc='Progressive Generation', + total=timesteps) if verbose else reversed( + range(0, timesteps)) + if type(temperature) == float: + temperature = [temperature] * timesteps + + for i in iterator: + ts = torch.full((b,), i, device=self.device, dtype=torch.long) + if self.shorten_cond_schedule: + assert self.model.conditioning_key != 'hybrid' + tc = self.cond_ids[ts].to(cond.device) + cond = self.q_sample(x_start=cond, t=tc, noise=torch.randn_like(cond)) + + img, x0_partial = self.p_sample(img, cond, ts, + clip_denoised=self.clip_denoised, + quantize_denoised=quantize_denoised, return_x0=True, + temperature=temperature[i], noise_dropout=noise_dropout, + score_corrector=score_corrector, corrector_kwargs=corrector_kwargs) + if mask is not None: + assert x0 is not None + img_orig = self.q_sample(x0, ts) + img = img_orig * mask + (1. - mask) * img + + if i % log_every_t == 0 or i == timesteps - 1: + intermediates.append(x0_partial) + if callback: callback(i) + if img_callback: img_callback(img, i) + return img, intermediates + + @torch.no_grad() + def p_sample_loop(self, cond, shape, return_intermediates=False, + x_T=None, verbose=True, callback=None, timesteps=None, quantize_denoised=False, + mask=None, x0=None, img_callback=None, start_T=None, + log_every_t=None): + + if not log_every_t: + log_every_t = self.log_every_t + device = self.betas.device + b = shape[0] + if x_T is None: + img = torch.randn(shape, device=device) + else: + img = x_T + + intermediates = [img] + if timesteps is None: + timesteps = self.num_timesteps + + if start_T is not None: + timesteps = min(timesteps, start_T) + iterator = tqdm(reversed(range(0, timesteps)), desc='Sampling t', total=timesteps) if verbose else reversed( + range(0, timesteps)) + + if mask is not None: + assert x0 is not None + assert x0.shape[2:3] == mask.shape[2:3] # spatial size has to match + + for i in iterator: + ts = torch.full((b,), i, device=device, dtype=torch.long) + if self.shorten_cond_schedule: + assert self.model.conditioning_key != 'hybrid' + tc = self.cond_ids[ts].to(cond.device) + cond = self.q_sample(x_start=cond, t=tc, noise=torch.randn_like(cond)) + + img = self.p_sample(img, cond, ts, + clip_denoised=self.clip_denoised, + quantize_denoised=quantize_denoised) + if mask is not None: + img_orig = self.q_sample(x0, ts) + img = img_orig * mask + (1. - mask) * img + + if i % log_every_t == 0 or i == timesteps - 1: + intermediates.append(img) + if callback: callback(i) + if img_callback: img_callback(img, i) + + if return_intermediates: + return img, intermediates + return img + + @torch.no_grad() + def sample(self, cond, batch_size=16, return_intermediates=False, x_T=None, + verbose=True, timesteps=None, quantize_denoised=False, + mask=None, x0=None, shape=None, **kwargs): + if shape is None: + shape = (batch_size, self.channels, self.image_size, self.image_size) + if cond is not None: + if isinstance(cond, dict): + cond = {key: cond[key][:batch_size] if not isinstance(cond[key], list) else + list(map(lambda x: x[:batch_size], cond[key])) for key in cond} + else: + cond = [c[:batch_size] for c in cond] if isinstance(cond, list) else cond[:batch_size] + return self.p_sample_loop(cond, + shape, + return_intermediates=return_intermediates, x_T=x_T, + verbose=verbose, timesteps=timesteps, quantize_denoised=quantize_denoised, + mask=mask, x0=x0) + + @torch.no_grad() + def sample_log(self, cond, batch_size, ddim, ddim_steps, **kwargs): + if ddim: + ddim_sampler = DDIMSampler(self) + shape = (self.channels, self.image_size, self.image_size) + samples, intermediates = ddim_sampler.sample(ddim_steps, batch_size, + shape, cond, verbose=False, **kwargs) + + else: + samples, intermediates = self.sample(cond=cond, batch_size=batch_size, + return_intermediates=True, **kwargs) + + return samples, intermediates + + @torch.no_grad() + def get_unconditional_conditioning(self, batch_size, null_label=None): + if null_label is not None: + xc = null_label + if isinstance(xc, ListConfig): + xc = list(xc) + if isinstance(xc, dict) or isinstance(xc, list): + c = self.get_learned_conditioning(xc) + else: + if hasattr(xc, "to"): + xc = xc.to(self.device) + c = self.get_learned_conditioning(xc) + else: + if self.cond_stage_key in ["class_label", "cls"]: + xc = self.cond_stage_model.get_unconditional_conditioning(batch_size, device=self.device) + return self.get_learned_conditioning(xc) + else: + raise NotImplementedError("todo") + if isinstance(c, list): # in case the encoder gives us a list + for i in range(len(c)): + c[i] = repeat(c[i], '1 ... -> b ...', b=batch_size).to(self.device) + else: + c = repeat(c, '1 ... -> b ...', b=batch_size).to(self.device) + return c + + @torch.no_grad() + def log_images(self, batch, N=8, n_row=4, sample=True, ddim_steps=50, ddim_eta=0., return_keys=None, + quantize_denoised=True, inpaint=True, plot_denoise_rows=False, plot_progressive_rows=True, + plot_diffusion_rows=True, unconditional_guidance_scale=1., unconditional_guidance_label=None, + use_ema_scope=True, + **kwargs): + ema_scope = self.ema_scope if use_ema_scope else nullcontext + use_ddim = ddim_steps is not None + + log = dict() + z, c, x, xrec, xc = self.get_input(batch, self.first_stage_key, + return_first_stage_outputs=True, + force_c_encode=True, + return_original_cond=True, + bs=N) + N = min(x.shape[0], N) + n_row = min(x.shape[0], n_row) + log["inputs"] = x + log["reconstruction"] = xrec + if self.model.conditioning_key is not None: + if hasattr(self.cond_stage_model, "decode"): + xc = self.cond_stage_model.decode(c) + log["conditioning"] = xc + elif self.cond_stage_key in ["caption", "txt"]: + xc = log_txt_as_img((x.shape[2], x.shape[3]), batch[self.cond_stage_key], size=x.shape[2] // 25) + log["conditioning"] = xc + elif self.cond_stage_key in ['class_label', "cls"]: + try: + xc = log_txt_as_img((x.shape[2], x.shape[3]), batch["human_label"], size=x.shape[2] // 25) + log['conditioning'] = xc + except KeyError: + # probably no "human_label" in batch + pass + elif isimage(xc): + log["conditioning"] = xc + if ismap(xc): + log["original_conditioning"] = self.to_rgb(xc) + + if plot_diffusion_rows: + # get diffusion row + diffusion_row = list() + z_start = z[:n_row] + for t in range(self.num_timesteps): + if t % self.log_every_t == 0 or t == self.num_timesteps - 1: + t = repeat(torch.tensor([t]), '1 -> b', b=n_row) + t = t.to(self.device).long() + noise = torch.randn_like(z_start) + z_noisy = self.q_sample(x_start=z_start, t=t, noise=noise) + diffusion_row.append(self.decode_first_stage(z_noisy)) + + diffusion_row = torch.stack(diffusion_row) # n_log_step, n_row, C, H, W + diffusion_grid = rearrange(diffusion_row, 'n b c h w -> b n c h w') + diffusion_grid = rearrange(diffusion_grid, 'b n c h w -> (b n) c h w') + diffusion_grid = make_grid(diffusion_grid, nrow=diffusion_row.shape[0]) + log["diffusion_row"] = diffusion_grid + + if sample: + # get denoise row + with ema_scope("Sampling"): + samples, z_denoise_row = self.sample_log(cond=c, batch_size=N, ddim=use_ddim, + ddim_steps=ddim_steps, eta=ddim_eta) + # samples, z_denoise_row = self.sample(cond=c, batch_size=N, return_intermediates=True) + x_samples = self.decode_first_stage(samples) + log["samples"] = x_samples + if plot_denoise_rows: + denoise_grid = self._get_denoise_row_from_list(z_denoise_row) + log["denoise_row"] = denoise_grid + + if quantize_denoised and not isinstance(self.first_stage_model, AutoencoderKL) and not isinstance( + self.first_stage_model, IdentityFirstStage): + # also display when quantizing x0 while sampling + with ema_scope("Plotting Quantized Denoised"): + samples, z_denoise_row = self.sample_log(cond=c, batch_size=N, ddim=use_ddim, + ddim_steps=ddim_steps, eta=ddim_eta, + quantize_denoised=True) + # samples, z_denoise_row = self.sample(cond=c, batch_size=N, return_intermediates=True, + # quantize_denoised=True) + x_samples = self.decode_first_stage(samples.to(self.device)) + log["samples_x0_quantized"] = x_samples + + if unconditional_guidance_scale > 1.0: + uc = self.get_unconditional_conditioning(N, unconditional_guidance_label) + if self.model.conditioning_key == "crossattn-adm": + uc = {"c_crossattn": [uc], "c_adm": c["c_adm"]} + with ema_scope("Sampling with classifier-free guidance"): + samples_cfg, _ = self.sample_log(cond=c, batch_size=N, ddim=use_ddim, + ddim_steps=ddim_steps, eta=ddim_eta, + unconditional_guidance_scale=unconditional_guidance_scale, + unconditional_conditioning=uc, + ) + x_samples_cfg = self.decode_first_stage(samples_cfg) + log[f"samples_cfg_scale_{unconditional_guidance_scale:.2f}"] = x_samples_cfg + + if inpaint: + # make a simple center square + b, h, w = z.shape[0], z.shape[2], z.shape[3] + mask = torch.ones(N, h, w).to(self.device) + # zeros will be filled in + mask[:, h // 4:3 * h // 4, w // 4:3 * w // 4] = 0. + mask = mask[:, None, ...] + with ema_scope("Plotting Inpaint"): + samples, _ = self.sample_log(cond=c, batch_size=N, ddim=use_ddim, eta=ddim_eta, + ddim_steps=ddim_steps, x0=z[:N], mask=mask) + x_samples = self.decode_first_stage(samples.to(self.device)) + log["samples_inpainting"] = x_samples + log["mask"] = mask + + # outpaint + mask = 1. - mask + with ema_scope("Plotting Outpaint"): + samples, _ = self.sample_log(cond=c, batch_size=N, ddim=use_ddim, eta=ddim_eta, + ddim_steps=ddim_steps, x0=z[:N], mask=mask) + x_samples = self.decode_first_stage(samples.to(self.device)) + log["samples_outpainting"] = x_samples + + if plot_progressive_rows: + with ema_scope("Plotting Progressives"): + img, progressives = self.progressive_denoising(c, + shape=(self.channels, self.image_size, self.image_size), + batch_size=N) + prog_row = self._get_denoise_row_from_list(progressives, desc="Progressive Generation") + log["progressive_row"] = prog_row + + if return_keys: + if np.intersect1d(list(log.keys()), return_keys).shape[0] == 0: + return log + else: + return {key: log[key] for key in return_keys} + return log + + def configure_optimizers(self): + lr = self.learning_rate + params = list(self.model.parameters()) + if self.cond_stage_trainable: + print(f"{self.__class__.__name__}: Also optimizing conditioner params!") + params = params + list(self.cond_stage_model.parameters()) + if self.learn_logvar: + print('Diffusion model optimizing logvar') + params.append(self.logvar) + opt = torch.optim.AdamW(params, lr=lr) + if self.use_scheduler: + assert 'target' in self.scheduler_config + scheduler = instantiate_from_config(self.scheduler_config) + + print("Setting up LambdaLR scheduler...") + scheduler = [ + { + 'scheduler': LambdaLR(opt, lr_lambda=scheduler.schedule), + 'interval': 'step', + 'frequency': 1 + }] + return [opt], scheduler + return opt + + @torch.no_grad() + def to_rgb(self, x): + x = x.float() + if not hasattr(self, "colorize"): + self.colorize = torch.randn(3, x.shape[1], 1, 1).to(x) + x = nn.functional.conv2d(x, weight=self.colorize) + x = 2. * (x - x.min()) / (x.max() - x.min()) - 1. + return x + + +class DiffusionWrapper(pl.LightningModule): + def __init__(self, diff_model_config, conditioning_key): + super().__init__() + self.sequential_cross_attn = diff_model_config.pop("sequential_crossattn", False) + self.diffusion_model = instantiate_from_config(diff_model_config) + self.conditioning_key = conditioning_key + assert self.conditioning_key in [None, 'concat', 'crossattn', 'hybrid', 'adm', 'hybrid-adm', 'crossattn-adm'] + + def forward(self, x, t, c_concat: list = None, c_crossattn: list = None, c_adm=None): + if self.conditioning_key is None: + out = self.diffusion_model(x, t) + elif self.conditioning_key == 'concat': + xc = torch.cat([x] + c_concat, dim=1) + out = self.diffusion_model(xc, t) + elif self.conditioning_key == 'crossattn': + if not self.sequential_cross_attn: + cc = torch.cat(c_crossattn, 1) + else: + cc = c_crossattn + if hasattr(self, "scripted_diffusion_model"): + # TorchScript changes names of the arguments + # with argument cc defined as context=cc scripted model will produce + # an error: RuntimeError: forward() is missing value for argument 'argument_3'. + out = self.scripted_diffusion_model(x, t, cc) + else: + out = self.diffusion_model(x, t, context=cc) + elif self.conditioning_key == 'hybrid': + xc = torch.cat([x] + c_concat, dim=1) + cc = torch.cat(c_crossattn, 1) + out = self.diffusion_model(xc, t, context=cc) + elif self.conditioning_key == 'hybrid-adm': + assert c_adm is not None + xc = torch.cat([x] + c_concat, dim=1) + cc = torch.cat(c_crossattn, 1) + out = self.diffusion_model(xc, t, context=cc, y=c_adm) + elif self.conditioning_key == 'crossattn-adm': + assert c_adm is not None + cc = torch.cat(c_crossattn, 1) + out = self.diffusion_model(x, t, context=cc, y=c_adm) + elif self.conditioning_key == 'adm': + cc = c_crossattn[0] + out = self.diffusion_model(x, t, y=cc) + else: + raise NotImplementedError() + + return out + + +class LatentUpscaleDiffusion(LatentDiffusion): + def __init__(self, *args, low_scale_config, low_scale_key="LR", noise_level_key=None, **kwargs): + super().__init__(*args, **kwargs) + # assumes that neither the cond_stage nor the low_scale_model contain trainable params + assert not self.cond_stage_trainable + self.instantiate_low_stage(low_scale_config) + self.low_scale_key = low_scale_key + self.noise_level_key = noise_level_key + + def instantiate_low_stage(self, config): + model = instantiate_from_config(config) + self.low_scale_model = model.eval() + self.low_scale_model.train = disabled_train + for param in self.low_scale_model.parameters(): + param.requires_grad = False + + @torch.no_grad() + def get_input(self, batch, k, cond_key=None, bs=None, log_mode=False): + if not log_mode: + z, c = super().get_input(batch, k, force_c_encode=True, bs=bs) + else: + z, c, x, xrec, xc = super().get_input(batch, self.first_stage_key, return_first_stage_outputs=True, + force_c_encode=True, return_original_cond=True, bs=bs) + x_low = batch[self.low_scale_key][:bs] + x_low = rearrange(x_low, 'b h w c -> b c h w') + x_low = x_low.to(memory_format=torch.contiguous_format).float() + zx, noise_level = self.low_scale_model(x_low) + if self.noise_level_key is not None: + # get noise level from batch instead, e.g. when extracting a custom noise level for bsr + raise NotImplementedError('TODO') + + all_conds = {"c_concat": [zx], "c_crossattn": [c], "c_adm": noise_level} + if log_mode: + # TODO: maybe disable if too expensive + x_low_rec = self.low_scale_model.decode(zx) + return z, all_conds, x, xrec, xc, x_low, x_low_rec, noise_level + return z, all_conds + + @torch.no_grad() + def log_images(self, batch, N=8, n_row=4, sample=True, ddim_steps=200, ddim_eta=1., return_keys=None, + plot_denoise_rows=False, plot_progressive_rows=True, plot_diffusion_rows=True, + unconditional_guidance_scale=1., unconditional_guidance_label=None, use_ema_scope=True, + **kwargs): + ema_scope = self.ema_scope if use_ema_scope else nullcontext + use_ddim = ddim_steps is not None + + log = dict() + z, c, x, xrec, xc, x_low, x_low_rec, noise_level = self.get_input(batch, self.first_stage_key, bs=N, + log_mode=True) + N = min(x.shape[0], N) + n_row = min(x.shape[0], n_row) + log["inputs"] = x + log["reconstruction"] = xrec + log["x_lr"] = x_low + log[f"x_lr_rec_@noise_levels{'-'.join(map(lambda x: str(x), list(noise_level.cpu().numpy())))}"] = x_low_rec + if self.model.conditioning_key is not None: + if hasattr(self.cond_stage_model, "decode"): + xc = self.cond_stage_model.decode(c) + log["conditioning"] = xc + elif self.cond_stage_key in ["caption", "txt"]: + xc = log_txt_as_img((x.shape[2], x.shape[3]), batch[self.cond_stage_key], size=x.shape[2] // 25) + log["conditioning"] = xc + elif self.cond_stage_key in ['class_label', 'cls']: + xc = log_txt_as_img((x.shape[2], x.shape[3]), batch["human_label"], size=x.shape[2] // 25) + log['conditioning'] = xc + elif isimage(xc): + log["conditioning"] = xc + if ismap(xc): + log["original_conditioning"] = self.to_rgb(xc) + + if plot_diffusion_rows: + # get diffusion row + diffusion_row = list() + z_start = z[:n_row] + for t in range(self.num_timesteps): + if t % self.log_every_t == 0 or t == self.num_timesteps - 1: + t = repeat(torch.tensor([t]), '1 -> b', b=n_row) + t = t.to(self.device).long() + noise = torch.randn_like(z_start) + z_noisy = self.q_sample(x_start=z_start, t=t, noise=noise) + diffusion_row.append(self.decode_first_stage(z_noisy)) + + diffusion_row = torch.stack(diffusion_row) # n_log_step, n_row, C, H, W + diffusion_grid = rearrange(diffusion_row, 'n b c h w -> b n c h w') + diffusion_grid = rearrange(diffusion_grid, 'b n c h w -> (b n) c h w') + diffusion_grid = make_grid(diffusion_grid, nrow=diffusion_row.shape[0]) + log["diffusion_row"] = diffusion_grid + + if sample: + # get denoise row + with ema_scope("Sampling"): + samples, z_denoise_row = self.sample_log(cond=c, batch_size=N, ddim=use_ddim, + ddim_steps=ddim_steps, eta=ddim_eta) + # samples, z_denoise_row = self.sample(cond=c, batch_size=N, return_intermediates=True) + x_samples = self.decode_first_stage(samples) + log["samples"] = x_samples + if plot_denoise_rows: + denoise_grid = self._get_denoise_row_from_list(z_denoise_row) + log["denoise_row"] = denoise_grid + + if unconditional_guidance_scale > 1.0: + uc_tmp = self.get_unconditional_conditioning(N, unconditional_guidance_label) + # TODO explore better "unconditional" choices for the other keys + # maybe guide away from empty text label and highest noise level and maximally degraded zx? + uc = dict() + for k in c: + if k == "c_crossattn": + assert isinstance(c[k], list) and len(c[k]) == 1 + uc[k] = [uc_tmp] + elif k == "c_adm": # todo: only run with text-based guidance? + assert isinstance(c[k], torch.Tensor) + #uc[k] = torch.ones_like(c[k]) * self.low_scale_model.max_noise_level + uc[k] = c[k] + elif isinstance(c[k], list): + uc[k] = [c[k][i] for i in range(len(c[k]))] + else: + uc[k] = c[k] + + with ema_scope("Sampling with classifier-free guidance"): + samples_cfg, _ = self.sample_log(cond=c, batch_size=N, ddim=use_ddim, + ddim_steps=ddim_steps, eta=ddim_eta, + unconditional_guidance_scale=unconditional_guidance_scale, + unconditional_conditioning=uc, + ) + x_samples_cfg = self.decode_first_stage(samples_cfg) + log[f"samples_cfg_scale_{unconditional_guidance_scale:.2f}"] = x_samples_cfg + + if plot_progressive_rows: + with ema_scope("Plotting Progressives"): + img, progressives = self.progressive_denoising(c, + shape=(self.channels, self.image_size, self.image_size), + batch_size=N) + prog_row = self._get_denoise_row_from_list(progressives, desc="Progressive Generation") + log["progressive_row"] = prog_row + + return log + + +class LatentFinetuneDiffusion(LatentDiffusion): + """ + Basis for different finetunas, such as inpainting or depth2image + To disable finetuning mode, set finetune_keys to None + """ + + def __init__(self, + concat_keys: tuple, + finetune_keys=("model.diffusion_model.input_blocks.0.0.weight", + "model_ema.diffusion_modelinput_blocks00weight" + ), + keep_finetune_dims=4, + # if model was trained without concat mode before and we would like to keep these channels + c_concat_log_start=None, # to log reconstruction of c_concat codes + c_concat_log_end=None, + *args, **kwargs + ): + ckpt_path = kwargs.pop("ckpt_path", None) + ignore_keys = kwargs.pop("ignore_keys", list()) + super().__init__(*args, **kwargs) + self.finetune_keys = finetune_keys + self.concat_keys = concat_keys + self.keep_dims = keep_finetune_dims + self.c_concat_log_start = c_concat_log_start + self.c_concat_log_end = c_concat_log_end + if exists(self.finetune_keys): assert exists(ckpt_path), 'can only finetune from a given checkpoint' + if exists(ckpt_path): + self.init_from_ckpt(ckpt_path, ignore_keys) + + def init_from_ckpt(self, path, ignore_keys=list(), only_model=False): + sd = torch.load(path, map_location="cpu") + if "state_dict" in list(sd.keys()): + sd = sd["state_dict"] + keys = list(sd.keys()) + for k in keys: + for ik in ignore_keys: + if k.startswith(ik): + print("Deleting key {} from state_dict.".format(k)) + del sd[k] + + # make it explicit, finetune by including extra input channels + if exists(self.finetune_keys) and k in self.finetune_keys: + new_entry = None + for name, param in self.named_parameters(): + if name in self.finetune_keys: + print( + f"modifying key '{name}' and keeping its original {self.keep_dims} (channels) dimensions only") + new_entry = torch.zeros_like(param) # zero init + assert exists(new_entry), 'did not find matching parameter to modify' + new_entry[:, :self.keep_dims, ...] = sd[k] + sd[k] = new_entry + + missing, unexpected = self.load_state_dict(sd, strict=False) if not only_model else self.model.load_state_dict( + sd, strict=False) + print(f"Restored from {path} with {len(missing)} missing and {len(unexpected)} unexpected keys") + if len(missing) > 0: + print(f"Missing Keys: {missing}") + if len(unexpected) > 0: + print(f"Unexpected Keys: {unexpected}") + + @torch.no_grad() + def log_images(self, batch, N=8, n_row=4, sample=True, ddim_steps=200, ddim_eta=1., return_keys=None, + quantize_denoised=True, inpaint=True, plot_denoise_rows=False, plot_progressive_rows=True, + plot_diffusion_rows=True, unconditional_guidance_scale=1., unconditional_guidance_label=None, + use_ema_scope=True, + **kwargs): + ema_scope = self.ema_scope if use_ema_scope else nullcontext + use_ddim = ddim_steps is not None + + log = dict() + z, c, x, xrec, xc = self.get_input(batch, self.first_stage_key, bs=N, return_first_stage_outputs=True) + c_cat, c = c["c_concat"][0], c["c_crossattn"][0] + N = min(x.shape[0], N) + n_row = min(x.shape[0], n_row) + log["inputs"] = x + log["reconstruction"] = xrec + if self.model.conditioning_key is not None: + if hasattr(self.cond_stage_model, "decode"): + xc = self.cond_stage_model.decode(c) + log["conditioning"] = xc + elif self.cond_stage_key in ["caption", "txt"]: + xc = log_txt_as_img((x.shape[2], x.shape[3]), batch[self.cond_stage_key], size=x.shape[2] // 25) + log["conditioning"] = xc + elif self.cond_stage_key in ['class_label', 'cls']: + xc = log_txt_as_img((x.shape[2], x.shape[3]), batch["human_label"], size=x.shape[2] // 25) + log['conditioning'] = xc + elif isimage(xc): + log["conditioning"] = xc + if ismap(xc): + log["original_conditioning"] = self.to_rgb(xc) + + if not (self.c_concat_log_start is None and self.c_concat_log_end is None): + log["c_concat_decoded"] = self.decode_first_stage(c_cat[:, self.c_concat_log_start:self.c_concat_log_end]) + + if plot_diffusion_rows: + # get diffusion row + diffusion_row = list() + z_start = z[:n_row] + for t in range(self.num_timesteps): + if t % self.log_every_t == 0 or t == self.num_timesteps - 1: + t = repeat(torch.tensor([t]), '1 -> b', b=n_row) + t = t.to(self.device).long() + noise = torch.randn_like(z_start) + z_noisy = self.q_sample(x_start=z_start, t=t, noise=noise) + diffusion_row.append(self.decode_first_stage(z_noisy)) + + diffusion_row = torch.stack(diffusion_row) # n_log_step, n_row, C, H, W + diffusion_grid = rearrange(diffusion_row, 'n b c h w -> b n c h w') + diffusion_grid = rearrange(diffusion_grid, 'b n c h w -> (b n) c h w') + diffusion_grid = make_grid(diffusion_grid, nrow=diffusion_row.shape[0]) + log["diffusion_row"] = diffusion_grid + + if sample: + # get denoise row + with ema_scope("Sampling"): + samples, z_denoise_row = self.sample_log(cond={"c_concat": [c_cat], "c_crossattn": [c]}, + batch_size=N, ddim=use_ddim, + ddim_steps=ddim_steps, eta=ddim_eta) + # samples, z_denoise_row = self.sample(cond=c, batch_size=N, return_intermediates=True) + x_samples = self.decode_first_stage(samples) + log["samples"] = x_samples + if plot_denoise_rows: + denoise_grid = self._get_denoise_row_from_list(z_denoise_row) + log["denoise_row"] = denoise_grid + + if unconditional_guidance_scale > 1.0: + uc_cross = self.get_unconditional_conditioning(N, unconditional_guidance_label) + uc_cat = c_cat + uc_full = {"c_concat": [uc_cat], "c_crossattn": [uc_cross]} + with ema_scope("Sampling with classifier-free guidance"): + samples_cfg, _ = self.sample_log(cond={"c_concat": [c_cat], "c_crossattn": [c]}, + batch_size=N, ddim=use_ddim, + ddim_steps=ddim_steps, eta=ddim_eta, + unconditional_guidance_scale=unconditional_guidance_scale, + unconditional_conditioning=uc_full, + ) + x_samples_cfg = self.decode_first_stage(samples_cfg) + log[f"samples_cfg_scale_{unconditional_guidance_scale:.2f}"] = x_samples_cfg + + return log + + +class LatentInpaintDiffusion(LatentFinetuneDiffusion): + """ + can either run as pure inpainting model (only concat mode) or with mixed conditionings, + e.g. mask as concat and text via cross-attn. + To disable finetuning mode, set finetune_keys to None + """ + + def __init__(self, + concat_keys=("mask", "masked_image"), + masked_image_key="masked_image", + *args, **kwargs + ): + super().__init__(concat_keys, *args, **kwargs) + self.masked_image_key = masked_image_key + assert self.masked_image_key in concat_keys + + @torch.no_grad() + def get_input(self, batch, k, cond_key=None, bs=None, return_first_stage_outputs=False): + # note: restricted to non-trainable encoders currently + assert not self.cond_stage_trainable, 'trainable cond stages not yet supported for inpainting' + z, c, x, xrec, xc = super().get_input(batch, self.first_stage_key, return_first_stage_outputs=True, + force_c_encode=True, return_original_cond=True, bs=bs) + + assert exists(self.concat_keys) + c_cat = list() + for ck in self.concat_keys: + cc = rearrange(batch[ck], 'b h w c -> b c h w').to(memory_format=torch.contiguous_format).float() + if bs is not None: + cc = cc[:bs] + cc = cc.to(self.device) + bchw = z.shape + if ck != self.masked_image_key: + cc = torch.nn.functional.interpolate(cc, size=bchw[-2:]) + else: + cc = self.get_first_stage_encoding(self.encode_first_stage(cc)) + c_cat.append(cc) + c_cat = torch.cat(c_cat, dim=1) + all_conds = {"c_concat": [c_cat], "c_crossattn": [c]} + if return_first_stage_outputs: + return z, all_conds, x, xrec, xc + return z, all_conds + + @torch.no_grad() + def log_images(self, *args, **kwargs): + log = super(LatentInpaintDiffusion, self).log_images(*args, **kwargs) + log["masked_image"] = rearrange(args[0]["masked_image"], + 'b h w c -> b c h w').to(memory_format=torch.contiguous_format).float() + return log + + +class LatentDepth2ImageDiffusion(LatentFinetuneDiffusion): + """ + condition on monocular depth estimation + """ + + def __init__(self, depth_stage_config, concat_keys=("midas_in",), *args, **kwargs): + super().__init__(concat_keys=concat_keys, *args, **kwargs) + self.depth_model = instantiate_from_config(depth_stage_config) + self.depth_stage_key = concat_keys[0] + + @torch.no_grad() + def get_input(self, batch, k, cond_key=None, bs=None, return_first_stage_outputs=False): + # note: restricted to non-trainable encoders currently + assert not self.cond_stage_trainable, 'trainable cond stages not yet supported for depth2img' + z, c, x, xrec, xc = super().get_input(batch, self.first_stage_key, return_first_stage_outputs=True, + force_c_encode=True, return_original_cond=True, bs=bs) + + assert exists(self.concat_keys) + assert len(self.concat_keys) == 1 + c_cat = list() + for ck in self.concat_keys: + cc = batch[ck] + if bs is not None: + cc = cc[:bs] + cc = cc.to(self.device) + cc = self.depth_model(cc) + cc = torch.nn.functional.interpolate( + cc, + size=z.shape[2:], + mode="bicubic", + align_corners=False, + ) + + depth_min, depth_max = torch.amin(cc, dim=[1, 2, 3], keepdim=True), torch.amax(cc, dim=[1, 2, 3], + keepdim=True) + cc = 2. * (cc - depth_min) / (depth_max - depth_min + 0.001) - 1. + c_cat.append(cc) + c_cat = torch.cat(c_cat, dim=1) + all_conds = {"c_concat": [c_cat], "c_crossattn": [c]} + if return_first_stage_outputs: + return z, all_conds, x, xrec, xc + return z, all_conds + + @torch.no_grad() + def log_images(self, *args, **kwargs): + log = super().log_images(*args, **kwargs) + depth = self.depth_model(args[0][self.depth_stage_key]) + depth_min, depth_max = torch.amin(depth, dim=[1, 2, 3], keepdim=True), \ + torch.amax(depth, dim=[1, 2, 3], keepdim=True) + log["depth"] = 2. * (depth - depth_min) / (depth_max - depth_min) - 1. + return log + + +class LatentUpscaleFinetuneDiffusion(LatentFinetuneDiffusion): + """ + condition on low-res image (and optionally on some spatial noise augmentation) + """ + def __init__(self, concat_keys=("lr",), reshuffle_patch_size=None, + low_scale_config=None, low_scale_key=None, *args, **kwargs): + super().__init__(concat_keys=concat_keys, *args, **kwargs) + self.reshuffle_patch_size = reshuffle_patch_size + self.low_scale_model = None + if low_scale_config is not None: + print("Initializing a low-scale model") + assert exists(low_scale_key) + self.instantiate_low_stage(low_scale_config) + self.low_scale_key = low_scale_key + + def instantiate_low_stage(self, config): + model = instantiate_from_config(config) + self.low_scale_model = model.eval() + self.low_scale_model.train = disabled_train + for param in self.low_scale_model.parameters(): + param.requires_grad = False + + @torch.no_grad() + def get_input(self, batch, k, cond_key=None, bs=None, return_first_stage_outputs=False): + # note: restricted to non-trainable encoders currently + assert not self.cond_stage_trainable, 'trainable cond stages not yet supported for upscaling-ft' + z, c, x, xrec, xc = super().get_input(batch, self.first_stage_key, return_first_stage_outputs=True, + force_c_encode=True, return_original_cond=True, bs=bs) + + assert exists(self.concat_keys) + assert len(self.concat_keys) == 1 + # optionally make spatial noise_level here + c_cat = list() + noise_level = None + for ck in self.concat_keys: + cc = batch[ck] + cc = rearrange(cc, 'b h w c -> b c h w') + if exists(self.reshuffle_patch_size): + assert isinstance(self.reshuffle_patch_size, int) + cc = rearrange(cc, 'b c (p1 h) (p2 w) -> b (p1 p2 c) h w', + p1=self.reshuffle_patch_size, p2=self.reshuffle_patch_size) + if bs is not None: + cc = cc[:bs] + cc = cc.to(self.device) + if exists(self.low_scale_model) and ck == self.low_scale_key: + cc, noise_level = self.low_scale_model(cc) + c_cat.append(cc) + c_cat = torch.cat(c_cat, dim=1) + if exists(noise_level): + all_conds = {"c_concat": [c_cat], "c_crossattn": [c], "c_adm": noise_level} + else: + all_conds = {"c_concat": [c_cat], "c_crossattn": [c]} + if return_first_stage_outputs: + return z, all_conds, x, xrec, xc + return z, all_conds + + @torch.no_grad() + def log_images(self, *args, **kwargs): + log = super().log_images(*args, **kwargs) + log["lr"] = rearrange(args[0]["lr"], 'b h w c -> b c h w') + return log + + +class ImageEmbeddingConditionedLatentDiffusion(LatentDiffusion): + def __init__(self, embedder_config, embedding_key="jpg", embedding_dropout=0.5, + freeze_embedder=True, noise_aug_config=None, *args, **kwargs): + super().__init__(*args, **kwargs) + self.embed_key = embedding_key + self.embedding_dropout = embedding_dropout + self._init_embedder(embedder_config, freeze_embedder) + self._init_noise_aug(noise_aug_config) + + def _init_embedder(self, config, freeze=True): + embedder = instantiate_from_config(config) + if freeze: + self.embedder = embedder.eval() + self.embedder.train = disabled_train + for param in self.embedder.parameters(): + param.requires_grad = False + + def _init_noise_aug(self, config): + if config is not None: + # use the KARLO schedule for noise augmentation on CLIP image embeddings + noise_augmentor = instantiate_from_config(config) + assert isinstance(noise_augmentor, nn.Module) + noise_augmentor = noise_augmentor.eval() + noise_augmentor.train = disabled_train + self.noise_augmentor = noise_augmentor + else: + self.noise_augmentor = None + + def get_input(self, batch, k, cond_key=None, bs=None, **kwargs): + outputs = LatentDiffusion.get_input(self, batch, k, bs=bs, **kwargs) + z, c = outputs[0], outputs[1] + img = batch[self.embed_key][:bs] + img = rearrange(img, 'b h w c -> b c h w') + c_adm = self.embedder(img) + if self.noise_augmentor is not None: + c_adm, noise_level_emb = self.noise_augmentor(c_adm) + # assume this gives embeddings of noise levels + c_adm = torch.cat((c_adm, noise_level_emb), 1) + if self.training: + c_adm = torch.bernoulli((1. - self.embedding_dropout) * torch.ones(c_adm.shape[0], + device=c_adm.device)[:, None]) * c_adm + all_conds = {"c_crossattn": [c], "c_adm": c_adm} + noutputs = [z, all_conds] + noutputs.extend(outputs[2:]) + return noutputs + + @torch.no_grad() + def log_images(self, batch, N=8, n_row=4, **kwargs): + log = dict() + z, c, x, xrec, xc = self.get_input(batch, self.first_stage_key, bs=N, return_first_stage_outputs=True, + return_original_cond=True) + log["inputs"] = x + log["reconstruction"] = xrec + assert self.model.conditioning_key is not None + assert self.cond_stage_key in ["caption", "txt"] + xc = log_txt_as_img((x.shape[2], x.shape[3]), batch[self.cond_stage_key], size=x.shape[2] // 25) + log["conditioning"] = xc + uc = self.get_unconditional_conditioning(N, kwargs.get('unconditional_guidance_label', '')) + unconditional_guidance_scale = kwargs.get('unconditional_guidance_scale', 5.) + + uc_ = {"c_crossattn": [uc], "c_adm": c["c_adm"]} + ema_scope = self.ema_scope if kwargs.get('use_ema_scope', True) else nullcontext + with ema_scope(f"Sampling"): + samples_cfg, _ = self.sample_log(cond=c, batch_size=N, ddim=True, + ddim_steps=kwargs.get('ddim_steps', 50), eta=kwargs.get('ddim_eta', 0.), + unconditional_guidance_scale=unconditional_guidance_scale, + unconditional_conditioning=uc_, ) + x_samples_cfg = self.decode_first_stage(samples_cfg) + log[f"samplescfg_scale_{unconditional_guidance_scale:.2f}"] = x_samples_cfg + return log diff --git a/ldm/models/diffusion/dpm_solver/__init__.py b/ldm/models/diffusion/dpm_solver/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..7427f38c07530afbab79154ea8aaf88c4bf70a08 --- /dev/null +++ b/ldm/models/diffusion/dpm_solver/__init__.py @@ -0,0 +1 @@ +from .sampler import DPMSolverSampler \ No newline at end of file diff --git a/ldm/models/diffusion/dpm_solver/dpm_solver.py b/ldm/models/diffusion/dpm_solver/dpm_solver.py new file mode 100644 index 0000000000000000000000000000000000000000..da8d41f9c5e93d2f9e07a22aeef9aeb06d0b7dd3 --- /dev/null +++ b/ldm/models/diffusion/dpm_solver/dpm_solver.py @@ -0,0 +1,1163 @@ +import torch +import torch.nn.functional as F +import math +from tqdm import tqdm + + +class NoiseScheduleVP: + def __init__( + self, + schedule='discrete', + betas=None, + alphas_cumprod=None, + continuous_beta_0=0.1, + continuous_beta_1=20., + ): + """Create a wrapper class for the forward SDE (VP type). + *** + Update: We support discrete-time diffusion models by implementing a picewise linear interpolation for log_alpha_t. + We recommend to use schedule='discrete' for the discrete-time diffusion models, especially for high-resolution images. + *** + The forward SDE ensures that the condition distribution q_{t|0}(x_t | x_0) = N ( alpha_t * x_0, sigma_t^2 * I ). + We further define lambda_t = log(alpha_t) - log(sigma_t), which is the half-logSNR (described in the DPM-Solver paper). + Therefore, we implement the functions for computing alpha_t, sigma_t and lambda_t. For t in [0, T], we have: + log_alpha_t = self.marginal_log_mean_coeff(t) + sigma_t = self.marginal_std(t) + lambda_t = self.marginal_lambda(t) + Moreover, as lambda(t) is an invertible function, we also support its inverse function: + t = self.inverse_lambda(lambda_t) + =============================================================== + We support both discrete-time DPMs (trained on n = 0, 1, ..., N-1) and continuous-time DPMs (trained on t in [t_0, T]). + 1. For discrete-time DPMs: + For discrete-time DPMs trained on n = 0, 1, ..., N-1, we convert the discrete steps to continuous time steps by: + t_i = (i + 1) / N + e.g. for N = 1000, we have t_0 = 1e-3 and T = t_{N-1} = 1. + We solve the corresponding diffusion ODE from time T = 1 to time t_0 = 1e-3. + Args: + betas: A `torch.Tensor`. The beta array for the discrete-time DPM. (See the original DDPM paper for details) + alphas_cumprod: A `torch.Tensor`. The cumprod alphas for the discrete-time DPM. (See the original DDPM paper for details) + Note that we always have alphas_cumprod = cumprod(betas). Therefore, we only need to set one of `betas` and `alphas_cumprod`. + **Important**: Please pay special attention for the args for `alphas_cumprod`: + The `alphas_cumprod` is the \hat{alpha_n} arrays in the notations of DDPM. Specifically, DDPMs assume that + q_{t_n | 0}(x_{t_n} | x_0) = N ( \sqrt{\hat{alpha_n}} * x_0, (1 - \hat{alpha_n}) * I ). + Therefore, the notation \hat{alpha_n} is different from the notation alpha_t in DPM-Solver. In fact, we have + alpha_{t_n} = \sqrt{\hat{alpha_n}}, + and + log(alpha_{t_n}) = 0.5 * log(\hat{alpha_n}). + 2. For continuous-time DPMs: + We support two types of VPSDEs: linear (DDPM) and cosine (improved-DDPM). The hyperparameters for the noise + schedule are the default settings in DDPM and improved-DDPM: + Args: + beta_min: A `float` number. The smallest beta for the linear schedule. + beta_max: A `float` number. The largest beta for the linear schedule. + cosine_s: A `float` number. The hyperparameter in the cosine schedule. + cosine_beta_max: A `float` number. The hyperparameter in the cosine schedule. + T: A `float` number. The ending time of the forward process. + =============================================================== + Args: + schedule: A `str`. The noise schedule of the forward SDE. 'discrete' for discrete-time DPMs, + 'linear' or 'cosine' for continuous-time DPMs. + Returns: + A wrapper object of the forward SDE (VP type). + + =============================================================== + Example: + # For discrete-time DPMs, given betas (the beta array for n = 0, 1, ..., N - 1): + >>> ns = NoiseScheduleVP('discrete', betas=betas) + # For discrete-time DPMs, given alphas_cumprod (the \hat{alpha_n} array for n = 0, 1, ..., N - 1): + >>> ns = NoiseScheduleVP('discrete', alphas_cumprod=alphas_cumprod) + # For continuous-time DPMs (VPSDE), linear schedule: + >>> ns = NoiseScheduleVP('linear', continuous_beta_0=0.1, continuous_beta_1=20.) + """ + + if schedule not in ['discrete', 'linear', 'cosine']: + raise ValueError( + "Unsupported noise schedule {}. The schedule needs to be 'discrete' or 'linear' or 'cosine'".format( + schedule)) + + self.schedule = schedule + if schedule == 'discrete': + if betas is not None: + log_alphas = 0.5 * torch.log(1 - betas).cumsum(dim=0) + else: + assert alphas_cumprod is not None + log_alphas = 0.5 * torch.log(alphas_cumprod) + self.total_N = len(log_alphas) + self.T = 1. + self.t_array = torch.linspace(0., 1., self.total_N + 1)[1:].reshape((1, -1)) + self.log_alpha_array = log_alphas.reshape((1, -1,)) + else: + self.total_N = 1000 + self.beta_0 = continuous_beta_0 + self.beta_1 = continuous_beta_1 + self.cosine_s = 0.008 + self.cosine_beta_max = 999. + self.cosine_t_max = math.atan(self.cosine_beta_max * (1. + self.cosine_s) / math.pi) * 2. * ( + 1. + self.cosine_s) / math.pi - self.cosine_s + self.cosine_log_alpha_0 = math.log(math.cos(self.cosine_s / (1. + self.cosine_s) * math.pi / 2.)) + self.schedule = schedule + if schedule == 'cosine': + # For the cosine schedule, T = 1 will have numerical issues. So we manually set the ending time T. + # Note that T = 0.9946 may be not the optimal setting. However, we find it works well. + self.T = 0.9946 + else: + self.T = 1. + + def marginal_log_mean_coeff(self, t): + """ + Compute log(alpha_t) of a given continuous-time label t in [0, T]. + """ + if self.schedule == 'discrete': + return interpolate_fn(t.reshape((-1, 1)), self.t_array.to(t.device), + self.log_alpha_array.to(t.device)).reshape((-1)) + elif self.schedule == 'linear': + return -0.25 * t ** 2 * (self.beta_1 - self.beta_0) - 0.5 * t * self.beta_0 + elif self.schedule == 'cosine': + log_alpha_fn = lambda s: torch.log(torch.cos((s + self.cosine_s) / (1. + self.cosine_s) * math.pi / 2.)) + log_alpha_t = log_alpha_fn(t) - self.cosine_log_alpha_0 + return log_alpha_t + + def marginal_alpha(self, t): + """ + Compute alpha_t of a given continuous-time label t in [0, T]. + """ + return torch.exp(self.marginal_log_mean_coeff(t)) + + def marginal_std(self, t): + """ + Compute sigma_t of a given continuous-time label t in [0, T]. + """ + return torch.sqrt(1. - torch.exp(2. * self.marginal_log_mean_coeff(t))) + + def marginal_lambda(self, t): + """ + Compute lambda_t = log(alpha_t) - log(sigma_t) of a given continuous-time label t in [0, T]. + """ + log_mean_coeff = self.marginal_log_mean_coeff(t) + log_std = 0.5 * torch.log(1. - torch.exp(2. * log_mean_coeff)) + return log_mean_coeff - log_std + + def inverse_lambda(self, lamb): + """ + Compute the continuous-time label t in [0, T] of a given half-logSNR lambda_t. + """ + if self.schedule == 'linear': + tmp = 2. * (self.beta_1 - self.beta_0) * torch.logaddexp(-2. * lamb, torch.zeros((1,)).to(lamb)) + Delta = self.beta_0 ** 2 + tmp + return tmp / (torch.sqrt(Delta) + self.beta_0) / (self.beta_1 - self.beta_0) + elif self.schedule == 'discrete': + log_alpha = -0.5 * torch.logaddexp(torch.zeros((1,)).to(lamb.device), -2. * lamb) + t = interpolate_fn(log_alpha.reshape((-1, 1)), torch.flip(self.log_alpha_array.to(lamb.device), [1]), + torch.flip(self.t_array.to(lamb.device), [1])) + return t.reshape((-1,)) + else: + log_alpha = -0.5 * torch.logaddexp(-2. * lamb, torch.zeros((1,)).to(lamb)) + t_fn = lambda log_alpha_t: torch.arccos(torch.exp(log_alpha_t + self.cosine_log_alpha_0)) * 2. * ( + 1. + self.cosine_s) / math.pi - self.cosine_s + t = t_fn(log_alpha) + return t + + +def model_wrapper( + model, + noise_schedule, + model_type="noise", + model_kwargs={}, + guidance_type="uncond", + condition=None, + unconditional_condition=None, + guidance_scale=1., + classifier_fn=None, + classifier_kwargs={}, +): + """Create a wrapper function for the noise prediction model. + DPM-Solver needs to solve the continuous-time diffusion ODEs. For DPMs trained on discrete-time labels, we need to + firstly wrap the model function to a noise prediction model that accepts the continuous time as the input. + We support four types of the diffusion model by setting `model_type`: + 1. "noise": noise prediction model. (Trained by predicting noise). + 2. "x_start": data prediction model. (Trained by predicting the data x_0 at time 0). + 3. "v": velocity prediction model. (Trained by predicting the velocity). + The "v" prediction is derivation detailed in Appendix D of [1], and is used in Imagen-Video [2]. + [1] Salimans, Tim, and Jonathan Ho. "Progressive distillation for fast sampling of diffusion models." + arXiv preprint arXiv:2202.00512 (2022). + [2] Ho, Jonathan, et al. "Imagen Video: High Definition Video Generation with Diffusion Models." + arXiv preprint arXiv:2210.02303 (2022). + + 4. "score": marginal score function. (Trained by denoising score matching). + Note that the score function and the noise prediction model follows a simple relationship: + ``` + noise(x_t, t) = -sigma_t * score(x_t, t) + ``` + We support three types of guided sampling by DPMs by setting `guidance_type`: + 1. "uncond": unconditional sampling by DPMs. + The input `model` has the following format: + `` + model(x, t_input, **model_kwargs) -> noise | x_start | v | score + `` + 2. "classifier": classifier guidance sampling [3] by DPMs and another classifier. + The input `model` has the following format: + `` + model(x, t_input, **model_kwargs) -> noise | x_start | v | score + `` + The input `classifier_fn` has the following format: + `` + classifier_fn(x, t_input, cond, **classifier_kwargs) -> logits(x, t_input, cond) + `` + [3] P. Dhariwal and A. Q. Nichol, "Diffusion models beat GANs on image synthesis," + in Advances in Neural Information Processing Systems, vol. 34, 2021, pp. 8780-8794. + 3. "classifier-free": classifier-free guidance sampling by conditional DPMs. + The input `model` has the following format: + `` + model(x, t_input, cond, **model_kwargs) -> noise | x_start | v | score + `` + And if cond == `unconditional_condition`, the model output is the unconditional DPM output. + [4] Ho, Jonathan, and Tim Salimans. "Classifier-free diffusion guidance." + arXiv preprint arXiv:2207.12598 (2022). + + The `t_input` is the time label of the model, which may be discrete-time labels (i.e. 0 to 999) + or continuous-time labels (i.e. epsilon to T). + We wrap the model function to accept only `x` and `t_continuous` as inputs, and outputs the predicted noise: + `` + def model_fn(x, t_continuous) -> noise: + t_input = get_model_input_time(t_continuous) + return noise_pred(model, x, t_input, **model_kwargs) + `` + where `t_continuous` is the continuous time labels (i.e. epsilon to T). And we use `model_fn` for DPM-Solver. + =============================================================== + Args: + model: A diffusion model with the corresponding format described above. + noise_schedule: A noise schedule object, such as NoiseScheduleVP. + model_type: A `str`. The parameterization type of the diffusion model. + "noise" or "x_start" or "v" or "score". + model_kwargs: A `dict`. A dict for the other inputs of the model function. + guidance_type: A `str`. The type of the guidance for sampling. + "uncond" or "classifier" or "classifier-free". + condition: A pytorch tensor. The condition for the guided sampling. + Only used for "classifier" or "classifier-free" guidance type. + unconditional_condition: A pytorch tensor. The condition for the unconditional sampling. + Only used for "classifier-free" guidance type. + guidance_scale: A `float`. The scale for the guided sampling. + classifier_fn: A classifier function. Only used for the classifier guidance. + classifier_kwargs: A `dict`. A dict for the other inputs of the classifier function. + Returns: + A noise prediction model that accepts the noised data and the continuous time as the inputs. + """ + + def get_model_input_time(t_continuous): + """ + Convert the continuous-time `t_continuous` (in [epsilon, T]) to the model input time. + For discrete-time DPMs, we convert `t_continuous` in [1 / N, 1] to `t_input` in [0, 1000 * (N - 1) / N]. + For continuous-time DPMs, we just use `t_continuous`. + """ + if noise_schedule.schedule == 'discrete': + return (t_continuous - 1. / noise_schedule.total_N) * 1000. + else: + return t_continuous + + def noise_pred_fn(x, t_continuous, cond=None): + if t_continuous.reshape((-1,)).shape[0] == 1: + t_continuous = t_continuous.expand((x.shape[0])) + t_input = get_model_input_time(t_continuous) + if cond is None: + output = model(x, t_input, **model_kwargs) + else: + output = model(x, t_input, cond, **model_kwargs) + if model_type == "noise": + return output + elif model_type == "x_start": + alpha_t, sigma_t = noise_schedule.marginal_alpha(t_continuous), noise_schedule.marginal_std(t_continuous) + dims = x.dim() + return (x - expand_dims(alpha_t, dims) * output) / expand_dims(sigma_t, dims) + elif model_type == "v": + alpha_t, sigma_t = noise_schedule.marginal_alpha(t_continuous), noise_schedule.marginal_std(t_continuous) + dims = x.dim() + return expand_dims(alpha_t, dims) * output + expand_dims(sigma_t, dims) * x + elif model_type == "score": + sigma_t = noise_schedule.marginal_std(t_continuous) + dims = x.dim() + return -expand_dims(sigma_t, dims) * output + + def cond_grad_fn(x, t_input): + """ + Compute the gradient of the classifier, i.e. nabla_{x} log p_t(cond | x_t). + """ + with torch.enable_grad(): + x_in = x.detach().requires_grad_(True) + log_prob = classifier_fn(x_in, t_input, condition, **classifier_kwargs) + return torch.autograd.grad(log_prob.sum(), x_in)[0] + + def model_fn(x, t_continuous): + """ + The noise predicition model function that is used for DPM-Solver. + """ + if t_continuous.reshape((-1,)).shape[0] == 1: + t_continuous = t_continuous.expand((x.shape[0])) + if guidance_type == "uncond": + return noise_pred_fn(x, t_continuous) + elif guidance_type == "classifier": + assert classifier_fn is not None + t_input = get_model_input_time(t_continuous) + cond_grad = cond_grad_fn(x, t_input) + sigma_t = noise_schedule.marginal_std(t_continuous) + noise = noise_pred_fn(x, t_continuous) + return noise - guidance_scale * expand_dims(sigma_t, dims=cond_grad.dim()) * cond_grad + elif guidance_type == "classifier-free": + if guidance_scale == 1. or unconditional_condition is None: + return noise_pred_fn(x, t_continuous, cond=condition) + else: + x_in = torch.cat([x] * 2) + t_in = torch.cat([t_continuous] * 2) + if isinstance(condition, dict): + assert isinstance(unconditional_condition, dict) + c_in = dict() + for k in condition: + if isinstance(condition[k], list): + c_in[k] = [torch.cat([unconditional_condition[k][i], condition[k][i]]) for i in range(len(condition[k]))] + else: + c_in[k] = torch.cat([unconditional_condition[k], condition[k]]) + else: + c_in = torch.cat([unconditional_condition, condition]) + noise_uncond, noise = noise_pred_fn(x_in, t_in, cond=c_in).chunk(2) + return noise_uncond + guidance_scale * (noise - noise_uncond) + + assert model_type in ["noise", "x_start", "v"] + assert guidance_type in ["uncond", "classifier", "classifier-free"] + return model_fn + + +class DPM_Solver: + def __init__(self, model_fn, noise_schedule, predict_x0=False, thresholding=False, max_val=1.): + """Construct a DPM-Solver. + We support both the noise prediction model ("predicting epsilon") and the data prediction model ("predicting x0"). + If `predict_x0` is False, we use the solver for the noise prediction model (DPM-Solver). + If `predict_x0` is True, we use the solver for the data prediction model (DPM-Solver++). + In such case, we further support the "dynamic thresholding" in [1] when `thresholding` is True. + The "dynamic thresholding" can greatly improve the sample quality for pixel-space DPMs with large guidance scales. + Args: + model_fn: A noise prediction model function which accepts the continuous-time input (t in [epsilon, T]): + `` + def model_fn(x, t_continuous): + return noise + `` + noise_schedule: A noise schedule object, such as NoiseScheduleVP. + predict_x0: A `bool`. If true, use the data prediction model; else, use the noise prediction model. + thresholding: A `bool`. Valid when `predict_x0` is True. Whether to use the "dynamic thresholding" in [1]. + max_val: A `float`. Valid when both `predict_x0` and `thresholding` are True. The max value for thresholding. + + [1] Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022b. + """ + self.model = model_fn + self.noise_schedule = noise_schedule + self.predict_x0 = predict_x0 + self.thresholding = thresholding + self.max_val = max_val + + def noise_prediction_fn(self, x, t): + """ + Return the noise prediction model. + """ + return self.model(x, t) + + def data_prediction_fn(self, x, t): + """ + Return the data prediction model (with thresholding). + """ + noise = self.noise_prediction_fn(x, t) + dims = x.dim() + alpha_t, sigma_t = self.noise_schedule.marginal_alpha(t), self.noise_schedule.marginal_std(t) + x0 = (x - expand_dims(sigma_t, dims) * noise) / expand_dims(alpha_t, dims) + if self.thresholding: + p = 0.995 # A hyperparameter in the paper of "Imagen" [1]. + s = torch.quantile(torch.abs(x0).reshape((x0.shape[0], -1)), p, dim=1) + s = expand_dims(torch.maximum(s, self.max_val * torch.ones_like(s).to(s.device)), dims) + x0 = torch.clamp(x0, -s, s) / s + return x0 + + def model_fn(self, x, t): + """ + Convert the model to the noise prediction model or the data prediction model. + """ + if self.predict_x0: + return self.data_prediction_fn(x, t) + else: + return self.noise_prediction_fn(x, t) + + def get_time_steps(self, skip_type, t_T, t_0, N, device): + """Compute the intermediate time steps for sampling. + Args: + skip_type: A `str`. The type for the spacing of the time steps. We support three types: + - 'logSNR': uniform logSNR for the time steps. + - 'time_uniform': uniform time for the time steps. (**Recommended for high-resolutional data**.) + - 'time_quadratic': quadratic time for the time steps. (Used in DDIM for low-resolutional data.) + t_T: A `float`. The starting time of the sampling (default is T). + t_0: A `float`. The ending time of the sampling (default is epsilon). + N: A `int`. The total number of the spacing of the time steps. + device: A torch device. + Returns: + A pytorch tensor of the time steps, with the shape (N + 1,). + """ + if skip_type == 'logSNR': + lambda_T = self.noise_schedule.marginal_lambda(torch.tensor(t_T).to(device)) + lambda_0 = self.noise_schedule.marginal_lambda(torch.tensor(t_0).to(device)) + logSNR_steps = torch.linspace(lambda_T.cpu().item(), lambda_0.cpu().item(), N + 1).to(device) + return self.noise_schedule.inverse_lambda(logSNR_steps) + elif skip_type == 'time_uniform': + return torch.linspace(t_T, t_0, N + 1).to(device) + elif skip_type == 'time_quadratic': + t_order = 2 + t = torch.linspace(t_T ** (1. / t_order), t_0 ** (1. / t_order), N + 1).pow(t_order).to(device) + return t + else: + raise ValueError( + "Unsupported skip_type {}, need to be 'logSNR' or 'time_uniform' or 'time_quadratic'".format(skip_type)) + + def get_orders_and_timesteps_for_singlestep_solver(self, steps, order, skip_type, t_T, t_0, device): + """ + Get the order of each step for sampling by the singlestep DPM-Solver. + We combine both DPM-Solver-1,2,3 to use all the function evaluations, which is named as "DPM-Solver-fast". + Given a fixed number of function evaluations by `steps`, the sampling procedure by DPM-Solver-fast is: + - If order == 1: + We take `steps` of DPM-Solver-1 (i.e. DDIM). + - If order == 2: + - Denote K = (steps // 2). We take K or (K + 1) intermediate time steps for sampling. + - If steps % 2 == 0, we use K steps of DPM-Solver-2. + - If steps % 2 == 1, we use K steps of DPM-Solver-2 and 1 step of DPM-Solver-1. + - If order == 3: + - Denote K = (steps // 3 + 1). We take K intermediate time steps for sampling. + - If steps % 3 == 0, we use (K - 2) steps of DPM-Solver-3, and 1 step of DPM-Solver-2 and 1 step of DPM-Solver-1. + - If steps % 3 == 1, we use (K - 1) steps of DPM-Solver-3 and 1 step of DPM-Solver-1. + - If steps % 3 == 2, we use (K - 1) steps of DPM-Solver-3 and 1 step of DPM-Solver-2. + ============================================ + Args: + order: A `int`. The max order for the solver (2 or 3). + steps: A `int`. The total number of function evaluations (NFE). + skip_type: A `str`. The type for the spacing of the time steps. We support three types: + - 'logSNR': uniform logSNR for the time steps. + - 'time_uniform': uniform time for the time steps. (**Recommended for high-resolutional data**.) + - 'time_quadratic': quadratic time for the time steps. (Used in DDIM for low-resolutional data.) + t_T: A `float`. The starting time of the sampling (default is T). + t_0: A `float`. The ending time of the sampling (default is epsilon). + device: A torch device. + Returns: + orders: A list of the solver order of each step. + """ + if order == 3: + K = steps // 3 + 1 + if steps % 3 == 0: + orders = [3, ] * (K - 2) + [2, 1] + elif steps % 3 == 1: + orders = [3, ] * (K - 1) + [1] + else: + orders = [3, ] * (K - 1) + [2] + elif order == 2: + if steps % 2 == 0: + K = steps // 2 + orders = [2, ] * K + else: + K = steps // 2 + 1 + orders = [2, ] * (K - 1) + [1] + elif order == 1: + K = 1 + orders = [1, ] * steps + else: + raise ValueError("'order' must be '1' or '2' or '3'.") + if skip_type == 'logSNR': + # To reproduce the results in DPM-Solver paper + timesteps_outer = self.get_time_steps(skip_type, t_T, t_0, K, device) + else: + timesteps_outer = self.get_time_steps(skip_type, t_T, t_0, steps, device)[ + torch.cumsum(torch.tensor([0, ] + orders)).to(device)] + return timesteps_outer, orders + + def denoise_to_zero_fn(self, x, s): + """ + Denoise at the final step, which is equivalent to solve the ODE from lambda_s to infty by first-order discretization. + """ + return self.data_prediction_fn(x, s) + + def dpm_solver_first_update(self, x, s, t, model_s=None, return_intermediate=False): + """ + DPM-Solver-1 (equivalent to DDIM) from time `s` to time `t`. + Args: + x: A pytorch tensor. The initial value at time `s`. + s: A pytorch tensor. The starting time, with the shape (x.shape[0],). + t: A pytorch tensor. The ending time, with the shape (x.shape[0],). + model_s: A pytorch tensor. The model function evaluated at time `s`. + If `model_s` is None, we evaluate the model by `x` and `s`; otherwise we directly use it. + return_intermediate: A `bool`. If true, also return the model value at time `s`. + Returns: + x_t: A pytorch tensor. The approximated solution at time `t`. + """ + ns = self.noise_schedule + dims = x.dim() + lambda_s, lambda_t = ns.marginal_lambda(s), ns.marginal_lambda(t) + h = lambda_t - lambda_s + log_alpha_s, log_alpha_t = ns.marginal_log_mean_coeff(s), ns.marginal_log_mean_coeff(t) + sigma_s, sigma_t = ns.marginal_std(s), ns.marginal_std(t) + alpha_t = torch.exp(log_alpha_t) + + if self.predict_x0: + phi_1 = torch.expm1(-h) + if model_s is None: + model_s = self.model_fn(x, s) + x_t = ( + expand_dims(sigma_t / sigma_s, dims) * x + - expand_dims(alpha_t * phi_1, dims) * model_s + ) + if return_intermediate: + return x_t, {'model_s': model_s} + else: + return x_t + else: + phi_1 = torch.expm1(h) + if model_s is None: + model_s = self.model_fn(x, s) + x_t = ( + expand_dims(torch.exp(log_alpha_t - log_alpha_s), dims) * x + - expand_dims(sigma_t * phi_1, dims) * model_s + ) + if return_intermediate: + return x_t, {'model_s': model_s} + else: + return x_t + + def singlestep_dpm_solver_second_update(self, x, s, t, r1=0.5, model_s=None, return_intermediate=False, + solver_type='dpm_solver'): + """ + Singlestep solver DPM-Solver-2 from time `s` to time `t`. + Args: + x: A pytorch tensor. The initial value at time `s`. + s: A pytorch tensor. The starting time, with the shape (x.shape[0],). + t: A pytorch tensor. The ending time, with the shape (x.shape[0],). + r1: A `float`. The hyperparameter of the second-order solver. + model_s: A pytorch tensor. The model function evaluated at time `s`. + If `model_s` is None, we evaluate the model by `x` and `s`; otherwise we directly use it. + return_intermediate: A `bool`. If true, also return the model value at time `s` and `s1` (the intermediate time). + solver_type: either 'dpm_solver' or 'taylor'. The type for the high-order solvers. + The type slightly impacts the performance. We recommend to use 'dpm_solver' type. + Returns: + x_t: A pytorch tensor. The approximated solution at time `t`. + """ + if solver_type not in ['dpm_solver', 'taylor']: + raise ValueError("'solver_type' must be either 'dpm_solver' or 'taylor', got {}".format(solver_type)) + if r1 is None: + r1 = 0.5 + ns = self.noise_schedule + dims = x.dim() + lambda_s, lambda_t = ns.marginal_lambda(s), ns.marginal_lambda(t) + h = lambda_t - lambda_s + lambda_s1 = lambda_s + r1 * h + s1 = ns.inverse_lambda(lambda_s1) + log_alpha_s, log_alpha_s1, log_alpha_t = ns.marginal_log_mean_coeff(s), ns.marginal_log_mean_coeff( + s1), ns.marginal_log_mean_coeff(t) + sigma_s, sigma_s1, sigma_t = ns.marginal_std(s), ns.marginal_std(s1), ns.marginal_std(t) + alpha_s1, alpha_t = torch.exp(log_alpha_s1), torch.exp(log_alpha_t) + + if self.predict_x0: + phi_11 = torch.expm1(-r1 * h) + phi_1 = torch.expm1(-h) + + if model_s is None: + model_s = self.model_fn(x, s) + x_s1 = ( + expand_dims(sigma_s1 / sigma_s, dims) * x + - expand_dims(alpha_s1 * phi_11, dims) * model_s + ) + model_s1 = self.model_fn(x_s1, s1) + if solver_type == 'dpm_solver': + x_t = ( + expand_dims(sigma_t / sigma_s, dims) * x + - expand_dims(alpha_t * phi_1, dims) * model_s + - (0.5 / r1) * expand_dims(alpha_t * phi_1, dims) * (model_s1 - model_s) + ) + elif solver_type == 'taylor': + x_t = ( + expand_dims(sigma_t / sigma_s, dims) * x + - expand_dims(alpha_t * phi_1, dims) * model_s + + (1. / r1) * expand_dims(alpha_t * ((torch.exp(-h) - 1.) / h + 1.), dims) * ( + model_s1 - model_s) + ) + else: + phi_11 = torch.expm1(r1 * h) + phi_1 = torch.expm1(h) + + if model_s is None: + model_s = self.model_fn(x, s) + x_s1 = ( + expand_dims(torch.exp(log_alpha_s1 - log_alpha_s), dims) * x + - expand_dims(sigma_s1 * phi_11, dims) * model_s + ) + model_s1 = self.model_fn(x_s1, s1) + if solver_type == 'dpm_solver': + x_t = ( + expand_dims(torch.exp(log_alpha_t - log_alpha_s), dims) * x + - expand_dims(sigma_t * phi_1, dims) * model_s + - (0.5 / r1) * expand_dims(sigma_t * phi_1, dims) * (model_s1 - model_s) + ) + elif solver_type == 'taylor': + x_t = ( + expand_dims(torch.exp(log_alpha_t - log_alpha_s), dims) * x + - expand_dims(sigma_t * phi_1, dims) * model_s + - (1. / r1) * expand_dims(sigma_t * ((torch.exp(h) - 1.) / h - 1.), dims) * (model_s1 - model_s) + ) + if return_intermediate: + return x_t, {'model_s': model_s, 'model_s1': model_s1} + else: + return x_t + + def singlestep_dpm_solver_third_update(self, x, s, t, r1=1. / 3., r2=2. / 3., model_s=None, model_s1=None, + return_intermediate=False, solver_type='dpm_solver'): + """ + Singlestep solver DPM-Solver-3 from time `s` to time `t`. + Args: + x: A pytorch tensor. The initial value at time `s`. + s: A pytorch tensor. The starting time, with the shape (x.shape[0],). + t: A pytorch tensor. The ending time, with the shape (x.shape[0],). + r1: A `float`. The hyperparameter of the third-order solver. + r2: A `float`. The hyperparameter of the third-order solver. + model_s: A pytorch tensor. The model function evaluated at time `s`. + If `model_s` is None, we evaluate the model by `x` and `s`; otherwise we directly use it. + model_s1: A pytorch tensor. The model function evaluated at time `s1` (the intermediate time given by `r1`). + If `model_s1` is None, we evaluate the model at `s1`; otherwise we directly use it. + return_intermediate: A `bool`. If true, also return the model value at time `s`, `s1` and `s2` (the intermediate times). + solver_type: either 'dpm_solver' or 'taylor'. The type for the high-order solvers. + The type slightly impacts the performance. We recommend to use 'dpm_solver' type. + Returns: + x_t: A pytorch tensor. The approximated solution at time `t`. + """ + if solver_type not in ['dpm_solver', 'taylor']: + raise ValueError("'solver_type' must be either 'dpm_solver' or 'taylor', got {}".format(solver_type)) + if r1 is None: + r1 = 1. / 3. + if r2 is None: + r2 = 2. / 3. + ns = self.noise_schedule + dims = x.dim() + lambda_s, lambda_t = ns.marginal_lambda(s), ns.marginal_lambda(t) + h = lambda_t - lambda_s + lambda_s1 = lambda_s + r1 * h + lambda_s2 = lambda_s + r2 * h + s1 = ns.inverse_lambda(lambda_s1) + s2 = ns.inverse_lambda(lambda_s2) + log_alpha_s, log_alpha_s1, log_alpha_s2, log_alpha_t = ns.marginal_log_mean_coeff( + s), ns.marginal_log_mean_coeff(s1), ns.marginal_log_mean_coeff(s2), ns.marginal_log_mean_coeff(t) + sigma_s, sigma_s1, sigma_s2, sigma_t = ns.marginal_std(s), ns.marginal_std(s1), ns.marginal_std( + s2), ns.marginal_std(t) + alpha_s1, alpha_s2, alpha_t = torch.exp(log_alpha_s1), torch.exp(log_alpha_s2), torch.exp(log_alpha_t) + + if self.predict_x0: + phi_11 = torch.expm1(-r1 * h) + phi_12 = torch.expm1(-r2 * h) + phi_1 = torch.expm1(-h) + phi_22 = torch.expm1(-r2 * h) / (r2 * h) + 1. + phi_2 = phi_1 / h + 1. + phi_3 = phi_2 / h - 0.5 + + if model_s is None: + model_s = self.model_fn(x, s) + if model_s1 is None: + x_s1 = ( + expand_dims(sigma_s1 / sigma_s, dims) * x + - expand_dims(alpha_s1 * phi_11, dims) * model_s + ) + model_s1 = self.model_fn(x_s1, s1) + x_s2 = ( + expand_dims(sigma_s2 / sigma_s, dims) * x + - expand_dims(alpha_s2 * phi_12, dims) * model_s + + r2 / r1 * expand_dims(alpha_s2 * phi_22, dims) * (model_s1 - model_s) + ) + model_s2 = self.model_fn(x_s2, s2) + if solver_type == 'dpm_solver': + x_t = ( + expand_dims(sigma_t / sigma_s, dims) * x + - expand_dims(alpha_t * phi_1, dims) * model_s + + (1. / r2) * expand_dims(alpha_t * phi_2, dims) * (model_s2 - model_s) + ) + elif solver_type == 'taylor': + D1_0 = (1. / r1) * (model_s1 - model_s) + D1_1 = (1. / r2) * (model_s2 - model_s) + D1 = (r2 * D1_0 - r1 * D1_1) / (r2 - r1) + D2 = 2. * (D1_1 - D1_0) / (r2 - r1) + x_t = ( + expand_dims(sigma_t / sigma_s, dims) * x + - expand_dims(alpha_t * phi_1, dims) * model_s + + expand_dims(alpha_t * phi_2, dims) * D1 + - expand_dims(alpha_t * phi_3, dims) * D2 + ) + else: + phi_11 = torch.expm1(r1 * h) + phi_12 = torch.expm1(r2 * h) + phi_1 = torch.expm1(h) + phi_22 = torch.expm1(r2 * h) / (r2 * h) - 1. + phi_2 = phi_1 / h - 1. + phi_3 = phi_2 / h - 0.5 + + if model_s is None: + model_s = self.model_fn(x, s) + if model_s1 is None: + x_s1 = ( + expand_dims(torch.exp(log_alpha_s1 - log_alpha_s), dims) * x + - expand_dims(sigma_s1 * phi_11, dims) * model_s + ) + model_s1 = self.model_fn(x_s1, s1) + x_s2 = ( + expand_dims(torch.exp(log_alpha_s2 - log_alpha_s), dims) * x + - expand_dims(sigma_s2 * phi_12, dims) * model_s + - r2 / r1 * expand_dims(sigma_s2 * phi_22, dims) * (model_s1 - model_s) + ) + model_s2 = self.model_fn(x_s2, s2) + if solver_type == 'dpm_solver': + x_t = ( + expand_dims(torch.exp(log_alpha_t - log_alpha_s), dims) * x + - expand_dims(sigma_t * phi_1, dims) * model_s + - (1. / r2) * expand_dims(sigma_t * phi_2, dims) * (model_s2 - model_s) + ) + elif solver_type == 'taylor': + D1_0 = (1. / r1) * (model_s1 - model_s) + D1_1 = (1. / r2) * (model_s2 - model_s) + D1 = (r2 * D1_0 - r1 * D1_1) / (r2 - r1) + D2 = 2. * (D1_1 - D1_0) / (r2 - r1) + x_t = ( + expand_dims(torch.exp(log_alpha_t - log_alpha_s), dims) * x + - expand_dims(sigma_t * phi_1, dims) * model_s + - expand_dims(sigma_t * phi_2, dims) * D1 + - expand_dims(sigma_t * phi_3, dims) * D2 + ) + + if return_intermediate: + return x_t, {'model_s': model_s, 'model_s1': model_s1, 'model_s2': model_s2} + else: + return x_t + + def multistep_dpm_solver_second_update(self, x, model_prev_list, t_prev_list, t, solver_type="dpm_solver"): + """ + Multistep solver DPM-Solver-2 from time `t_prev_list[-1]` to time `t`. + Args: + x: A pytorch tensor. The initial value at time `s`. + model_prev_list: A list of pytorch tensor. The previous computed model values. + t_prev_list: A list of pytorch tensor. The previous times, each time has the shape (x.shape[0],) + t: A pytorch tensor. The ending time, with the shape (x.shape[0],). + solver_type: either 'dpm_solver' or 'taylor'. The type for the high-order solvers. + The type slightly impacts the performance. We recommend to use 'dpm_solver' type. + Returns: + x_t: A pytorch tensor. The approximated solution at time `t`. + """ + if solver_type not in ['dpm_solver', 'taylor']: + raise ValueError("'solver_type' must be either 'dpm_solver' or 'taylor', got {}".format(solver_type)) + ns = self.noise_schedule + dims = x.dim() + model_prev_1, model_prev_0 = model_prev_list + t_prev_1, t_prev_0 = t_prev_list + lambda_prev_1, lambda_prev_0, lambda_t = ns.marginal_lambda(t_prev_1), ns.marginal_lambda( + t_prev_0), ns.marginal_lambda(t) + log_alpha_prev_0, log_alpha_t = ns.marginal_log_mean_coeff(t_prev_0), ns.marginal_log_mean_coeff(t) + sigma_prev_0, sigma_t = ns.marginal_std(t_prev_0), ns.marginal_std(t) + alpha_t = torch.exp(log_alpha_t) + + h_0 = lambda_prev_0 - lambda_prev_1 + h = lambda_t - lambda_prev_0 + r0 = h_0 / h + D1_0 = expand_dims(1. / r0, dims) * (model_prev_0 - model_prev_1) + if self.predict_x0: + if solver_type == 'dpm_solver': + x_t = ( + expand_dims(sigma_t / sigma_prev_0, dims) * x + - expand_dims(alpha_t * (torch.exp(-h) - 1.), dims) * model_prev_0 + - 0.5 * expand_dims(alpha_t * (torch.exp(-h) - 1.), dims) * D1_0 + ) + elif solver_type == 'taylor': + x_t = ( + expand_dims(sigma_t / sigma_prev_0, dims) * x + - expand_dims(alpha_t * (torch.exp(-h) - 1.), dims) * model_prev_0 + + expand_dims(alpha_t * ((torch.exp(-h) - 1.) / h + 1.), dims) * D1_0 + ) + else: + if solver_type == 'dpm_solver': + x_t = ( + expand_dims(torch.exp(log_alpha_t - log_alpha_prev_0), dims) * x + - expand_dims(sigma_t * (torch.exp(h) - 1.), dims) * model_prev_0 + - 0.5 * expand_dims(sigma_t * (torch.exp(h) - 1.), dims) * D1_0 + ) + elif solver_type == 'taylor': + x_t = ( + expand_dims(torch.exp(log_alpha_t - log_alpha_prev_0), dims) * x + - expand_dims(sigma_t * (torch.exp(h) - 1.), dims) * model_prev_0 + - expand_dims(sigma_t * ((torch.exp(h) - 1.) / h - 1.), dims) * D1_0 + ) + return x_t + + def multistep_dpm_solver_third_update(self, x, model_prev_list, t_prev_list, t, solver_type='dpm_solver'): + """ + Multistep solver DPM-Solver-3 from time `t_prev_list[-1]` to time `t`. + Args: + x: A pytorch tensor. The initial value at time `s`. + model_prev_list: A list of pytorch tensor. The previous computed model values. + t_prev_list: A list of pytorch tensor. The previous times, each time has the shape (x.shape[0],) + t: A pytorch tensor. The ending time, with the shape (x.shape[0],). + solver_type: either 'dpm_solver' or 'taylor'. The type for the high-order solvers. + The type slightly impacts the performance. We recommend to use 'dpm_solver' type. + Returns: + x_t: A pytorch tensor. The approximated solution at time `t`. + """ + ns = self.noise_schedule + dims = x.dim() + model_prev_2, model_prev_1, model_prev_0 = model_prev_list + t_prev_2, t_prev_1, t_prev_0 = t_prev_list + lambda_prev_2, lambda_prev_1, lambda_prev_0, lambda_t = ns.marginal_lambda(t_prev_2), ns.marginal_lambda( + t_prev_1), ns.marginal_lambda(t_prev_0), ns.marginal_lambda(t) + log_alpha_prev_0, log_alpha_t = ns.marginal_log_mean_coeff(t_prev_0), ns.marginal_log_mean_coeff(t) + sigma_prev_0, sigma_t = ns.marginal_std(t_prev_0), ns.marginal_std(t) + alpha_t = torch.exp(log_alpha_t) + + h_1 = lambda_prev_1 - lambda_prev_2 + h_0 = lambda_prev_0 - lambda_prev_1 + h = lambda_t - lambda_prev_0 + r0, r1 = h_0 / h, h_1 / h + D1_0 = expand_dims(1. / r0, dims) * (model_prev_0 - model_prev_1) + D1_1 = expand_dims(1. / r1, dims) * (model_prev_1 - model_prev_2) + D1 = D1_0 + expand_dims(r0 / (r0 + r1), dims) * (D1_0 - D1_1) + D2 = expand_dims(1. / (r0 + r1), dims) * (D1_0 - D1_1) + if self.predict_x0: + x_t = ( + expand_dims(sigma_t / sigma_prev_0, dims) * x + - expand_dims(alpha_t * (torch.exp(-h) - 1.), dims) * model_prev_0 + + expand_dims(alpha_t * ((torch.exp(-h) - 1.) / h + 1.), dims) * D1 + - expand_dims(alpha_t * ((torch.exp(-h) - 1. + h) / h ** 2 - 0.5), dims) * D2 + ) + else: + x_t = ( + expand_dims(torch.exp(log_alpha_t - log_alpha_prev_0), dims) * x + - expand_dims(sigma_t * (torch.exp(h) - 1.), dims) * model_prev_0 + - expand_dims(sigma_t * ((torch.exp(h) - 1.) / h - 1.), dims) * D1 + - expand_dims(sigma_t * ((torch.exp(h) - 1. - h) / h ** 2 - 0.5), dims) * D2 + ) + return x_t + + def singlestep_dpm_solver_update(self, x, s, t, order, return_intermediate=False, solver_type='dpm_solver', r1=None, + r2=None): + """ + Singlestep DPM-Solver with the order `order` from time `s` to time `t`. + Args: + x: A pytorch tensor. The initial value at time `s`. + s: A pytorch tensor. The starting time, with the shape (x.shape[0],). + t: A pytorch tensor. The ending time, with the shape (x.shape[0],). + order: A `int`. The order of DPM-Solver. We only support order == 1 or 2 or 3. + return_intermediate: A `bool`. If true, also return the model value at time `s`, `s1` and `s2` (the intermediate times). + solver_type: either 'dpm_solver' or 'taylor'. The type for the high-order solvers. + The type slightly impacts the performance. We recommend to use 'dpm_solver' type. + r1: A `float`. The hyperparameter of the second-order or third-order solver. + r2: A `float`. The hyperparameter of the third-order solver. + Returns: + x_t: A pytorch tensor. The approximated solution at time `t`. + """ + if order == 1: + return self.dpm_solver_first_update(x, s, t, return_intermediate=return_intermediate) + elif order == 2: + return self.singlestep_dpm_solver_second_update(x, s, t, return_intermediate=return_intermediate, + solver_type=solver_type, r1=r1) + elif order == 3: + return self.singlestep_dpm_solver_third_update(x, s, t, return_intermediate=return_intermediate, + solver_type=solver_type, r1=r1, r2=r2) + else: + raise ValueError("Solver order must be 1 or 2 or 3, got {}".format(order)) + + def multistep_dpm_solver_update(self, x, model_prev_list, t_prev_list, t, order, solver_type='dpm_solver'): + """ + Multistep DPM-Solver with the order `order` from time `t_prev_list[-1]` to time `t`. + Args: + x: A pytorch tensor. The initial value at time `s`. + model_prev_list: A list of pytorch tensor. The previous computed model values. + t_prev_list: A list of pytorch tensor. The previous times, each time has the shape (x.shape[0],) + t: A pytorch tensor. The ending time, with the shape (x.shape[0],). + order: A `int`. The order of DPM-Solver. We only support order == 1 or 2 or 3. + solver_type: either 'dpm_solver' or 'taylor'. The type for the high-order solvers. + The type slightly impacts the performance. We recommend to use 'dpm_solver' type. + Returns: + x_t: A pytorch tensor. The approximated solution at time `t`. + """ + if order == 1: + return self.dpm_solver_first_update(x, t_prev_list[-1], t, model_s=model_prev_list[-1]) + elif order == 2: + return self.multistep_dpm_solver_second_update(x, model_prev_list, t_prev_list, t, solver_type=solver_type) + elif order == 3: + return self.multistep_dpm_solver_third_update(x, model_prev_list, t_prev_list, t, solver_type=solver_type) + else: + raise ValueError("Solver order must be 1 or 2 or 3, got {}".format(order)) + + def dpm_solver_adaptive(self, x, order, t_T, t_0, h_init=0.05, atol=0.0078, rtol=0.05, theta=0.9, t_err=1e-5, + solver_type='dpm_solver'): + """ + The adaptive step size solver based on singlestep DPM-Solver. + Args: + x: A pytorch tensor. The initial value at time `t_T`. + order: A `int`. The (higher) order of the solver. We only support order == 2 or 3. + t_T: A `float`. The starting time of the sampling (default is T). + t_0: A `float`. The ending time of the sampling (default is epsilon). + h_init: A `float`. The initial step size (for logSNR). + atol: A `float`. The absolute tolerance of the solver. For image data, the default setting is 0.0078, followed [1]. + rtol: A `float`. The relative tolerance of the solver. The default setting is 0.05. + theta: A `float`. The safety hyperparameter for adapting the step size. The default setting is 0.9, followed [1]. + t_err: A `float`. The tolerance for the time. We solve the diffusion ODE until the absolute error between the + current time and `t_0` is less than `t_err`. The default setting is 1e-5. + solver_type: either 'dpm_solver' or 'taylor'. The type for the high-order solvers. + The type slightly impacts the performance. We recommend to use 'dpm_solver' type. + Returns: + x_0: A pytorch tensor. The approximated solution at time `t_0`. + [1] A. Jolicoeur-Martineau, K. Li, R. Piché-Taillefer, T. Kachman, and I. Mitliagkas, "Gotta go fast when generating data with score-based models," arXiv preprint arXiv:2105.14080, 2021. + """ + ns = self.noise_schedule + s = t_T * torch.ones((x.shape[0],)).to(x) + lambda_s = ns.marginal_lambda(s) + lambda_0 = ns.marginal_lambda(t_0 * torch.ones_like(s).to(x)) + h = h_init * torch.ones_like(s).to(x) + x_prev = x + nfe = 0 + if order == 2: + r1 = 0.5 + lower_update = lambda x, s, t: self.dpm_solver_first_update(x, s, t, return_intermediate=True) + higher_update = lambda x, s, t, **kwargs: self.singlestep_dpm_solver_second_update(x, s, t, r1=r1, + solver_type=solver_type, + **kwargs) + elif order == 3: + r1, r2 = 1. / 3., 2. / 3. + lower_update = lambda x, s, t: self.singlestep_dpm_solver_second_update(x, s, t, r1=r1, + return_intermediate=True, + solver_type=solver_type) + higher_update = lambda x, s, t, **kwargs: self.singlestep_dpm_solver_third_update(x, s, t, r1=r1, r2=r2, + solver_type=solver_type, + **kwargs) + else: + raise ValueError("For adaptive step size solver, order must be 2 or 3, got {}".format(order)) + while torch.abs((s - t_0)).mean() > t_err: + t = ns.inverse_lambda(lambda_s + h) + x_lower, lower_noise_kwargs = lower_update(x, s, t) + x_higher = higher_update(x, s, t, **lower_noise_kwargs) + delta = torch.max(torch.ones_like(x).to(x) * atol, rtol * torch.max(torch.abs(x_lower), torch.abs(x_prev))) + norm_fn = lambda v: torch.sqrt(torch.square(v.reshape((v.shape[0], -1))).mean(dim=-1, keepdim=True)) + E = norm_fn((x_higher - x_lower) / delta).max() + if torch.all(E <= 1.): + x = x_higher + s = t + x_prev = x_lower + lambda_s = ns.marginal_lambda(s) + h = torch.min(theta * h * torch.float_power(E, -1. / order).float(), lambda_0 - lambda_s) + nfe += order + print('adaptive solver nfe', nfe) + return x + + def sample(self, x, steps=20, t_start=None, t_end=None, order=3, skip_type='time_uniform', + method='singlestep', lower_order_final=True, denoise_to_zero=False, solver_type='dpm_solver', + atol=0.0078, rtol=0.05, + ): + """ + Compute the sample at time `t_end` by DPM-Solver, given the initial `x` at time `t_start`. + ===================================================== + We support the following algorithms for both noise prediction model and data prediction model: + - 'singlestep': + Singlestep DPM-Solver (i.e. "DPM-Solver-fast" in the paper), which combines different orders of singlestep DPM-Solver. + We combine all the singlestep solvers with order <= `order` to use up all the function evaluations (steps). + The total number of function evaluations (NFE) == `steps`. + Given a fixed NFE == `steps`, the sampling procedure is: + - If `order` == 1: + - Denote K = steps. We use K steps of DPM-Solver-1 (i.e. DDIM). + - If `order` == 2: + - Denote K = (steps // 2) + (steps % 2). We take K intermediate time steps for sampling. + - If steps % 2 == 0, we use K steps of singlestep DPM-Solver-2. + - If steps % 2 == 1, we use (K - 1) steps of singlestep DPM-Solver-2 and 1 step of DPM-Solver-1. + - If `order` == 3: + - Denote K = (steps // 3 + 1). We take K intermediate time steps for sampling. + - If steps % 3 == 0, we use (K - 2) steps of singlestep DPM-Solver-3, and 1 step of singlestep DPM-Solver-2 and 1 step of DPM-Solver-1. + - If steps % 3 == 1, we use (K - 1) steps of singlestep DPM-Solver-3 and 1 step of DPM-Solver-1. + - If steps % 3 == 2, we use (K - 1) steps of singlestep DPM-Solver-3 and 1 step of singlestep DPM-Solver-2. + - 'multistep': + Multistep DPM-Solver with the order of `order`. The total number of function evaluations (NFE) == `steps`. + We initialize the first `order` values by lower order multistep solvers. + Given a fixed NFE == `steps`, the sampling procedure is: + Denote K = steps. + - If `order` == 1: + - We use K steps of DPM-Solver-1 (i.e. DDIM). + - If `order` == 2: + - We firstly use 1 step of DPM-Solver-1, then use (K - 1) step of multistep DPM-Solver-2. + - If `order` == 3: + - We firstly use 1 step of DPM-Solver-1, then 1 step of multistep DPM-Solver-2, then (K - 2) step of multistep DPM-Solver-3. + - 'singlestep_fixed': + Fixed order singlestep DPM-Solver (i.e. DPM-Solver-1 or singlestep DPM-Solver-2 or singlestep DPM-Solver-3). + We use singlestep DPM-Solver-`order` for `order`=1 or 2 or 3, with total [`steps` // `order`] * `order` NFE. + - 'adaptive': + Adaptive step size DPM-Solver (i.e. "DPM-Solver-12" and "DPM-Solver-23" in the paper). + We ignore `steps` and use adaptive step size DPM-Solver with a higher order of `order`. + You can adjust the absolute tolerance `atol` and the relative tolerance `rtol` to balance the computatation costs + (NFE) and the sample quality. + - If `order` == 2, we use DPM-Solver-12 which combines DPM-Solver-1 and singlestep DPM-Solver-2. + - If `order` == 3, we use DPM-Solver-23 which combines singlestep DPM-Solver-2 and singlestep DPM-Solver-3. + ===================================================== + Some advices for choosing the algorithm: + - For **unconditional sampling** or **guided sampling with small guidance scale** by DPMs: + Use singlestep DPM-Solver ("DPM-Solver-fast" in the paper) with `order = 3`. + e.g. + >>> dpm_solver = DPM_Solver(model_fn, noise_schedule, predict_x0=False) + >>> x_sample = dpm_solver.sample(x, steps=steps, t_start=t_start, t_end=t_end, order=3, + skip_type='time_uniform', method='singlestep') + - For **guided sampling with large guidance scale** by DPMs: + Use multistep DPM-Solver with `predict_x0 = True` and `order = 2`. + e.g. + >>> dpm_solver = DPM_Solver(model_fn, noise_schedule, predict_x0=True) + >>> x_sample = dpm_solver.sample(x, steps=steps, t_start=t_start, t_end=t_end, order=2, + skip_type='time_uniform', method='multistep') + We support three types of `skip_type`: + - 'logSNR': uniform logSNR for the time steps. **Recommended for low-resolutional images** + - 'time_uniform': uniform time for the time steps. **Recommended for high-resolutional images**. + - 'time_quadratic': quadratic time for the time steps. + ===================================================== + Args: + x: A pytorch tensor. The initial value at time `t_start` + e.g. if `t_start` == T, then `x` is a sample from the standard normal distribution. + steps: A `int`. The total number of function evaluations (NFE). + t_start: A `float`. The starting time of the sampling. + If `T` is None, we use self.noise_schedule.T (default is 1.0). + t_end: A `float`. The ending time of the sampling. + If `t_end` is None, we use 1. / self.noise_schedule.total_N. + e.g. if total_N == 1000, we have `t_end` == 1e-3. + For discrete-time DPMs: + - We recommend `t_end` == 1. / self.noise_schedule.total_N. + For continuous-time DPMs: + - We recommend `t_end` == 1e-3 when `steps` <= 15; and `t_end` == 1e-4 when `steps` > 15. + order: A `int`. The order of DPM-Solver. + skip_type: A `str`. The type for the spacing of the time steps. 'time_uniform' or 'logSNR' or 'time_quadratic'. + method: A `str`. The method for sampling. 'singlestep' or 'multistep' or 'singlestep_fixed' or 'adaptive'. + denoise_to_zero: A `bool`. Whether to denoise to time 0 at the final step. + Default is `False`. If `denoise_to_zero` is `True`, the total NFE is (`steps` + 1). + This trick is firstly proposed by DDPM (https://arxiv.org/abs/2006.11239) and + score_sde (https://arxiv.org/abs/2011.13456). Such trick can improve the FID + for diffusion models sampling by diffusion SDEs for low-resolutional images + (such as CIFAR-10). However, we observed that such trick does not matter for + high-resolutional images. As it needs an additional NFE, we do not recommend + it for high-resolutional images. + lower_order_final: A `bool`. Whether to use lower order solvers at the final steps. + Only valid for `method=multistep` and `steps < 15`. We empirically find that + this trick is a key to stabilizing the sampling by DPM-Solver with very few steps + (especially for steps <= 10). So we recommend to set it to be `True`. + solver_type: A `str`. The taylor expansion type for the solver. `dpm_solver` or `taylor`. We recommend `dpm_solver`. + atol: A `float`. The absolute tolerance of the adaptive step size solver. Valid when `method` == 'adaptive'. + rtol: A `float`. The relative tolerance of the adaptive step size solver. Valid when `method` == 'adaptive'. + Returns: + x_end: A pytorch tensor. The approximated solution at time `t_end`. + """ + t_0 = 1. / self.noise_schedule.total_N if t_end is None else t_end + t_T = self.noise_schedule.T if t_start is None else t_start + device = x.device + if method == 'adaptive': + with torch.no_grad(): + x = self.dpm_solver_adaptive(x, order=order, t_T=t_T, t_0=t_0, atol=atol, rtol=rtol, + solver_type=solver_type) + elif method == 'multistep': + assert steps >= order + timesteps = self.get_time_steps(skip_type=skip_type, t_T=t_T, t_0=t_0, N=steps, device=device) + assert timesteps.shape[0] - 1 == steps + with torch.no_grad(): + vec_t = timesteps[0].expand((x.shape[0])) + model_prev_list = [self.model_fn(x, vec_t)] + t_prev_list = [vec_t] + # Init the first `order` values by lower order multistep DPM-Solver. + for init_order in tqdm(range(1, order), desc="DPM init order"): + vec_t = timesteps[init_order].expand(x.shape[0]) + x = self.multistep_dpm_solver_update(x, model_prev_list, t_prev_list, vec_t, init_order, + solver_type=solver_type) + model_prev_list.append(self.model_fn(x, vec_t)) + t_prev_list.append(vec_t) + # Compute the remaining values by `order`-th order multistep DPM-Solver. + for step in tqdm(range(order, steps + 1), desc="DPM multistep"): + vec_t = timesteps[step].expand(x.shape[0]) + if lower_order_final and steps < 15: + step_order = min(order, steps + 1 - step) + else: + step_order = order + x = self.multistep_dpm_solver_update(x, model_prev_list, t_prev_list, vec_t, step_order, + solver_type=solver_type) + for i in range(order - 1): + t_prev_list[i] = t_prev_list[i + 1] + model_prev_list[i] = model_prev_list[i + 1] + t_prev_list[-1] = vec_t + # We do not need to evaluate the final model value. + if step < steps: + model_prev_list[-1] = self.model_fn(x, vec_t) + elif method in ['singlestep', 'singlestep_fixed']: + if method == 'singlestep': + timesteps_outer, orders = self.get_orders_and_timesteps_for_singlestep_solver(steps=steps, order=order, + skip_type=skip_type, + t_T=t_T, t_0=t_0, + device=device) + elif method == 'singlestep_fixed': + K = steps // order + orders = [order, ] * K + timesteps_outer = self.get_time_steps(skip_type=skip_type, t_T=t_T, t_0=t_0, N=K, device=device) + for i, order in enumerate(orders): + t_T_inner, t_0_inner = timesteps_outer[i], timesteps_outer[i + 1] + timesteps_inner = self.get_time_steps(skip_type=skip_type, t_T=t_T_inner.item(), t_0=t_0_inner.item(), + N=order, device=device) + lambda_inner = self.noise_schedule.marginal_lambda(timesteps_inner) + vec_s, vec_t = t_T_inner.tile(x.shape[0]), t_0_inner.tile(x.shape[0]) + h = lambda_inner[-1] - lambda_inner[0] + r1 = None if order <= 1 else (lambda_inner[1] - lambda_inner[0]) / h + r2 = None if order <= 2 else (lambda_inner[2] - lambda_inner[0]) / h + x = self.singlestep_dpm_solver_update(x, vec_s, vec_t, order, solver_type=solver_type, r1=r1, r2=r2) + if denoise_to_zero: + x = self.denoise_to_zero_fn(x, torch.ones((x.shape[0],)).to(device) * t_0) + return x + + +############################################################# +# other utility functions +############################################################# + +def interpolate_fn(x, xp, yp): + """ + A piecewise linear function y = f(x), using xp and yp as keypoints. + We implement f(x) in a differentiable way (i.e. applicable for autograd). + The function f(x) is well-defined for all x-axis. (For x beyond the bounds of xp, we use the outmost points of xp to define the linear function.) + Args: + x: PyTorch tensor with shape [N, C], where N is the batch size, C is the number of channels (we use C = 1 for DPM-Solver). + xp: PyTorch tensor with shape [C, K], where K is the number of keypoints. + yp: PyTorch tensor with shape [C, K]. + Returns: + The function values f(x), with shape [N, C]. + """ + N, K = x.shape[0], xp.shape[1] + all_x = torch.cat([x.unsqueeze(2), xp.unsqueeze(0).repeat((N, 1, 1))], dim=2) + sorted_all_x, x_indices = torch.sort(all_x, dim=2) + x_idx = torch.argmin(x_indices, dim=2) + cand_start_idx = x_idx - 1 + start_idx = torch.where( + torch.eq(x_idx, 0), + torch.tensor(1, device=x.device), + torch.where( + torch.eq(x_idx, K), torch.tensor(K - 2, device=x.device), cand_start_idx, + ), + ) + end_idx = torch.where(torch.eq(start_idx, cand_start_idx), start_idx + 2, start_idx + 1) + start_x = torch.gather(sorted_all_x, dim=2, index=start_idx.unsqueeze(2)).squeeze(2) + end_x = torch.gather(sorted_all_x, dim=2, index=end_idx.unsqueeze(2)).squeeze(2) + start_idx2 = torch.where( + torch.eq(x_idx, 0), + torch.tensor(0, device=x.device), + torch.where( + torch.eq(x_idx, K), torch.tensor(K - 2, device=x.device), cand_start_idx, + ), + ) + y_positions_expanded = yp.unsqueeze(0).expand(N, -1, -1) + start_y = torch.gather(y_positions_expanded, dim=2, index=start_idx2.unsqueeze(2)).squeeze(2) + end_y = torch.gather(y_positions_expanded, dim=2, index=(start_idx2 + 1).unsqueeze(2)).squeeze(2) + cand = start_y + (x - start_x) * (end_y - start_y) / (end_x - start_x) + return cand + + +def expand_dims(v, dims): + """ + Expand the tensor `v` to the dim `dims`. + Args: + `v`: a PyTorch tensor with shape [N]. + `dim`: a `int`. + Returns: + a PyTorch tensor with shape [N, 1, 1, ..., 1] and the total dimension is `dims`. + """ + return v[(...,) + (None,) * (dims - 1)] \ No newline at end of file diff --git a/ldm/models/diffusion/dpm_solver/sampler.py b/ldm/models/diffusion/dpm_solver/sampler.py new file mode 100644 index 0000000000000000000000000000000000000000..e4d0d0a387548a280d872b60344d0a74dac5e1f0 --- /dev/null +++ b/ldm/models/diffusion/dpm_solver/sampler.py @@ -0,0 +1,96 @@ +"""SAMPLING ONLY.""" +import torch + +from .dpm_solver import NoiseScheduleVP, model_wrapper, DPM_Solver + +MODEL_TYPES = { + "eps": "noise", + "v": "v" +} + + +class DPMSolverSampler(object): + def __init__(self, model, device=torch.device("cuda"), **kwargs): + super().__init__() + self.model = model + self.device = device + to_torch = lambda x: x.clone().detach().to(torch.float32).to(model.device) + self.register_buffer('alphas_cumprod', to_torch(model.alphas_cumprod)) + + def register_buffer(self, name, attr): + if type(attr) == torch.Tensor: + if attr.device != self.device: + attr = attr.to(self.device) + setattr(self, name, attr) + + @torch.no_grad() + def sample(self, + S, + batch_size, + shape, + conditioning=None, + callback=None, + normals_sequence=None, + img_callback=None, + quantize_x0=False, + eta=0., + mask=None, + x0=None, + temperature=1., + noise_dropout=0., + score_corrector=None, + corrector_kwargs=None, + verbose=True, + x_T=None, + log_every_t=100, + unconditional_guidance_scale=1., + unconditional_conditioning=None, + # this has to come in the same format as the conditioning, # e.g. as encoded tokens, ... + **kwargs + ): + if conditioning is not None: + if isinstance(conditioning, dict): + ctmp = conditioning[list(conditioning.keys())[0]] + while isinstance(ctmp, list): ctmp = ctmp[0] + if isinstance(ctmp, torch.Tensor): + cbs = ctmp.shape[0] + if cbs != batch_size: + print(f"Warning: Got {cbs} conditionings but batch-size is {batch_size}") + elif isinstance(conditioning, list): + for ctmp in conditioning: + if ctmp.shape[0] != batch_size: + print(f"Warning: Got {ctmp.shape[0]} conditionings but batch-size is {batch_size}") + else: + if isinstance(conditioning, torch.Tensor): + if conditioning.shape[0] != batch_size: + print(f"Warning: Got {conditioning.shape[0]} conditionings but batch-size is {batch_size}") + + # sampling + C, H, W = shape + size = (batch_size, C, H, W) + + print(f'Data shape for DPM-Solver sampling is {size}, sampling steps {S}') + + device = self.model.betas.device + if x_T is None: + img = torch.randn(size, device=device) + else: + img = x_T + + ns = NoiseScheduleVP('discrete', alphas_cumprod=self.alphas_cumprod) + + model_fn = model_wrapper( + lambda x, t, c: self.model.apply_model(x, t, c), + ns, + model_type=MODEL_TYPES[self.model.parameterization], + guidance_type="classifier-free", + condition=conditioning, + unconditional_condition=unconditional_conditioning, + guidance_scale=unconditional_guidance_scale, + ) + + dpm_solver = DPM_Solver(model_fn, ns, predict_x0=True, thresholding=False) + x = dpm_solver.sample(img, steps=S, skip_type="time_uniform", method="multistep", order=2, + lower_order_final=True) + + return x.to(device), None diff --git a/ldm/models/diffusion/plms.py b/ldm/models/diffusion/plms.py new file mode 100644 index 0000000000000000000000000000000000000000..9d31b3994ed283e9d97ed0ae275d89046442cc89 --- /dev/null +++ b/ldm/models/diffusion/plms.py @@ -0,0 +1,245 @@ +"""SAMPLING ONLY.""" + +import torch +import numpy as np +from tqdm import tqdm +from functools import partial + +from ldm.modules.diffusionmodules.util import make_ddim_sampling_parameters, make_ddim_timesteps, noise_like +from ldm.models.diffusion.sampling_util import norm_thresholding + + +class PLMSSampler(object): + def __init__(self, model, schedule="linear", device=torch.device("cuda"), **kwargs): + super().__init__() + self.model = model + self.ddpm_num_timesteps = model.num_timesteps + self.schedule = schedule + self.device = device + + def register_buffer(self, name, attr): + if type(attr) == torch.Tensor: + if attr.device != self.device: + attr = attr.to(self.device) + setattr(self, name, attr) + + def make_schedule(self, ddim_num_steps, ddim_discretize="uniform", ddim_eta=0., verbose=True): + if ddim_eta != 0: + raise ValueError('ddim_eta must be 0 for PLMS') + self.ddim_timesteps = make_ddim_timesteps(ddim_discr_method=ddim_discretize, num_ddim_timesteps=ddim_num_steps, + num_ddpm_timesteps=self.ddpm_num_timesteps,verbose=verbose) + alphas_cumprod = self.model.alphas_cumprod + assert alphas_cumprod.shape[0] == self.ddpm_num_timesteps, 'alphas have to be defined for each timestep' + to_torch = lambda x: x.clone().detach().to(torch.float32).to(self.model.device) + + self.register_buffer('betas', to_torch(self.model.betas)) + self.register_buffer('alphas_cumprod', to_torch(alphas_cumprod)) + self.register_buffer('alphas_cumprod_prev', to_torch(self.model.alphas_cumprod_prev)) + + # calculations for diffusion q(x_t | x_{t-1}) and others + self.register_buffer('sqrt_alphas_cumprod', to_torch(np.sqrt(alphas_cumprod.cpu()))) + self.register_buffer('sqrt_one_minus_alphas_cumprod', to_torch(np.sqrt(1. - alphas_cumprod.cpu()))) + self.register_buffer('log_one_minus_alphas_cumprod', to_torch(np.log(1. - alphas_cumprod.cpu()))) + self.register_buffer('sqrt_recip_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod.cpu()))) + self.register_buffer('sqrt_recipm1_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod.cpu() - 1))) + + # ddim sampling parameters + ddim_sigmas, ddim_alphas, ddim_alphas_prev = make_ddim_sampling_parameters(alphacums=alphas_cumprod.cpu(), + ddim_timesteps=self.ddim_timesteps, + eta=ddim_eta,verbose=verbose) + self.register_buffer('ddim_sigmas', ddim_sigmas) + self.register_buffer('ddim_alphas', ddim_alphas) + self.register_buffer('ddim_alphas_prev', ddim_alphas_prev) + self.register_buffer('ddim_sqrt_one_minus_alphas', np.sqrt(1. - ddim_alphas)) + sigmas_for_original_sampling_steps = ddim_eta * torch.sqrt( + (1 - self.alphas_cumprod_prev) / (1 - self.alphas_cumprod) * ( + 1 - self.alphas_cumprod / self.alphas_cumprod_prev)) + self.register_buffer('ddim_sigmas_for_original_num_steps', sigmas_for_original_sampling_steps) + + @torch.no_grad() + def sample(self, + S, + batch_size, + shape, + conditioning=None, + callback=None, + normals_sequence=None, + img_callback=None, + quantize_x0=False, + eta=0., + mask=None, + x0=None, + temperature=1., + noise_dropout=0., + score_corrector=None, + corrector_kwargs=None, + verbose=True, + x_T=None, + log_every_t=100, + unconditional_guidance_scale=1., + unconditional_conditioning=None, + # this has to come in the same format as the conditioning, # e.g. as encoded tokens, ... + dynamic_threshold=None, + **kwargs + ): + if conditioning is not None: + if isinstance(conditioning, dict): + cbs = conditioning[list(conditioning.keys())[0]].shape[0] + if cbs != batch_size: + print(f"Warning: Got {cbs} conditionings but batch-size is {batch_size}") + else: + if conditioning.shape[0] != batch_size: + print(f"Warning: Got {conditioning.shape[0]} conditionings but batch-size is {batch_size}") + + self.make_schedule(ddim_num_steps=S, ddim_eta=eta, verbose=verbose) + # sampling + C, H, W = shape + size = (batch_size, C, H, W) + print(f'Data shape for PLMS sampling is {size}') + + samples, intermediates = self.plms_sampling(conditioning, size, + callback=callback, + img_callback=img_callback, + quantize_denoised=quantize_x0, + mask=mask, x0=x0, + ddim_use_original_steps=False, + noise_dropout=noise_dropout, + temperature=temperature, + score_corrector=score_corrector, + corrector_kwargs=corrector_kwargs, + x_T=x_T, + log_every_t=log_every_t, + unconditional_guidance_scale=unconditional_guidance_scale, + unconditional_conditioning=unconditional_conditioning, + dynamic_threshold=dynamic_threshold, + ) + return samples, intermediates + + @torch.no_grad() + def plms_sampling(self, cond, shape, + x_T=None, ddim_use_original_steps=False, + callback=None, timesteps=None, quantize_denoised=False, + mask=None, x0=None, img_callback=None, log_every_t=100, + temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None, + unconditional_guidance_scale=1., unconditional_conditioning=None, + dynamic_threshold=None): + device = self.model.betas.device + b = shape[0] + if x_T is None: + img = torch.randn(shape, device=device) + else: + img = x_T + + if timesteps is None: + timesteps = self.ddpm_num_timesteps if ddim_use_original_steps else self.ddim_timesteps + elif timesteps is not None and not ddim_use_original_steps: + subset_end = int(min(timesteps / self.ddim_timesteps.shape[0], 1) * self.ddim_timesteps.shape[0]) - 1 + timesteps = self.ddim_timesteps[:subset_end] + + intermediates = {'x_inter': [img], 'pred_x0': [img]} + time_range = list(reversed(range(0,timesteps))) if ddim_use_original_steps else np.flip(timesteps) + total_steps = timesteps if ddim_use_original_steps else timesteps.shape[0] + print(f"Running PLMS Sampling with {total_steps} timesteps") + + iterator = tqdm(time_range, desc='PLMS Sampler', total=total_steps) + old_eps = [] + + for i, step in enumerate(iterator): + index = total_steps - i - 1 + ts = torch.full((b,), step, device=device, dtype=torch.long) + ts_next = torch.full((b,), time_range[min(i + 1, len(time_range) - 1)], device=device, dtype=torch.long) + + if mask is not None: + assert x0 is not None + img_orig = self.model.q_sample(x0, ts) # TODO: deterministic forward pass? + img = img_orig * mask + (1. - mask) * img + + outs = self.p_sample_plms(img, cond, ts, index=index, use_original_steps=ddim_use_original_steps, + quantize_denoised=quantize_denoised, temperature=temperature, + noise_dropout=noise_dropout, score_corrector=score_corrector, + corrector_kwargs=corrector_kwargs, + unconditional_guidance_scale=unconditional_guidance_scale, + unconditional_conditioning=unconditional_conditioning, + old_eps=old_eps, t_next=ts_next, + dynamic_threshold=dynamic_threshold) + img, pred_x0, e_t = outs + old_eps.append(e_t) + if len(old_eps) >= 4: + old_eps.pop(0) + if callback: callback(i) + if img_callback: img_callback(pred_x0, i) + + if index % log_every_t == 0 or index == total_steps - 1: + intermediates['x_inter'].append(img) + intermediates['pred_x0'].append(pred_x0) + + return img, intermediates + + @torch.no_grad() + def p_sample_plms(self, x, c, t, index, repeat_noise=False, use_original_steps=False, quantize_denoised=False, + temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None, + unconditional_guidance_scale=1., unconditional_conditioning=None, old_eps=None, t_next=None, + dynamic_threshold=None): + b, *_, device = *x.shape, x.device + + def get_model_output(x, t): + if unconditional_conditioning is None or unconditional_guidance_scale == 1.: + e_t = self.model.apply_model(x, t, c) + else: + x_in = torch.cat([x] * 2) + t_in = torch.cat([t] * 2) + c_in = torch.cat([unconditional_conditioning, c]) + e_t_uncond, e_t = self.model.apply_model(x_in, t_in, c_in).chunk(2) + e_t = e_t_uncond + unconditional_guidance_scale * (e_t - e_t_uncond) + + if score_corrector is not None: + assert self.model.parameterization == "eps" + e_t = score_corrector.modify_score(self.model, e_t, x, t, c, **corrector_kwargs) + + return e_t + + alphas = self.model.alphas_cumprod if use_original_steps else self.ddim_alphas + alphas_prev = self.model.alphas_cumprod_prev if use_original_steps else self.ddim_alphas_prev + sqrt_one_minus_alphas = self.model.sqrt_one_minus_alphas_cumprod if use_original_steps else self.ddim_sqrt_one_minus_alphas + sigmas = self.model.ddim_sigmas_for_original_num_steps if use_original_steps else self.ddim_sigmas + + def get_x_prev_and_pred_x0(e_t, index): + # select parameters corresponding to the currently considered timestep + a_t = torch.full((b, 1, 1, 1), alphas[index], device=device) + a_prev = torch.full((b, 1, 1, 1), alphas_prev[index], device=device) + sigma_t = torch.full((b, 1, 1, 1), sigmas[index], device=device) + sqrt_one_minus_at = torch.full((b, 1, 1, 1), sqrt_one_minus_alphas[index],device=device) + + # current prediction for x_0 + pred_x0 = (x - sqrt_one_minus_at * e_t) / a_t.sqrt() + if quantize_denoised: + pred_x0, _, *_ = self.model.first_stage_model.quantize(pred_x0) + if dynamic_threshold is not None: + pred_x0 = norm_thresholding(pred_x0, dynamic_threshold) + # direction pointing to x_t + dir_xt = (1. - a_prev - sigma_t**2).sqrt() * e_t + noise = sigma_t * noise_like(x.shape, device, repeat_noise) * temperature + if noise_dropout > 0.: + noise = torch.nn.functional.dropout(noise, p=noise_dropout) + x_prev = a_prev.sqrt() * pred_x0 + dir_xt + noise + return x_prev, pred_x0 + + e_t = get_model_output(x, t) + if len(old_eps) == 0: + # Pseudo Improved Euler (2nd order) + x_prev, pred_x0 = get_x_prev_and_pred_x0(e_t, index) + e_t_next = get_model_output(x_prev, t_next) + e_t_prime = (e_t + e_t_next) / 2 + elif len(old_eps) == 1: + # 2nd order Pseudo Linear Multistep (Adams-Bashforth) + e_t_prime = (3 * e_t - old_eps[-1]) / 2 + elif len(old_eps) == 2: + # 3nd order Pseudo Linear Multistep (Adams-Bashforth) + e_t_prime = (23 * e_t - 16 * old_eps[-1] + 5 * old_eps[-2]) / 12 + elif len(old_eps) >= 3: + # 4nd order Pseudo Linear Multistep (Adams-Bashforth) + e_t_prime = (55 * e_t - 59 * old_eps[-1] + 37 * old_eps[-2] - 9 * old_eps[-3]) / 24 + + x_prev, pred_x0 = get_x_prev_and_pred_x0(e_t_prime, index) + + return x_prev, pred_x0, e_t diff --git a/ldm/models/diffusion/sampling_util.py b/ldm/models/diffusion/sampling_util.py new file mode 100644 index 0000000000000000000000000000000000000000..7eff02be6d7c54d43ee6680636ac0698dd3b3f33 --- /dev/null +++ b/ldm/models/diffusion/sampling_util.py @@ -0,0 +1,22 @@ +import torch +import numpy as np + + +def append_dims(x, target_dims): + """Appends dimensions to the end of a tensor until it has target_dims dimensions. + From https://github.com/crowsonkb/k-diffusion/blob/master/k_diffusion/utils.py""" + dims_to_append = target_dims - x.ndim + if dims_to_append < 0: + raise ValueError(f'input has {x.ndim} dims but target_dims is {target_dims}, which is less') + return x[(...,) + (None,) * dims_to_append] + + +def norm_thresholding(x0, value): + s = append_dims(x0.pow(2).flatten(1).mean(1).sqrt().clamp(min=value), x0.ndim) + return x0 * (value / s) + + +def spatial_norm_thresholding(x0, value): + # b c h w + s = x0.pow(2).mean(1, keepdim=True).sqrt().clamp(min=value) + return x0 * (value / s) \ No newline at end of file diff --git a/ldm/modules/attention.py b/ldm/modules/attention.py new file mode 100644 index 0000000000000000000000000000000000000000..509cd873768f0dd75a75ab3fcdd652822b12b59f --- /dev/null +++ b/ldm/modules/attention.py @@ -0,0 +1,341 @@ +from inspect import isfunction +import math +import torch +import torch.nn.functional as F +from torch import nn, einsum +from einops import rearrange, repeat +from typing import Optional, Any + +from ldm.modules.diffusionmodules.util import checkpoint + + +try: + import xformers + import xformers.ops + XFORMERS_IS_AVAILBLE = True +except: + XFORMERS_IS_AVAILBLE = False + +# CrossAttn precision handling +import os +_ATTN_PRECISION = os.environ.get("ATTN_PRECISION", "fp32") + +def exists(val): + return val is not None + + +def uniq(arr): + return{el: True for el in arr}.keys() + + +def default(val, d): + if exists(val): + return val + return d() if isfunction(d) else d + + +def max_neg_value(t): + return -torch.finfo(t.dtype).max + + +def init_(tensor): + dim = tensor.shape[-1] + std = 1 / math.sqrt(dim) + tensor.uniform_(-std, std) + return tensor + + +# feedforward +class GEGLU(nn.Module): + def __init__(self, dim_in, dim_out): + super().__init__() + self.proj = nn.Linear(dim_in, dim_out * 2) + + def forward(self, x): + x, gate = self.proj(x).chunk(2, dim=-1) + return x * F.gelu(gate) + + +class FeedForward(nn.Module): + def __init__(self, dim, dim_out=None, mult=4, glu=False, dropout=0.): + super().__init__() + inner_dim = int(dim * mult) + dim_out = default(dim_out, dim) + project_in = nn.Sequential( + nn.Linear(dim, inner_dim), + nn.GELU() + ) if not glu else GEGLU(dim, inner_dim) + + self.net = nn.Sequential( + project_in, + nn.Dropout(dropout), + nn.Linear(inner_dim, dim_out) + ) + + def forward(self, x): + return self.net(x) + + +def zero_module(module): + """ + Zero out the parameters of a module and return it. + """ + for p in module.parameters(): + p.detach().zero_() + return module + + +def Normalize(in_channels): + return torch.nn.GroupNorm(num_groups=32, num_channels=in_channels, eps=1e-6, affine=True) + + +class SpatialSelfAttention(nn.Module): + def __init__(self, in_channels): + super().__init__() + self.in_channels = in_channels + + self.norm = Normalize(in_channels) + self.q = torch.nn.Conv2d(in_channels, + in_channels, + kernel_size=1, + stride=1, + padding=0) + self.k = torch.nn.Conv2d(in_channels, + in_channels, + kernel_size=1, + stride=1, + padding=0) + self.v = torch.nn.Conv2d(in_channels, + in_channels, + kernel_size=1, + stride=1, + padding=0) + self.proj_out = torch.nn.Conv2d(in_channels, + in_channels, + kernel_size=1, + stride=1, + padding=0) + + def forward(self, x): + h_ = x + h_ = self.norm(h_) + q = self.q(h_) + k = self.k(h_) + v = self.v(h_) + + # compute attention + b,c,h,w = q.shape + q = rearrange(q, 'b c h w -> b (h w) c') + k = rearrange(k, 'b c h w -> b c (h w)') + w_ = torch.einsum('bij,bjk->bik', q, k) + + w_ = w_ * (int(c)**(-0.5)) + w_ = torch.nn.functional.softmax(w_, dim=2) + + # attend to values + v = rearrange(v, 'b c h w -> b c (h w)') + w_ = rearrange(w_, 'b i j -> b j i') + h_ = torch.einsum('bij,bjk->bik', v, w_) + h_ = rearrange(h_, 'b c (h w) -> b c h w', h=h) + h_ = self.proj_out(h_) + + return x+h_ + + +class CrossAttention(nn.Module): + def __init__(self, query_dim, context_dim=None, heads=8, dim_head=64, dropout=0.): + super().__init__() + inner_dim = dim_head * heads + context_dim = default(context_dim, query_dim) + + self.scale = dim_head ** -0.5 + self.heads = heads + + self.to_q = nn.Linear(query_dim, inner_dim, bias=False) + self.to_k = nn.Linear(context_dim, inner_dim, bias=False) + self.to_v = nn.Linear(context_dim, inner_dim, bias=False) + + self.to_out = nn.Sequential( + nn.Linear(inner_dim, query_dim), + nn.Dropout(dropout) + ) + + def forward(self, x, context=None, mask=None): + h = self.heads + + q = self.to_q(x) + context = default(context, x) + k = self.to_k(context) + v = self.to_v(context) + + q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> (b h) n d', h=h), (q, k, v)) + + # force cast to fp32 to avoid overflowing + if _ATTN_PRECISION =="fp32": + with torch.autocast(enabled=False, device_type = 'cuda'): + q, k = q.float(), k.float() + sim = einsum('b i d, b j d -> b i j', q, k) * self.scale + else: + sim = einsum('b i d, b j d -> b i j', q, k) * self.scale + + del q, k + + if exists(mask): + mask = rearrange(mask, 'b ... -> b (...)') + max_neg_value = -torch.finfo(sim.dtype).max + mask = repeat(mask, 'b j -> (b h) () j', h=h) + sim.masked_fill_(~mask, max_neg_value) + + # attention, what we cannot get enough of + sim = sim.softmax(dim=-1) + + out = einsum('b i j, b j d -> b i d', sim, v) + out = rearrange(out, '(b h) n d -> b n (h d)', h=h) + return self.to_out(out) + + +class MemoryEfficientCrossAttention(nn.Module): + # https://github.com/MatthieuTPHR/diffusers/blob/d80b531ff8060ec1ea982b65a1b8df70f73aa67c/src/diffusers/models/attention.py#L223 + def __init__(self, query_dim, context_dim=None, heads=8, dim_head=64, dropout=0.0): + super().__init__() + print(f"Setting up {self.__class__.__name__}. Query dim is {query_dim}, context_dim is {context_dim} and using " + f"{heads} heads.") + inner_dim = dim_head * heads + context_dim = default(context_dim, query_dim) + + self.heads = heads + self.dim_head = dim_head + + self.to_q = nn.Linear(query_dim, inner_dim, bias=False) + self.to_k = nn.Linear(context_dim, inner_dim, bias=False) + self.to_v = nn.Linear(context_dim, inner_dim, bias=False) + + self.to_out = nn.Sequential(nn.Linear(inner_dim, query_dim), nn.Dropout(dropout)) + self.attention_op: Optional[Any] = None + + def forward(self, x, context=None, mask=None): + q = self.to_q(x) + context = default(context, x) + k = self.to_k(context) + v = self.to_v(context) + + b, _, _ = q.shape + q, k, v = map( + lambda t: t.unsqueeze(3) + .reshape(b, t.shape[1], self.heads, self.dim_head) + .permute(0, 2, 1, 3) + .reshape(b * self.heads, t.shape[1], self.dim_head) + .contiguous(), + (q, k, v), + ) + + # actually compute the attention, what we cannot get enough of + out = xformers.ops.memory_efficient_attention(q, k, v, attn_bias=None, op=self.attention_op) + + if exists(mask): + raise NotImplementedError + out = ( + out.unsqueeze(0) + .reshape(b, self.heads, out.shape[1], self.dim_head) + .permute(0, 2, 1, 3) + .reshape(b, out.shape[1], self.heads * self.dim_head) + ) + return self.to_out(out) + + +class BasicTransformerBlock(nn.Module): + ATTENTION_MODES = { + "softmax": CrossAttention, # vanilla attention + "softmax-xformers": MemoryEfficientCrossAttention + } + def __init__(self, dim, n_heads, d_head, dropout=0., context_dim=None, gated_ff=True, checkpoint=True, + disable_self_attn=False): + super().__init__() + attn_mode = "softmax-xformers" if XFORMERS_IS_AVAILBLE else "softmax" + assert attn_mode in self.ATTENTION_MODES + attn_cls = self.ATTENTION_MODES[attn_mode] + self.disable_self_attn = disable_self_attn + self.attn1 = attn_cls(query_dim=dim, heads=n_heads, dim_head=d_head, dropout=dropout, + context_dim=context_dim if self.disable_self_attn else None) # is a self-attention if not self.disable_self_attn + self.ff = FeedForward(dim, dropout=dropout, glu=gated_ff) + self.attn2 = attn_cls(query_dim=dim, context_dim=context_dim, + heads=n_heads, dim_head=d_head, dropout=dropout) # is self-attn if context is none + self.norm1 = nn.LayerNorm(dim) + self.norm2 = nn.LayerNorm(dim) + self.norm3 = nn.LayerNorm(dim) + self.checkpoint = checkpoint + + def forward(self, x, context=None): + return checkpoint(self._forward, (x, context), self.parameters(), self.checkpoint) + + def _forward(self, x, context=None): + x = self.attn1(self.norm1(x), context=context if self.disable_self_attn else None) + x + x = self.attn2(self.norm2(x), context=context) + x + x = self.ff(self.norm3(x)) + x + return x + + +class SpatialTransformer(nn.Module): + """ + Transformer block for image-like data. + First, project the input (aka embedding) + and reshape to b, t, d. + Then apply standard transformer action. + Finally, reshape to image + NEW: use_linear for more efficiency instead of the 1x1 convs + """ + def __init__(self, in_channels, n_heads, d_head, + depth=1, dropout=0., context_dim=None, + disable_self_attn=False, use_linear=False, + use_checkpoint=True): + super().__init__() + if exists(context_dim) and not isinstance(context_dim, list): + context_dim = [context_dim] + self.in_channels = in_channels + inner_dim = n_heads * d_head + self.norm = Normalize(in_channels) + if not use_linear: + self.proj_in = nn.Conv2d(in_channels, + inner_dim, + kernel_size=1, + stride=1, + padding=0) + else: + self.proj_in = nn.Linear(in_channels, inner_dim) + + self.transformer_blocks = nn.ModuleList( + [BasicTransformerBlock(inner_dim, n_heads, d_head, dropout=dropout, context_dim=context_dim[d], + disable_self_attn=disable_self_attn, checkpoint=use_checkpoint) + for d in range(depth)] + ) + if not use_linear: + self.proj_out = zero_module(nn.Conv2d(inner_dim, + in_channels, + kernel_size=1, + stride=1, + padding=0)) + else: + self.proj_out = zero_module(nn.Linear(in_channels, inner_dim)) + self.use_linear = use_linear + + def forward(self, x, context=None): + # note: if no context is given, cross-attention defaults to self-attention + if not isinstance(context, list): + context = [context] + b, c, h, w = x.shape + x_in = x + x = self.norm(x) + if not self.use_linear: + x = self.proj_in(x) + x = rearrange(x, 'b c h w -> b (h w) c').contiguous() + if self.use_linear: + x = self.proj_in(x) + for i, block in enumerate(self.transformer_blocks): + x = block(x, context=context[i]) + if self.use_linear: + x = self.proj_out(x) + x = rearrange(x, 'b (h w) c -> b c h w', h=h, w=w).contiguous() + if not self.use_linear: + x = self.proj_out(x) + return x + x_in + diff --git a/ldm/modules/diffusionmodules/__init__.py b/ldm/modules/diffusionmodules/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/ldm/modules/diffusionmodules/model.py b/ldm/modules/diffusionmodules/model.py new file mode 100644 index 0000000000000000000000000000000000000000..b089eebbe1676d8249005bb9def002ff5180715b --- /dev/null +++ b/ldm/modules/diffusionmodules/model.py @@ -0,0 +1,852 @@ +# pytorch_diffusion + derived encoder decoder +import math +import torch +import torch.nn as nn +import numpy as np +from einops import rearrange +from typing import Optional, Any + +from ldm.modules.attention import MemoryEfficientCrossAttention + +try: + import xformers + import xformers.ops + XFORMERS_IS_AVAILBLE = True +except: + XFORMERS_IS_AVAILBLE = False + print("No module 'xformers'. Proceeding without it.") + + +def get_timestep_embedding(timesteps, embedding_dim): + """ + This matches the implementation in Denoising Diffusion Probabilistic Models: + From Fairseq. + Build sinusoidal embeddings. + This matches the implementation in tensor2tensor, but differs slightly + from the description in Section 3.5 of "Attention Is All You Need". + """ + assert len(timesteps.shape) == 1 + + half_dim = embedding_dim // 2 + emb = math.log(10000) / (half_dim - 1) + emb = torch.exp(torch.arange(half_dim, dtype=torch.float32) * -emb) + emb = emb.to(device=timesteps.device) + emb = timesteps.float()[:, None] * emb[None, :] + emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1) + if embedding_dim % 2 == 1: # zero pad + emb = torch.nn.functional.pad(emb, (0,1,0,0)) + return emb + + +def nonlinearity(x): + # swish + return x*torch.sigmoid(x) + + +def Normalize(in_channels, num_groups=32): + return torch.nn.GroupNorm(num_groups=num_groups, num_channels=in_channels, eps=1e-6, affine=True) + + +class Upsample(nn.Module): + def __init__(self, in_channels, with_conv): + super().__init__() + self.with_conv = with_conv + if self.with_conv: + self.conv = torch.nn.Conv2d(in_channels, + in_channels, + kernel_size=3, + stride=1, + padding=1) + + def forward(self, x): + x = torch.nn.functional.interpolate(x, scale_factor=2.0, mode="nearest") + if self.with_conv: + x = self.conv(x) + return x + + +class Downsample(nn.Module): + def __init__(self, in_channels, with_conv): + super().__init__() + self.with_conv = with_conv + if self.with_conv: + # no asymmetric padding in torch conv, must do it ourselves + self.conv = torch.nn.Conv2d(in_channels, + in_channels, + kernel_size=3, + stride=2, + padding=0) + + def forward(self, x): + if self.with_conv: + pad = (0,1,0,1) + x = torch.nn.functional.pad(x, pad, mode="constant", value=0) + x = self.conv(x) + else: + x = torch.nn.functional.avg_pool2d(x, kernel_size=2, stride=2) + return x + + +class ResnetBlock(nn.Module): + def __init__(self, *, in_channels, out_channels=None, conv_shortcut=False, + dropout, temb_channels=512): + super().__init__() + self.in_channels = in_channels + out_channels = in_channels if out_channels is None else out_channels + self.out_channels = out_channels + self.use_conv_shortcut = conv_shortcut + + self.norm1 = Normalize(in_channels) + self.conv1 = torch.nn.Conv2d(in_channels, + out_channels, + kernel_size=3, + stride=1, + padding=1) + if temb_channels > 0: + self.temb_proj = torch.nn.Linear(temb_channels, + out_channels) + self.norm2 = Normalize(out_channels) + self.dropout = torch.nn.Dropout(dropout) + self.conv2 = torch.nn.Conv2d(out_channels, + out_channels, + kernel_size=3, + stride=1, + padding=1) + if self.in_channels != self.out_channels: + if self.use_conv_shortcut: + self.conv_shortcut = torch.nn.Conv2d(in_channels, + out_channels, + kernel_size=3, + stride=1, + padding=1) + else: + self.nin_shortcut = torch.nn.Conv2d(in_channels, + out_channels, + kernel_size=1, + stride=1, + padding=0) + + def forward(self, x, temb): + h = x + h = self.norm1(h) + h = nonlinearity(h) + h = self.conv1(h) + + if temb is not None: + h = h + self.temb_proj(nonlinearity(temb))[:,:,None,None] + + h = self.norm2(h) + h = nonlinearity(h) + h = self.dropout(h) + h = self.conv2(h) + + if self.in_channels != self.out_channels: + if self.use_conv_shortcut: + x = self.conv_shortcut(x) + else: + x = self.nin_shortcut(x) + + return x+h + + +class AttnBlock(nn.Module): + def __init__(self, in_channels): + super().__init__() + self.in_channels = in_channels + + self.norm = Normalize(in_channels) + self.q = torch.nn.Conv2d(in_channels, + in_channels, + kernel_size=1, + stride=1, + padding=0) + self.k = torch.nn.Conv2d(in_channels, + in_channels, + kernel_size=1, + stride=1, + padding=0) + self.v = torch.nn.Conv2d(in_channels, + in_channels, + kernel_size=1, + stride=1, + padding=0) + self.proj_out = torch.nn.Conv2d(in_channels, + in_channels, + kernel_size=1, + stride=1, + padding=0) + + def forward(self, x): + h_ = x + h_ = self.norm(h_) + q = self.q(h_) + k = self.k(h_) + v = self.v(h_) + + # compute attention + b,c,h,w = q.shape + q = q.reshape(b,c,h*w) + q = q.permute(0,2,1) # b,hw,c + k = k.reshape(b,c,h*w) # b,c,hw + w_ = torch.bmm(q,k) # b,hw,hw w[b,i,j]=sum_c q[b,i,c]k[b,c,j] + w_ = w_ * (int(c)**(-0.5)) + w_ = torch.nn.functional.softmax(w_, dim=2) + + # attend to values + v = v.reshape(b,c,h*w) + w_ = w_.permute(0,2,1) # b,hw,hw (first hw of k, second of q) + h_ = torch.bmm(v,w_) # b, c,hw (hw of q) h_[b,c,j] = sum_i v[b,c,i] w_[b,i,j] + h_ = h_.reshape(b,c,h,w) + + h_ = self.proj_out(h_) + + return x+h_ + +class MemoryEfficientAttnBlock(nn.Module): + """ + Uses xformers efficient implementation, + see https://github.com/MatthieuTPHR/diffusers/blob/d80b531ff8060ec1ea982b65a1b8df70f73aa67c/src/diffusers/models/attention.py#L223 + Note: this is a single-head self-attention operation + """ + # + def __init__(self, in_channels): + super().__init__() + self.in_channels = in_channels + + self.norm = Normalize(in_channels) + self.q = torch.nn.Conv2d(in_channels, + in_channels, + kernel_size=1, + stride=1, + padding=0) + self.k = torch.nn.Conv2d(in_channels, + in_channels, + kernel_size=1, + stride=1, + padding=0) + self.v = torch.nn.Conv2d(in_channels, + in_channels, + kernel_size=1, + stride=1, + padding=0) + self.proj_out = torch.nn.Conv2d(in_channels, + in_channels, + kernel_size=1, + stride=1, + padding=0) + self.attention_op: Optional[Any] = None + + def forward(self, x): + h_ = x + h_ = self.norm(h_) + q = self.q(h_) + k = self.k(h_) + v = self.v(h_) + + # compute attention + B, C, H, W = q.shape + q, k, v = map(lambda x: rearrange(x, 'b c h w -> b (h w) c'), (q, k, v)) + + q, k, v = map( + lambda t: t.unsqueeze(3) + .reshape(B, t.shape[1], 1, C) + .permute(0, 2, 1, 3) + .reshape(B * 1, t.shape[1], C) + .contiguous(), + (q, k, v), + ) + out = xformers.ops.memory_efficient_attention(q, k, v, attn_bias=None, op=self.attention_op) + + out = ( + out.unsqueeze(0) + .reshape(B, 1, out.shape[1], C) + .permute(0, 2, 1, 3) + .reshape(B, out.shape[1], C) + ) + out = rearrange(out, 'b (h w) c -> b c h w', b=B, h=H, w=W, c=C) + out = self.proj_out(out) + return x+out + + +class MemoryEfficientCrossAttentionWrapper(MemoryEfficientCrossAttention): + def forward(self, x, context=None, mask=None): + b, c, h, w = x.shape + x = rearrange(x, 'b c h w -> b (h w) c') + out = super().forward(x, context=context, mask=mask) + out = rearrange(out, 'b (h w) c -> b c h w', h=h, w=w, c=c) + return x + out + + +def make_attn(in_channels, attn_type="vanilla", attn_kwargs=None): + assert attn_type in ["vanilla", "vanilla-xformers", "memory-efficient-cross-attn", "linear", "none"], f'attn_type {attn_type} unknown' + if XFORMERS_IS_AVAILBLE and attn_type == "vanilla": + attn_type = "vanilla-xformers" + print(f"making attention of type '{attn_type}' with {in_channels} in_channels") + if attn_type == "vanilla": + assert attn_kwargs is None + return AttnBlock(in_channels) + elif attn_type == "vanilla-xformers": + print(f"building MemoryEfficientAttnBlock with {in_channels} in_channels...") + return MemoryEfficientAttnBlock(in_channels) + elif type == "memory-efficient-cross-attn": + attn_kwargs["query_dim"] = in_channels + return MemoryEfficientCrossAttentionWrapper(**attn_kwargs) + elif attn_type == "none": + return nn.Identity(in_channels) + else: + raise NotImplementedError() + + +class Model(nn.Module): + def __init__(self, *, ch, out_ch, ch_mult=(1,2,4,8), num_res_blocks, + attn_resolutions, dropout=0.0, resamp_with_conv=True, in_channels, + resolution, use_timestep=True, use_linear_attn=False, attn_type="vanilla"): + super().__init__() + if use_linear_attn: attn_type = "linear" + self.ch = ch + self.temb_ch = self.ch*4 + self.num_resolutions = len(ch_mult) + self.num_res_blocks = num_res_blocks + self.resolution = resolution + self.in_channels = in_channels + + self.use_timestep = use_timestep + if self.use_timestep: + # timestep embedding + self.temb = nn.Module() + self.temb.dense = nn.ModuleList([ + torch.nn.Linear(self.ch, + self.temb_ch), + torch.nn.Linear(self.temb_ch, + self.temb_ch), + ]) + + # downsampling + self.conv_in = torch.nn.Conv2d(in_channels, + self.ch, + kernel_size=3, + stride=1, + padding=1) + + curr_res = resolution + in_ch_mult = (1,)+tuple(ch_mult) + self.down = nn.ModuleList() + for i_level in range(self.num_resolutions): + block = nn.ModuleList() + attn = nn.ModuleList() + block_in = ch*in_ch_mult[i_level] + block_out = ch*ch_mult[i_level] + for i_block in range(self.num_res_blocks): + block.append(ResnetBlock(in_channels=block_in, + out_channels=block_out, + temb_channels=self.temb_ch, + dropout=dropout)) + block_in = block_out + if curr_res in attn_resolutions: + attn.append(make_attn(block_in, attn_type=attn_type)) + down = nn.Module() + down.block = block + down.attn = attn + if i_level != self.num_resolutions-1: + down.downsample = Downsample(block_in, resamp_with_conv) + curr_res = curr_res // 2 + self.down.append(down) + + # middle + self.mid = nn.Module() + self.mid.block_1 = ResnetBlock(in_channels=block_in, + out_channels=block_in, + temb_channels=self.temb_ch, + dropout=dropout) + self.mid.attn_1 = make_attn(block_in, attn_type=attn_type) + self.mid.block_2 = ResnetBlock(in_channels=block_in, + out_channels=block_in, + temb_channels=self.temb_ch, + dropout=dropout) + + # upsampling + self.up = nn.ModuleList() + for i_level in reversed(range(self.num_resolutions)): + block = nn.ModuleList() + attn = nn.ModuleList() + block_out = ch*ch_mult[i_level] + skip_in = ch*ch_mult[i_level] + for i_block in range(self.num_res_blocks+1): + if i_block == self.num_res_blocks: + skip_in = ch*in_ch_mult[i_level] + block.append(ResnetBlock(in_channels=block_in+skip_in, + out_channels=block_out, + temb_channels=self.temb_ch, + dropout=dropout)) + block_in = block_out + if curr_res in attn_resolutions: + attn.append(make_attn(block_in, attn_type=attn_type)) + up = nn.Module() + up.block = block + up.attn = attn + if i_level != 0: + up.upsample = Upsample(block_in, resamp_with_conv) + curr_res = curr_res * 2 + self.up.insert(0, up) # prepend to get consistent order + + # end + self.norm_out = Normalize(block_in) + self.conv_out = torch.nn.Conv2d(block_in, + out_ch, + kernel_size=3, + stride=1, + padding=1) + + def forward(self, x, t=None, context=None): + #assert x.shape[2] == x.shape[3] == self.resolution + if context is not None: + # assume aligned context, cat along channel axis + x = torch.cat((x, context), dim=1) + if self.use_timestep: + # timestep embedding + assert t is not None + temb = get_timestep_embedding(t, self.ch) + temb = self.temb.dense[0](temb) + temb = nonlinearity(temb) + temb = self.temb.dense[1](temb) + else: + temb = None + + # downsampling + hs = [self.conv_in(x)] + for i_level in range(self.num_resolutions): + for i_block in range(self.num_res_blocks): + h = self.down[i_level].block[i_block](hs[-1], temb) + if len(self.down[i_level].attn) > 0: + h = self.down[i_level].attn[i_block](h) + hs.append(h) + if i_level != self.num_resolutions-1: + hs.append(self.down[i_level].downsample(hs[-1])) + + # middle + h = hs[-1] + h = self.mid.block_1(h, temb) + h = self.mid.attn_1(h) + h = self.mid.block_2(h, temb) + + # upsampling + for i_level in reversed(range(self.num_resolutions)): + for i_block in range(self.num_res_blocks+1): + h = self.up[i_level].block[i_block]( + torch.cat([h, hs.pop()], dim=1), temb) + if len(self.up[i_level].attn) > 0: + h = self.up[i_level].attn[i_block](h) + if i_level != 0: + h = self.up[i_level].upsample(h) + + # end + h = self.norm_out(h) + h = nonlinearity(h) + h = self.conv_out(h) + return h + + def get_last_layer(self): + return self.conv_out.weight + + +class Encoder(nn.Module): + def __init__(self, *, ch, out_ch, ch_mult=(1,2,4,8), num_res_blocks, + attn_resolutions, dropout=0.0, resamp_with_conv=True, in_channels, + resolution, z_channels, double_z=True, use_linear_attn=False, attn_type="vanilla", + **ignore_kwargs): + super().__init__() + if use_linear_attn: attn_type = "linear" + self.ch = ch + self.temb_ch = 0 + self.num_resolutions = len(ch_mult) + self.num_res_blocks = num_res_blocks + self.resolution = resolution + self.in_channels = in_channels + + # downsampling + self.conv_in = torch.nn.Conv2d(in_channels, + self.ch, + kernel_size=3, + stride=1, + padding=1) + + curr_res = resolution + in_ch_mult = (1,)+tuple(ch_mult) + self.in_ch_mult = in_ch_mult + self.down = nn.ModuleList() + for i_level in range(self.num_resolutions): + block = nn.ModuleList() + attn = nn.ModuleList() + block_in = ch*in_ch_mult[i_level] + block_out = ch*ch_mult[i_level] + for i_block in range(self.num_res_blocks): + block.append(ResnetBlock(in_channels=block_in, + out_channels=block_out, + temb_channels=self.temb_ch, + dropout=dropout)) + block_in = block_out + if curr_res in attn_resolutions: + attn.append(make_attn(block_in, attn_type=attn_type)) + down = nn.Module() + down.block = block + down.attn = attn + if i_level != self.num_resolutions-1: + down.downsample = Downsample(block_in, resamp_with_conv) + curr_res = curr_res // 2 + self.down.append(down) + + # middle + self.mid = nn.Module() + self.mid.block_1 = ResnetBlock(in_channels=block_in, + out_channels=block_in, + temb_channels=self.temb_ch, + dropout=dropout) + self.mid.attn_1 = make_attn(block_in, attn_type=attn_type) + self.mid.block_2 = ResnetBlock(in_channels=block_in, + out_channels=block_in, + temb_channels=self.temb_ch, + dropout=dropout) + + # end + self.norm_out = Normalize(block_in) + self.conv_out = torch.nn.Conv2d(block_in, + 2*z_channels if double_z else z_channels, + kernel_size=3, + stride=1, + padding=1) + + def forward(self, x): + # timestep embedding + temb = None + + # downsampling + hs = [self.conv_in(x)] + for i_level in range(self.num_resolutions): + for i_block in range(self.num_res_blocks): + h = self.down[i_level].block[i_block](hs[-1], temb) + if len(self.down[i_level].attn) > 0: + h = self.down[i_level].attn[i_block](h) + hs.append(h) + if i_level != self.num_resolutions-1: + hs.append(self.down[i_level].downsample(hs[-1])) + + # middle + h = hs[-1] + h = self.mid.block_1(h, temb) + h = self.mid.attn_1(h) + h = self.mid.block_2(h, temb) + + # end + h = self.norm_out(h) + h = nonlinearity(h) + h = self.conv_out(h) + return h + + +class Decoder(nn.Module): + def __init__(self, *, ch, out_ch, ch_mult=(1,2,4,8), num_res_blocks, + attn_resolutions, dropout=0.0, resamp_with_conv=True, in_channels, + resolution, z_channels, give_pre_end=False, tanh_out=False, use_linear_attn=False, + attn_type="vanilla", **ignorekwargs): + super().__init__() + if use_linear_attn: attn_type = "linear" + self.ch = ch + self.temb_ch = 0 + self.num_resolutions = len(ch_mult) + self.num_res_blocks = num_res_blocks + self.resolution = resolution + self.in_channels = in_channels + self.give_pre_end = give_pre_end + self.tanh_out = tanh_out + + # compute in_ch_mult, block_in and curr_res at lowest res + in_ch_mult = (1,)+tuple(ch_mult) + block_in = ch*ch_mult[self.num_resolutions-1] + curr_res = resolution // 2**(self.num_resolutions-1) + self.z_shape = (1,z_channels,curr_res,curr_res) + print("Working with z of shape {} = {} dimensions.".format( + self.z_shape, np.prod(self.z_shape))) + + # z to block_in + self.conv_in = torch.nn.Conv2d(z_channels, + block_in, + kernel_size=3, + stride=1, + padding=1) + + # middle + self.mid = nn.Module() + self.mid.block_1 = ResnetBlock(in_channels=block_in, + out_channels=block_in, + temb_channels=self.temb_ch, + dropout=dropout) + self.mid.attn_1 = make_attn(block_in, attn_type=attn_type) + self.mid.block_2 = ResnetBlock(in_channels=block_in, + out_channels=block_in, + temb_channels=self.temb_ch, + dropout=dropout) + + # upsampling + self.up = nn.ModuleList() + for i_level in reversed(range(self.num_resolutions)): + block = nn.ModuleList() + attn = nn.ModuleList() + block_out = ch*ch_mult[i_level] + for i_block in range(self.num_res_blocks+1): + block.append(ResnetBlock(in_channels=block_in, + out_channels=block_out, + temb_channels=self.temb_ch, + dropout=dropout)) + block_in = block_out + if curr_res in attn_resolutions: + attn.append(make_attn(block_in, attn_type=attn_type)) + up = nn.Module() + up.block = block + up.attn = attn + if i_level != 0: + up.upsample = Upsample(block_in, resamp_with_conv) + curr_res = curr_res * 2 + self.up.insert(0, up) # prepend to get consistent order + + # end + self.norm_out = Normalize(block_in) + self.conv_out = torch.nn.Conv2d(block_in, + out_ch, + kernel_size=3, + stride=1, + padding=1) + + def forward(self, z): + #assert z.shape[1:] == self.z_shape[1:] + self.last_z_shape = z.shape + + # timestep embedding + temb = None + + # z to block_in + h = self.conv_in(z) + + # middle + h = self.mid.block_1(h, temb) + h = self.mid.attn_1(h) + h = self.mid.block_2(h, temb) + + # upsampling + for i_level in reversed(range(self.num_resolutions)): + for i_block in range(self.num_res_blocks+1): + h = self.up[i_level].block[i_block](h, temb) + if len(self.up[i_level].attn) > 0: + h = self.up[i_level].attn[i_block](h) + if i_level != 0: + h = self.up[i_level].upsample(h) + + # end + if self.give_pre_end: + return h + + h = self.norm_out(h) + h = nonlinearity(h) + h = self.conv_out(h) + if self.tanh_out: + h = torch.tanh(h) + return h + + +class SimpleDecoder(nn.Module): + def __init__(self, in_channels, out_channels, *args, **kwargs): + super().__init__() + self.model = nn.ModuleList([nn.Conv2d(in_channels, in_channels, 1), + ResnetBlock(in_channels=in_channels, + out_channels=2 * in_channels, + temb_channels=0, dropout=0.0), + ResnetBlock(in_channels=2 * in_channels, + out_channels=4 * in_channels, + temb_channels=0, dropout=0.0), + ResnetBlock(in_channels=4 * in_channels, + out_channels=2 * in_channels, + temb_channels=0, dropout=0.0), + nn.Conv2d(2*in_channels, in_channels, 1), + Upsample(in_channels, with_conv=True)]) + # end + self.norm_out = Normalize(in_channels) + self.conv_out = torch.nn.Conv2d(in_channels, + out_channels, + kernel_size=3, + stride=1, + padding=1) + + def forward(self, x): + for i, layer in enumerate(self.model): + if i in [1,2,3]: + x = layer(x, None) + else: + x = layer(x) + + h = self.norm_out(x) + h = nonlinearity(h) + x = self.conv_out(h) + return x + + +class UpsampleDecoder(nn.Module): + def __init__(self, in_channels, out_channels, ch, num_res_blocks, resolution, + ch_mult=(2,2), dropout=0.0): + super().__init__() + # upsampling + self.temb_ch = 0 + self.num_resolutions = len(ch_mult) + self.num_res_blocks = num_res_blocks + block_in = in_channels + curr_res = resolution // 2 ** (self.num_resolutions - 1) + self.res_blocks = nn.ModuleList() + self.upsample_blocks = nn.ModuleList() + for i_level in range(self.num_resolutions): + res_block = [] + block_out = ch * ch_mult[i_level] + for i_block in range(self.num_res_blocks + 1): + res_block.append(ResnetBlock(in_channels=block_in, + out_channels=block_out, + temb_channels=self.temb_ch, + dropout=dropout)) + block_in = block_out + self.res_blocks.append(nn.ModuleList(res_block)) + if i_level != self.num_resolutions - 1: + self.upsample_blocks.append(Upsample(block_in, True)) + curr_res = curr_res * 2 + + # end + self.norm_out = Normalize(block_in) + self.conv_out = torch.nn.Conv2d(block_in, + out_channels, + kernel_size=3, + stride=1, + padding=1) + + def forward(self, x): + # upsampling + h = x + for k, i_level in enumerate(range(self.num_resolutions)): + for i_block in range(self.num_res_blocks + 1): + h = self.res_blocks[i_level][i_block](h, None) + if i_level != self.num_resolutions - 1: + h = self.upsample_blocks[k](h) + h = self.norm_out(h) + h = nonlinearity(h) + h = self.conv_out(h) + return h + + +class LatentRescaler(nn.Module): + def __init__(self, factor, in_channels, mid_channels, out_channels, depth=2): + super().__init__() + # residual block, interpolate, residual block + self.factor = factor + self.conv_in = nn.Conv2d(in_channels, + mid_channels, + kernel_size=3, + stride=1, + padding=1) + self.res_block1 = nn.ModuleList([ResnetBlock(in_channels=mid_channels, + out_channels=mid_channels, + temb_channels=0, + dropout=0.0) for _ in range(depth)]) + self.attn = AttnBlock(mid_channels) + self.res_block2 = nn.ModuleList([ResnetBlock(in_channels=mid_channels, + out_channels=mid_channels, + temb_channels=0, + dropout=0.0) for _ in range(depth)]) + + self.conv_out = nn.Conv2d(mid_channels, + out_channels, + kernel_size=1, + ) + + def forward(self, x): + x = self.conv_in(x) + for block in self.res_block1: + x = block(x, None) + x = torch.nn.functional.interpolate(x, size=(int(round(x.shape[2]*self.factor)), int(round(x.shape[3]*self.factor)))) + x = self.attn(x) + for block in self.res_block2: + x = block(x, None) + x = self.conv_out(x) + return x + + +class MergedRescaleEncoder(nn.Module): + def __init__(self, in_channels, ch, resolution, out_ch, num_res_blocks, + attn_resolutions, dropout=0.0, resamp_with_conv=True, + ch_mult=(1,2,4,8), rescale_factor=1.0, rescale_module_depth=1): + super().__init__() + intermediate_chn = ch * ch_mult[-1] + self.encoder = Encoder(in_channels=in_channels, num_res_blocks=num_res_blocks, ch=ch, ch_mult=ch_mult, + z_channels=intermediate_chn, double_z=False, resolution=resolution, + attn_resolutions=attn_resolutions, dropout=dropout, resamp_with_conv=resamp_with_conv, + out_ch=None) + self.rescaler = LatentRescaler(factor=rescale_factor, in_channels=intermediate_chn, + mid_channels=intermediate_chn, out_channels=out_ch, depth=rescale_module_depth) + + def forward(self, x): + x = self.encoder(x) + x = self.rescaler(x) + return x + + +class MergedRescaleDecoder(nn.Module): + def __init__(self, z_channels, out_ch, resolution, num_res_blocks, attn_resolutions, ch, ch_mult=(1,2,4,8), + dropout=0.0, resamp_with_conv=True, rescale_factor=1.0, rescale_module_depth=1): + super().__init__() + tmp_chn = z_channels*ch_mult[-1] + self.decoder = Decoder(out_ch=out_ch, z_channels=tmp_chn, attn_resolutions=attn_resolutions, dropout=dropout, + resamp_with_conv=resamp_with_conv, in_channels=None, num_res_blocks=num_res_blocks, + ch_mult=ch_mult, resolution=resolution, ch=ch) + self.rescaler = LatentRescaler(factor=rescale_factor, in_channels=z_channels, mid_channels=tmp_chn, + out_channels=tmp_chn, depth=rescale_module_depth) + + def forward(self, x): + x = self.rescaler(x) + x = self.decoder(x) + return x + + +class Upsampler(nn.Module): + def __init__(self, in_size, out_size, in_channels, out_channels, ch_mult=2): + super().__init__() + assert out_size >= in_size + num_blocks = int(np.log2(out_size//in_size))+1 + factor_up = 1.+ (out_size % in_size) + print(f"Building {self.__class__.__name__} with in_size: {in_size} --> out_size {out_size} and factor {factor_up}") + self.rescaler = LatentRescaler(factor=factor_up, in_channels=in_channels, mid_channels=2*in_channels, + out_channels=in_channels) + self.decoder = Decoder(out_ch=out_channels, resolution=out_size, z_channels=in_channels, num_res_blocks=2, + attn_resolutions=[], in_channels=None, ch=in_channels, + ch_mult=[ch_mult for _ in range(num_blocks)]) + + def forward(self, x): + x = self.rescaler(x) + x = self.decoder(x) + return x + + +class Resize(nn.Module): + def __init__(self, in_channels=None, learned=False, mode="bilinear"): + super().__init__() + self.with_conv = learned + self.mode = mode + if self.with_conv: + print(f"Note: {self.__class__.__name} uses learned downsampling and will ignore the fixed {mode} mode") + raise NotImplementedError() + assert in_channels is not None + # no asymmetric padding in torch conv, must do it ourselves + self.conv = torch.nn.Conv2d(in_channels, + in_channels, + kernel_size=4, + stride=2, + padding=1) + + def forward(self, x, scale_factor=1.0): + if scale_factor==1.0: + return x + else: + x = torch.nn.functional.interpolate(x, mode=self.mode, align_corners=False, scale_factor=scale_factor) + return x diff --git a/ldm/modules/diffusionmodules/openaimodel.py b/ldm/modules/diffusionmodules/openaimodel.py new file mode 100644 index 0000000000000000000000000000000000000000..cc3875c63c86fb2e33a83de70a9601cba6b39148 --- /dev/null +++ b/ldm/modules/diffusionmodules/openaimodel.py @@ -0,0 +1,807 @@ +from abc import abstractmethod +import math + +import numpy as np +import torch as th +import torch.nn as nn +import torch.nn.functional as F + +from ldm.modules.diffusionmodules.util import ( + checkpoint, + conv_nd, + linear, + avg_pool_nd, + zero_module, + normalization, + timestep_embedding, +) +from ldm.modules.attention import SpatialTransformer +from ldm.util import exists + + +# dummy replace +def convert_module_to_f16(x): + pass + +def convert_module_to_f32(x): + pass + + +## go +class AttentionPool2d(nn.Module): + """ + Adapted from CLIP: https://github.com/openai/CLIP/blob/main/clip/model.py + """ + + def __init__( + self, + spacial_dim: int, + embed_dim: int, + num_heads_channels: int, + output_dim: int = None, + ): + super().__init__() + self.positional_embedding = nn.Parameter(th.randn(embed_dim, spacial_dim ** 2 + 1) / embed_dim ** 0.5) + self.qkv_proj = conv_nd(1, embed_dim, 3 * embed_dim, 1) + self.c_proj = conv_nd(1, embed_dim, output_dim or embed_dim, 1) + self.num_heads = embed_dim // num_heads_channels + self.attention = QKVAttention(self.num_heads) + + def forward(self, x): + b, c, *_spatial = x.shape + x = x.reshape(b, c, -1) # NC(HW) + x = th.cat([x.mean(dim=-1, keepdim=True), x], dim=-1) # NC(HW+1) + x = x + self.positional_embedding[None, :, :].to(x.dtype) # NC(HW+1) + x = self.qkv_proj(x) + x = self.attention(x) + x = self.c_proj(x) + return x[:, :, 0] + + +class TimestepBlock(nn.Module): + """ + Any module where forward() takes timestep embeddings as a second argument. + """ + + @abstractmethod + def forward(self, x, emb): + """ + Apply the module to `x` given `emb` timestep embeddings. + """ + + +class TimestepEmbedSequential(nn.Sequential, TimestepBlock): + """ + A sequential module that passes timestep embeddings to the children that + support it as an extra input. + """ + + def forward(self, x, emb, context=None): + for layer in self: + if isinstance(layer, TimestepBlock): + x = layer(x, emb) + elif isinstance(layer, SpatialTransformer): + x = layer(x, context) + else: + x = layer(x) + return x + + +class Upsample(nn.Module): + """ + An upsampling layer with an optional convolution. + :param channels: channels in the inputs and outputs. + :param use_conv: a bool determining if a convolution is applied. + :param dims: determines if the signal is 1D, 2D, or 3D. If 3D, then + upsampling occurs in the inner-two dimensions. + """ + + def __init__(self, channels, use_conv, dims=2, out_channels=None, padding=1): + super().__init__() + self.channels = channels + self.out_channels = out_channels or channels + self.use_conv = use_conv + self.dims = dims + if use_conv: + self.conv = conv_nd(dims, self.channels, self.out_channels, 3, padding=padding) + + def forward(self, x): + assert x.shape[1] == self.channels + if self.dims == 3: + x = F.interpolate( + x, (x.shape[2], x.shape[3] * 2, x.shape[4] * 2), mode="nearest" + ) + else: + x = F.interpolate(x, scale_factor=2, mode="nearest") + if self.use_conv: + x = self.conv(x) + return x + +class TransposedUpsample(nn.Module): + 'Learned 2x upsampling without padding' + def __init__(self, channels, out_channels=None, ks=5): + super().__init__() + self.channels = channels + self.out_channels = out_channels or channels + + self.up = nn.ConvTranspose2d(self.channels,self.out_channels,kernel_size=ks,stride=2) + + def forward(self,x): + return self.up(x) + + +class Downsample(nn.Module): + """ + A downsampling layer with an optional convolution. + :param channels: channels in the inputs and outputs. + :param use_conv: a bool determining if a convolution is applied. + :param dims: determines if the signal is 1D, 2D, or 3D. If 3D, then + downsampling occurs in the inner-two dimensions. + """ + + def __init__(self, channels, use_conv, dims=2, out_channels=None,padding=1): + super().__init__() + self.channels = channels + self.out_channels = out_channels or channels + self.use_conv = use_conv + self.dims = dims + stride = 2 if dims != 3 else (1, 2, 2) + if use_conv: + self.op = conv_nd( + dims, self.channels, self.out_channels, 3, stride=stride, padding=padding + ) + else: + assert self.channels == self.out_channels + self.op = avg_pool_nd(dims, kernel_size=stride, stride=stride) + + def forward(self, x): + assert x.shape[1] == self.channels + return self.op(x) + + +class ResBlock(TimestepBlock): + """ + A residual block that can optionally change the number of channels. + :param channels: the number of input channels. + :param emb_channels: the number of timestep embedding channels. + :param dropout: the rate of dropout. + :param out_channels: if specified, the number of out channels. + :param use_conv: if True and out_channels is specified, use a spatial + convolution instead of a smaller 1x1 convolution to change the + channels in the skip connection. + :param dims: determines if the signal is 1D, 2D, or 3D. + :param use_checkpoint: if True, use gradient checkpointing on this module. + :param up: if True, use this block for upsampling. + :param down: if True, use this block for downsampling. + """ + + def __init__( + self, + channels, + emb_channels, + dropout, + out_channels=None, + use_conv=False, + use_scale_shift_norm=False, + dims=2, + use_checkpoint=False, + up=False, + down=False, + ): + super().__init__() + self.channels = channels + self.emb_channels = emb_channels + self.dropout = dropout + self.out_channels = out_channels or channels + self.use_conv = use_conv + self.use_checkpoint = use_checkpoint + self.use_scale_shift_norm = use_scale_shift_norm + + self.in_layers = nn.Sequential( + normalization(channels), + nn.SiLU(), + conv_nd(dims, channels, self.out_channels, 3, padding=1), + ) + + self.updown = up or down + + if up: + self.h_upd = Upsample(channels, False, dims) + self.x_upd = Upsample(channels, False, dims) + elif down: + self.h_upd = Downsample(channels, False, dims) + self.x_upd = Downsample(channels, False, dims) + else: + self.h_upd = self.x_upd = nn.Identity() + + self.emb_layers = nn.Sequential( + nn.SiLU(), + linear( + emb_channels, + 2 * self.out_channels if use_scale_shift_norm else self.out_channels, + ), + ) + self.out_layers = nn.Sequential( + normalization(self.out_channels), + nn.SiLU(), + nn.Dropout(p=dropout), + zero_module( + conv_nd(dims, self.out_channels, self.out_channels, 3, padding=1) + ), + ) + + if self.out_channels == channels: + self.skip_connection = nn.Identity() + elif use_conv: + self.skip_connection = conv_nd( + dims, channels, self.out_channels, 3, padding=1 + ) + else: + self.skip_connection = conv_nd(dims, channels, self.out_channels, 1) + + def forward(self, x, emb): + """ + Apply the block to a Tensor, conditioned on a timestep embedding. + :param x: an [N x C x ...] Tensor of features. + :param emb: an [N x emb_channels] Tensor of timestep embeddings. + :return: an [N x C x ...] Tensor of outputs. + """ + return checkpoint( + self._forward, (x, emb), self.parameters(), self.use_checkpoint + ) + + + def _forward(self, x, emb): + if self.updown: + in_rest, in_conv = self.in_layers[:-1], self.in_layers[-1] + h = in_rest(x) + h = self.h_upd(h) + x = self.x_upd(x) + h = in_conv(h) + else: + h = self.in_layers(x) + emb_out = self.emb_layers(emb).type(h.dtype) + while len(emb_out.shape) < len(h.shape): + emb_out = emb_out[..., None] + if self.use_scale_shift_norm: + out_norm, out_rest = self.out_layers[0], self.out_layers[1:] + scale, shift = th.chunk(emb_out, 2, dim=1) + h = out_norm(h) * (1 + scale) + shift + h = out_rest(h) + else: + h = h + emb_out + h = self.out_layers(h) + return self.skip_connection(x) + h + + +class AttentionBlock(nn.Module): + """ + An attention block that allows spatial positions to attend to each other. + Originally ported from here, but adapted to the N-d case. + https://github.com/hojonathanho/diffusion/blob/1e0dceb3b3495bbe19116a5e1b3596cd0706c543/diffusion_tf/models/unet.py#L66. + """ + + def __init__( + self, + channels, + num_heads=1, + num_head_channels=-1, + use_checkpoint=False, + use_new_attention_order=False, + ): + super().__init__() + self.channels = channels + if num_head_channels == -1: + self.num_heads = num_heads + else: + assert ( + channels % num_head_channels == 0 + ), f"q,k,v channels {channels} is not divisible by num_head_channels {num_head_channels}" + self.num_heads = channels // num_head_channels + self.use_checkpoint = use_checkpoint + self.norm = normalization(channels) + self.qkv = conv_nd(1, channels, channels * 3, 1) + if use_new_attention_order: + # split qkv before split heads + self.attention = QKVAttention(self.num_heads) + else: + # split heads before split qkv + self.attention = QKVAttentionLegacy(self.num_heads) + + self.proj_out = zero_module(conv_nd(1, channels, channels, 1)) + + def forward(self, x): + return checkpoint(self._forward, (x,), self.parameters(), True) # TODO: check checkpoint usage, is True # TODO: fix the .half call!!! + #return pt_checkpoint(self._forward, x) # pytorch + + def _forward(self, x): + b, c, *spatial = x.shape + x = x.reshape(b, c, -1) + qkv = self.qkv(self.norm(x)) + h = self.attention(qkv) + h = self.proj_out(h) + return (x + h).reshape(b, c, *spatial) + + +def count_flops_attn(model, _x, y): + """ + A counter for the `thop` package to count the operations in an + attention operation. + Meant to be used like: + macs, params = thop.profile( + model, + inputs=(inputs, timestamps), + custom_ops={QKVAttention: QKVAttention.count_flops}, + ) + """ + b, c, *spatial = y[0].shape + num_spatial = int(np.prod(spatial)) + # We perform two matmuls with the same number of ops. + # The first computes the weight matrix, the second computes + # the combination of the value vectors. + matmul_ops = 2 * b * (num_spatial ** 2) * c + model.total_ops += th.DoubleTensor([matmul_ops]) + + +class QKVAttentionLegacy(nn.Module): + """ + A module which performs QKV attention. Matches legacy QKVAttention + input/ouput heads shaping + """ + + def __init__(self, n_heads): + super().__init__() + self.n_heads = n_heads + + def forward(self, qkv): + """ + Apply QKV attention. + :param qkv: an [N x (H * 3 * C) x T] tensor of Qs, Ks, and Vs. + :return: an [N x (H * C) x T] tensor after attention. + """ + bs, width, length = qkv.shape + assert width % (3 * self.n_heads) == 0 + ch = width // (3 * self.n_heads) + q, k, v = qkv.reshape(bs * self.n_heads, ch * 3, length).split(ch, dim=1) + scale = 1 / math.sqrt(math.sqrt(ch)) + weight = th.einsum( + "bct,bcs->bts", q * scale, k * scale + ) # More stable with f16 than dividing afterwards + weight = th.softmax(weight.float(), dim=-1).type(weight.dtype) + a = th.einsum("bts,bcs->bct", weight, v) + return a.reshape(bs, -1, length) + + @staticmethod + def count_flops(model, _x, y): + return count_flops_attn(model, _x, y) + + +class QKVAttention(nn.Module): + """ + A module which performs QKV attention and splits in a different order. + """ + + def __init__(self, n_heads): + super().__init__() + self.n_heads = n_heads + + def forward(self, qkv): + """ + Apply QKV attention. + :param qkv: an [N x (3 * H * C) x T] tensor of Qs, Ks, and Vs. + :return: an [N x (H * C) x T] tensor after attention. + """ + bs, width, length = qkv.shape + assert width % (3 * self.n_heads) == 0 + ch = width // (3 * self.n_heads) + q, k, v = qkv.chunk(3, dim=1) + scale = 1 / math.sqrt(math.sqrt(ch)) + weight = th.einsum( + "bct,bcs->bts", + (q * scale).view(bs * self.n_heads, ch, length), + (k * scale).view(bs * self.n_heads, ch, length), + ) # More stable with f16 than dividing afterwards + weight = th.softmax(weight.float(), dim=-1).type(weight.dtype) + a = th.einsum("bts,bcs->bct", weight, v.reshape(bs * self.n_heads, ch, length)) + return a.reshape(bs, -1, length) + + @staticmethod + def count_flops(model, _x, y): + return count_flops_attn(model, _x, y) + + +class Timestep(nn.Module): + def __init__(self, dim): + super().__init__() + self.dim = dim + + def forward(self, t): + return timestep_embedding(t, self.dim) + + +class UNetModel(nn.Module): + """ + The full UNet model with attention and timestep embedding. + :param in_channels: channels in the input Tensor. + :param model_channels: base channel count for the model. + :param out_channels: channels in the output Tensor. + :param num_res_blocks: number of residual blocks per downsample. + :param attention_resolutions: a collection of downsample rates at which + attention will take place. May be a set, list, or tuple. + For example, if this contains 4, then at 4x downsampling, attention + will be used. + :param dropout: the dropout probability. + :param channel_mult: channel multiplier for each level of the UNet. + :param conv_resample: if True, use learned convolutions for upsampling and + downsampling. + :param dims: determines if the signal is 1D, 2D, or 3D. + :param num_classes: if specified (as an int), then this model will be + class-conditional with `num_classes` classes. + :param use_checkpoint: use gradient checkpointing to reduce memory usage. + :param num_heads: the number of attention heads in each attention layer. + :param num_heads_channels: if specified, ignore num_heads and instead use + a fixed channel width per attention head. + :param num_heads_upsample: works with num_heads to set a different number + of heads for upsampling. Deprecated. + :param use_scale_shift_norm: use a FiLM-like conditioning mechanism. + :param resblock_updown: use residual blocks for up/downsampling. + :param use_new_attention_order: use a different attention pattern for potentially + increased efficiency. + """ + + def __init__( + self, + image_size, + in_channels, + model_channels, + out_channels, + num_res_blocks, + attention_resolutions, + dropout=0, + channel_mult=(1, 2, 4, 8), + conv_resample=True, + dims=2, + num_classes=None, + use_checkpoint=False, + use_fp16=False, + use_bf16=False, + num_heads=-1, + num_head_channels=-1, + num_heads_upsample=-1, + use_scale_shift_norm=False, + resblock_updown=False, + use_new_attention_order=False, + use_spatial_transformer=False, # custom transformer support + transformer_depth=1, # custom transformer support + context_dim=None, # custom transformer support + n_embed=None, # custom support for prediction of discrete ids into codebook of first stage vq model + legacy=True, + disable_self_attentions=None, + num_attention_blocks=None, + disable_middle_self_attn=False, + use_linear_in_transformer=False, + adm_in_channels=None, + ): + super().__init__() + if use_spatial_transformer: + assert context_dim is not None, 'Fool!! You forgot to include the dimension of your cross-attention conditioning...' + + if context_dim is not None: + assert use_spatial_transformer, 'Fool!! You forgot to use the spatial transformer for your cross-attention conditioning...' + from omegaconf.listconfig import ListConfig + if type(context_dim) == ListConfig: + context_dim = list(context_dim) + + if num_heads_upsample == -1: + num_heads_upsample = num_heads + + if num_heads == -1: + assert num_head_channels != -1, 'Either num_heads or num_head_channels has to be set' + + if num_head_channels == -1: + assert num_heads != -1, 'Either num_heads or num_head_channels has to be set' + + self.image_size = image_size + self.in_channels = in_channels + self.model_channels = model_channels + self.out_channels = out_channels + if isinstance(num_res_blocks, int): + self.num_res_blocks = len(channel_mult) * [num_res_blocks] + else: + if len(num_res_blocks) != len(channel_mult): + raise ValueError("provide num_res_blocks either as an int (globally constant) or " + "as a list/tuple (per-level) with the same length as channel_mult") + self.num_res_blocks = num_res_blocks + if disable_self_attentions is not None: + # should be a list of booleans, indicating whether to disable self-attention in TransformerBlocks or not + assert len(disable_self_attentions) == len(channel_mult) + if num_attention_blocks is not None: + assert len(num_attention_blocks) == len(self.num_res_blocks) + assert all(map(lambda i: self.num_res_blocks[i] >= num_attention_blocks[i], range(len(num_attention_blocks)))) + print(f"Constructor of UNetModel received num_attention_blocks={num_attention_blocks}. " + f"This option has LESS priority than attention_resolutions {attention_resolutions}, " + f"i.e., in cases where num_attention_blocks[i] > 0 but 2**i not in attention_resolutions, " + f"attention will still not be set.") + + self.attention_resolutions = attention_resolutions + self.dropout = dropout + self.channel_mult = channel_mult + self.conv_resample = conv_resample + self.num_classes = num_classes + self.use_checkpoint = use_checkpoint + self.dtype = th.float16 if use_fp16 else th.float32 + self.dtype = th.bfloat16 if use_bf16 else self.dtype + self.num_heads = num_heads + self.num_head_channels = num_head_channels + self.num_heads_upsample = num_heads_upsample + self.predict_codebook_ids = n_embed is not None + + time_embed_dim = model_channels * 4 + self.time_embed = nn.Sequential( + linear(model_channels, time_embed_dim), + nn.SiLU(), + linear(time_embed_dim, time_embed_dim), + ) + + if self.num_classes is not None: + if isinstance(self.num_classes, int): + self.label_emb = nn.Embedding(num_classes, time_embed_dim) + elif self.num_classes == "continuous": + print("setting up linear c_adm embedding layer") + self.label_emb = nn.Linear(1, time_embed_dim) + elif self.num_classes == "sequential": + assert adm_in_channels is not None + self.label_emb = nn.Sequential( + nn.Sequential( + linear(adm_in_channels, time_embed_dim), + nn.SiLU(), + linear(time_embed_dim, time_embed_dim), + ) + ) + else: + raise ValueError() + + self.input_blocks = nn.ModuleList( + [ + TimestepEmbedSequential( + conv_nd(dims, in_channels, model_channels, 3, padding=1) + ) + ] + ) + self._feature_size = model_channels + input_block_chans = [model_channels] + ch = model_channels + ds = 1 + for level, mult in enumerate(channel_mult): + for nr in range(self.num_res_blocks[level]): + layers = [ + ResBlock( + ch, + time_embed_dim, + dropout, + out_channels=mult * model_channels, + dims=dims, + use_checkpoint=use_checkpoint, + use_scale_shift_norm=use_scale_shift_norm, + ) + ] + ch = mult * model_channels + if ds in attention_resolutions: + if num_head_channels == -1: + dim_head = ch // num_heads + else: + num_heads = ch // num_head_channels + dim_head = num_head_channels + if legacy: + #num_heads = 1 + dim_head = ch // num_heads if use_spatial_transformer else num_head_channels + if exists(disable_self_attentions): + disabled_sa = disable_self_attentions[level] + else: + disabled_sa = False + + if not exists(num_attention_blocks) or nr < num_attention_blocks[level]: + layers.append( + AttentionBlock( + ch, + use_checkpoint=use_checkpoint, + num_heads=num_heads, + num_head_channels=dim_head, + use_new_attention_order=use_new_attention_order, + ) if not use_spatial_transformer else SpatialTransformer( + ch, num_heads, dim_head, depth=transformer_depth, context_dim=context_dim, + disable_self_attn=disabled_sa, use_linear=use_linear_in_transformer, + use_checkpoint=use_checkpoint + ) + ) + self.input_blocks.append(TimestepEmbedSequential(*layers)) + self._feature_size += ch + input_block_chans.append(ch) + if level != len(channel_mult) - 1: + out_ch = ch + self.input_blocks.append( + TimestepEmbedSequential( + ResBlock( + ch, + time_embed_dim, + dropout, + out_channels=out_ch, + dims=dims, + use_checkpoint=use_checkpoint, + use_scale_shift_norm=use_scale_shift_norm, + down=True, + ) + if resblock_updown + else Downsample( + ch, conv_resample, dims=dims, out_channels=out_ch + ) + ) + ) + ch = out_ch + input_block_chans.append(ch) + ds *= 2 + self._feature_size += ch + + if num_head_channels == -1: + dim_head = ch // num_heads + else: + num_heads = ch // num_head_channels + dim_head = num_head_channels + if legacy: + #num_heads = 1 + dim_head = ch // num_heads if use_spatial_transformer else num_head_channels + self.middle_block = TimestepEmbedSequential( + ResBlock( + ch, + time_embed_dim, + dropout, + dims=dims, + use_checkpoint=use_checkpoint, + use_scale_shift_norm=use_scale_shift_norm, + ), + AttentionBlock( + ch, + use_checkpoint=use_checkpoint, + num_heads=num_heads, + num_head_channels=dim_head, + use_new_attention_order=use_new_attention_order, + ) if not use_spatial_transformer else SpatialTransformer( # always uses a self-attn + ch, num_heads, dim_head, depth=transformer_depth, context_dim=context_dim, + disable_self_attn=disable_middle_self_attn, use_linear=use_linear_in_transformer, + use_checkpoint=use_checkpoint + ), + ResBlock( + ch, + time_embed_dim, + dropout, + dims=dims, + use_checkpoint=use_checkpoint, + use_scale_shift_norm=use_scale_shift_norm, + ), + ) + self._feature_size += ch + + self.output_blocks = nn.ModuleList([]) + for level, mult in list(enumerate(channel_mult))[::-1]: + for i in range(self.num_res_blocks[level] + 1): + ich = input_block_chans.pop() + layers = [ + ResBlock( + ch + ich, + time_embed_dim, + dropout, + out_channels=model_channels * mult, + dims=dims, + use_checkpoint=use_checkpoint, + use_scale_shift_norm=use_scale_shift_norm, + ) + ] + ch = model_channels * mult + if ds in attention_resolutions: + if num_head_channels == -1: + dim_head = ch // num_heads + else: + num_heads = ch // num_head_channels + dim_head = num_head_channels + if legacy: + #num_heads = 1 + dim_head = ch // num_heads if use_spatial_transformer else num_head_channels + if exists(disable_self_attentions): + disabled_sa = disable_self_attentions[level] + else: + disabled_sa = False + + if not exists(num_attention_blocks) or i < num_attention_blocks[level]: + layers.append( + AttentionBlock( + ch, + use_checkpoint=use_checkpoint, + num_heads=num_heads_upsample, + num_head_channels=dim_head, + use_new_attention_order=use_new_attention_order, + ) if not use_spatial_transformer else SpatialTransformer( + ch, num_heads, dim_head, depth=transformer_depth, context_dim=context_dim, + disable_self_attn=disabled_sa, use_linear=use_linear_in_transformer, + use_checkpoint=use_checkpoint + ) + ) + if level and i == self.num_res_blocks[level]: + out_ch = ch + layers.append( + ResBlock( + ch, + time_embed_dim, + dropout, + out_channels=out_ch, + dims=dims, + use_checkpoint=use_checkpoint, + use_scale_shift_norm=use_scale_shift_norm, + up=True, + ) + if resblock_updown + else Upsample(ch, conv_resample, dims=dims, out_channels=out_ch) + ) + ds //= 2 + self.output_blocks.append(TimestepEmbedSequential(*layers)) + self._feature_size += ch + + self.out = nn.Sequential( + normalization(ch), + nn.SiLU(), + zero_module(conv_nd(dims, model_channels, out_channels, 3, padding=1)), + ) + if self.predict_codebook_ids: + self.id_predictor = nn.Sequential( + normalization(ch), + conv_nd(dims, model_channels, n_embed, 1), + #nn.LogSoftmax(dim=1) # change to cross_entropy and produce non-normalized logits + ) + + def convert_to_fp16(self): + """ + Convert the torso of the model to float16. + """ + self.input_blocks.apply(convert_module_to_f16) + self.middle_block.apply(convert_module_to_f16) + self.output_blocks.apply(convert_module_to_f16) + + def convert_to_fp32(self): + """ + Convert the torso of the model to float32. + """ + self.input_blocks.apply(convert_module_to_f32) + self.middle_block.apply(convert_module_to_f32) + self.output_blocks.apply(convert_module_to_f32) + + def forward(self, x, timesteps=None, context=None, y=None,**kwargs): + """ + Apply the model to an input batch. + :param x: an [N x C x ...] Tensor of inputs. + :param timesteps: a 1-D batch of timesteps. + :param context: conditioning plugged in via crossattn + :param y: an [N] Tensor of labels, if class-conditional. + :return: an [N x C x ...] Tensor of outputs. + """ + assert (y is not None) == ( + self.num_classes is not None + ), "must specify y if and only if the model is class-conditional" + hs = [] + t_emb = timestep_embedding(timesteps, self.model_channels, repeat_only=False) + emb = self.time_embed(t_emb) + + if self.num_classes is not None: + assert y.shape[0] == x.shape[0] + emb = emb + self.label_emb(y) + + h = x.type(self.dtype) + for module in self.input_blocks: + h = module(h, emb, context) + hs.append(h) + h = self.middle_block(h, emb, context) + for module in self.output_blocks: + h = th.cat([h, hs.pop()], dim=1) + h = module(h, emb, context) + h = h.type(x.dtype) + if self.predict_codebook_ids: + return self.id_predictor(h) + else: + return self.out(h) diff --git a/ldm/modules/diffusionmodules/upscaling.py b/ldm/modules/diffusionmodules/upscaling.py new file mode 100644 index 0000000000000000000000000000000000000000..03816662098ce1ffac79bd939b892e867ab91988 --- /dev/null +++ b/ldm/modules/diffusionmodules/upscaling.py @@ -0,0 +1,81 @@ +import torch +import torch.nn as nn +import numpy as np +from functools import partial + +from ldm.modules.diffusionmodules.util import extract_into_tensor, make_beta_schedule +from ldm.util import default + + +class AbstractLowScaleModel(nn.Module): + # for concatenating a downsampled image to the latent representation + def __init__(self, noise_schedule_config=None): + super(AbstractLowScaleModel, self).__init__() + if noise_schedule_config is not None: + self.register_schedule(**noise_schedule_config) + + def register_schedule(self, beta_schedule="linear", timesteps=1000, + linear_start=1e-4, linear_end=2e-2, cosine_s=8e-3): + betas = make_beta_schedule(beta_schedule, timesteps, linear_start=linear_start, linear_end=linear_end, + cosine_s=cosine_s) + alphas = 1. - betas + alphas_cumprod = np.cumprod(alphas, axis=0) + alphas_cumprod_prev = np.append(1., alphas_cumprod[:-1]) + + timesteps, = betas.shape + self.num_timesteps = int(timesteps) + self.linear_start = linear_start + self.linear_end = linear_end + assert alphas_cumprod.shape[0] == self.num_timesteps, 'alphas have to be defined for each timestep' + + to_torch = partial(torch.tensor, dtype=torch.float32) + + self.register_buffer('betas', to_torch(betas)) + self.register_buffer('alphas_cumprod', to_torch(alphas_cumprod)) + self.register_buffer('alphas_cumprod_prev', to_torch(alphas_cumprod_prev)) + + # calculations for diffusion q(x_t | x_{t-1}) and others + self.register_buffer('sqrt_alphas_cumprod', to_torch(np.sqrt(alphas_cumprod))) + self.register_buffer('sqrt_one_minus_alphas_cumprod', to_torch(np.sqrt(1. - alphas_cumprod))) + self.register_buffer('log_one_minus_alphas_cumprod', to_torch(np.log(1. - alphas_cumprod))) + self.register_buffer('sqrt_recip_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod))) + self.register_buffer('sqrt_recipm1_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod - 1))) + + def q_sample(self, x_start, t, noise=None): + noise = default(noise, lambda: torch.randn_like(x_start)) + return (extract_into_tensor(self.sqrt_alphas_cumprod, t, x_start.shape) * x_start + + extract_into_tensor(self.sqrt_one_minus_alphas_cumprod, t, x_start.shape) * noise) + + def forward(self, x): + return x, None + + def decode(self, x): + return x + + +class SimpleImageConcat(AbstractLowScaleModel): + # no noise level conditioning + def __init__(self): + super(SimpleImageConcat, self).__init__(noise_schedule_config=None) + self.max_noise_level = 0 + + def forward(self, x): + # fix to constant noise level + return x, torch.zeros(x.shape[0], device=x.device).long() + + +class ImageConcatWithNoiseAugmentation(AbstractLowScaleModel): + def __init__(self, noise_schedule_config, max_noise_level=1000, to_cuda=False): + super().__init__(noise_schedule_config=noise_schedule_config) + self.max_noise_level = max_noise_level + + def forward(self, x, noise_level=None): + if noise_level is None: + noise_level = torch.randint(0, self.max_noise_level, (x.shape[0],), device=x.device).long() + else: + assert isinstance(noise_level, torch.Tensor) + z = self.q_sample(x, noise_level) + return z, noise_level + + + diff --git a/ldm/modules/diffusionmodules/util.py b/ldm/modules/diffusionmodules/util.py new file mode 100644 index 0000000000000000000000000000000000000000..daf35da7bae05b9a2880dbc923b70e129e2c904d --- /dev/null +++ b/ldm/modules/diffusionmodules/util.py @@ -0,0 +1,278 @@ +# adopted from +# https://github.com/openai/improved-diffusion/blob/main/improved_diffusion/gaussian_diffusion.py +# and +# https://github.com/lucidrains/denoising-diffusion-pytorch/blob/7706bdfc6f527f58d33f84b7b522e61e6e3164b3/denoising_diffusion_pytorch/denoising_diffusion_pytorch.py +# and +# https://github.com/openai/guided-diffusion/blob/0ba878e517b276c45d1195eb29f6f5f72659a05b/guided_diffusion/nn.py +# +# thanks! + + +import os +import math +import torch +import torch.nn as nn +import numpy as np +from einops import repeat + +from ldm.util import instantiate_from_config + + +def make_beta_schedule(schedule, n_timestep, linear_start=1e-4, linear_end=2e-2, cosine_s=8e-3): + if schedule == "linear": + betas = ( + torch.linspace(linear_start ** 0.5, linear_end ** 0.5, n_timestep, dtype=torch.float64) ** 2 + ) + + elif schedule == "cosine": + timesteps = ( + torch.arange(n_timestep + 1, dtype=torch.float64) / n_timestep + cosine_s + ) + alphas = timesteps / (1 + cosine_s) * np.pi / 2 + alphas = torch.cos(alphas).pow(2) + alphas = alphas / alphas[0] + betas = 1 - alphas[1:] / alphas[:-1] + betas = np.clip(betas, a_min=0, a_max=0.999) + + elif schedule == "squaredcos_cap_v2": # used for karlo prior + # return early + return betas_for_alpha_bar( + n_timestep, + lambda t: math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2, + ) + + elif schedule == "sqrt_linear": + betas = torch.linspace(linear_start, linear_end, n_timestep, dtype=torch.float64) + elif schedule == "sqrt": + betas = torch.linspace(linear_start, linear_end, n_timestep, dtype=torch.float64) ** 0.5 + else: + raise ValueError(f"schedule '{schedule}' unknown.") + return betas.numpy() + + +def make_ddim_timesteps(ddim_discr_method, num_ddim_timesteps, num_ddpm_timesteps, verbose=True): + if ddim_discr_method == 'uniform': + c = num_ddpm_timesteps // num_ddim_timesteps + ddim_timesteps = np.asarray(list(range(0, num_ddpm_timesteps, c))) + elif ddim_discr_method == 'quad': + ddim_timesteps = ((np.linspace(0, np.sqrt(num_ddpm_timesteps * .8), num_ddim_timesteps)) ** 2).astype(int) + else: + raise NotImplementedError(f'There is no ddim discretization method called "{ddim_discr_method}"') + + # assert ddim_timesteps.shape[0] == num_ddim_timesteps + # add one to get the final alpha values right (the ones from first scale to data during sampling) + steps_out = ddim_timesteps + 1 + if verbose: + print(f'Selected timesteps for ddim sampler: {steps_out}') + return steps_out + + +def make_ddim_sampling_parameters(alphacums, ddim_timesteps, eta, verbose=True): + # select alphas for computing the variance schedule + alphas = alphacums[ddim_timesteps] + alphas_prev = np.asarray([alphacums[0]] + alphacums[ddim_timesteps[:-1]].tolist()) + + # according the the formula provided in https://arxiv.org/abs/2010.02502 + sigmas = eta * np.sqrt((1 - alphas_prev) / (1 - alphas) * (1 - alphas / alphas_prev)) + if verbose: + print(f'Selected alphas for ddim sampler: a_t: {alphas}; a_(t-1): {alphas_prev}') + print(f'For the chosen value of eta, which is {eta}, ' + f'this results in the following sigma_t schedule for ddim sampler {sigmas}') + return sigmas, alphas, alphas_prev + + +def betas_for_alpha_bar(num_diffusion_timesteps, alpha_bar, max_beta=0.999): + """ + Create a beta schedule that discretizes the given alpha_t_bar function, + which defines the cumulative product of (1-beta) over time from t = [0,1]. + :param num_diffusion_timesteps: the number of betas to produce. + :param alpha_bar: a lambda that takes an argument t from 0 to 1 and + produces the cumulative product of (1-beta) up to that + part of the diffusion process. + :param max_beta: the maximum beta to use; use values lower than 1 to + prevent singularities. + """ + betas = [] + for i in range(num_diffusion_timesteps): + t1 = i / num_diffusion_timesteps + t2 = (i + 1) / num_diffusion_timesteps + betas.append(min(1 - alpha_bar(t2) / alpha_bar(t1), max_beta)) + return np.array(betas) + + +def extract_into_tensor(a, t, x_shape): + b, *_ = t.shape + out = a.gather(-1, t) + return out.reshape(b, *((1,) * (len(x_shape) - 1))) + + +def checkpoint(func, inputs, params, flag): + """ + Evaluate a function without caching intermediate activations, allowing for + reduced memory at the expense of extra compute in the backward pass. + :param func: the function to evaluate. + :param inputs: the argument sequence to pass to `func`. + :param params: a sequence of parameters `func` depends on but does not + explicitly take as arguments. + :param flag: if False, disable gradient checkpointing. + """ + if flag: + args = tuple(inputs) + tuple(params) + return CheckpointFunction.apply(func, len(inputs), *args) + else: + return func(*inputs) + + +class CheckpointFunction(torch.autograd.Function): + @staticmethod + def forward(ctx, run_function, length, *args): + ctx.run_function = run_function + ctx.input_tensors = list(args[:length]) + ctx.input_params = list(args[length:]) + ctx.gpu_autocast_kwargs = {"enabled": torch.is_autocast_enabled(), + "dtype": torch.get_autocast_gpu_dtype(), + "cache_enabled": torch.is_autocast_cache_enabled()} + with torch.no_grad(): + output_tensors = ctx.run_function(*ctx.input_tensors) + return output_tensors + + @staticmethod + def backward(ctx, *output_grads): + ctx.input_tensors = [x.detach().requires_grad_(True) for x in ctx.input_tensors] + with torch.enable_grad(), \ + torch.cuda.amp.autocast(**ctx.gpu_autocast_kwargs): + # Fixes a bug where the first op in run_function modifies the + # Tensor storage in place, which is not allowed for detach()'d + # Tensors. + shallow_copies = [x.view_as(x) for x in ctx.input_tensors] + output_tensors = ctx.run_function(*shallow_copies) + input_grads = torch.autograd.grad( + output_tensors, + ctx.input_tensors + ctx.input_params, + output_grads, + allow_unused=True, + ) + del ctx.input_tensors + del ctx.input_params + del output_tensors + return (None, None) + input_grads + + +def timestep_embedding(timesteps, dim, max_period=10000, repeat_only=False): + """ + Create sinusoidal timestep embeddings. + :param timesteps: a 1-D Tensor of N indices, one per batch element. + These may be fractional. + :param dim: the dimension of the output. + :param max_period: controls the minimum frequency of the embeddings. + :return: an [N x dim] Tensor of positional embeddings. + """ + if not repeat_only: + half = dim // 2 + freqs = torch.exp( + -math.log(max_period) * torch.arange(start=0, end=half, dtype=torch.float32) / half + ).to(device=timesteps.device) + args = timesteps[:, None].float() * freqs[None] + embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1) + if dim % 2: + embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1) + else: + embedding = repeat(timesteps, 'b -> b d', d=dim) + return embedding + + +def zero_module(module): + """ + Zero out the parameters of a module and return it. + """ + for p in module.parameters(): + p.detach().zero_() + return module + + +def scale_module(module, scale): + """ + Scale the parameters of a module and return it. + """ + for p in module.parameters(): + p.detach().mul_(scale) + return module + + +def mean_flat(tensor): + """ + Take the mean over all non-batch dimensions. + """ + return tensor.mean(dim=list(range(1, len(tensor.shape)))) + + +def normalization(channels): + """ + Make a standard normalization layer. + :param channels: number of input channels. + :return: an nn.Module for normalization. + """ + return GroupNorm32(32, channels) + + +# PyTorch 1.7 has SiLU, but we support PyTorch 1.5. +class SiLU(nn.Module): + def forward(self, x): + return x * torch.sigmoid(x) + + +class GroupNorm32(nn.GroupNorm): + def forward(self, x): + return super().forward(x.float()).type(x.dtype) + + +def conv_nd(dims, *args, **kwargs): + """ + Create a 1D, 2D, or 3D convolution module. + """ + if dims == 1: + return nn.Conv1d(*args, **kwargs) + elif dims == 2: + return nn.Conv2d(*args, **kwargs) + elif dims == 3: + return nn.Conv3d(*args, **kwargs) + raise ValueError(f"unsupported dimensions: {dims}") + + +def linear(*args, **kwargs): + """ + Create a linear module. + """ + return nn.Linear(*args, **kwargs) + + +def avg_pool_nd(dims, *args, **kwargs): + """ + Create a 1D, 2D, or 3D average pooling module. + """ + if dims == 1: + return nn.AvgPool1d(*args, **kwargs) + elif dims == 2: + return nn.AvgPool2d(*args, **kwargs) + elif dims == 3: + return nn.AvgPool3d(*args, **kwargs) + raise ValueError(f"unsupported dimensions: {dims}") + + +class HybridConditioner(nn.Module): + + def __init__(self, c_concat_config, c_crossattn_config): + super().__init__() + self.concat_conditioner = instantiate_from_config(c_concat_config) + self.crossattn_conditioner = instantiate_from_config(c_crossattn_config) + + def forward(self, c_concat, c_crossattn): + c_concat = self.concat_conditioner(c_concat) + c_crossattn = self.crossattn_conditioner(c_crossattn) + return {'c_concat': [c_concat], 'c_crossattn': [c_crossattn]} + + +def noise_like(shape, device, repeat=False): + repeat_noise = lambda: torch.randn((1, *shape[1:]), device=device).repeat(shape[0], *((1,) * (len(shape) - 1))) + noise = lambda: torch.randn(shape, device=device) + return repeat_noise() if repeat else noise() diff --git a/ldm/modules/distributions/__init__.py b/ldm/modules/distributions/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/ldm/modules/distributions/distributions.py b/ldm/modules/distributions/distributions.py new file mode 100644 index 0000000000000000000000000000000000000000..f2b8ef901130efc171aa69742ca0244d94d3f2e9 --- /dev/null +++ b/ldm/modules/distributions/distributions.py @@ -0,0 +1,92 @@ +import torch +import numpy as np + + +class AbstractDistribution: + def sample(self): + raise NotImplementedError() + + def mode(self): + raise NotImplementedError() + + +class DiracDistribution(AbstractDistribution): + def __init__(self, value): + self.value = value + + def sample(self): + return self.value + + def mode(self): + return self.value + + +class DiagonalGaussianDistribution(object): + def __init__(self, parameters, deterministic=False): + self.parameters = parameters + self.mean, self.logvar = torch.chunk(parameters, 2, dim=1) + self.logvar = torch.clamp(self.logvar, -30.0, 20.0) + self.deterministic = deterministic + self.std = torch.exp(0.5 * self.logvar) + self.var = torch.exp(self.logvar) + if self.deterministic: + self.var = self.std = torch.zeros_like(self.mean).to(device=self.parameters.device) + + def sample(self): + x = self.mean + self.std * torch.randn(self.mean.shape).to(device=self.parameters.device) + return x + + def kl(self, other=None): + if self.deterministic: + return torch.Tensor([0.]) + else: + if other is None: + return 0.5 * torch.sum(torch.pow(self.mean, 2) + + self.var - 1.0 - self.logvar, + dim=[1, 2, 3]) + else: + return 0.5 * torch.sum( + torch.pow(self.mean - other.mean, 2) / other.var + + self.var / other.var - 1.0 - self.logvar + other.logvar, + dim=[1, 2, 3]) + + def nll(self, sample, dims=[1,2,3]): + if self.deterministic: + return torch.Tensor([0.]) + logtwopi = np.log(2.0 * np.pi) + return 0.5 * torch.sum( + logtwopi + self.logvar + torch.pow(sample - self.mean, 2) / self.var, + dim=dims) + + def mode(self): + return self.mean + + +def normal_kl(mean1, logvar1, mean2, logvar2): + """ + source: https://github.com/openai/guided-diffusion/blob/27c20a8fab9cb472df5d6bdd6c8d11c8f430b924/guided_diffusion/losses.py#L12 + Compute the KL divergence between two gaussians. + Shapes are automatically broadcasted, so batches can be compared to + scalars, among other use cases. + """ + tensor = None + for obj in (mean1, logvar1, mean2, logvar2): + if isinstance(obj, torch.Tensor): + tensor = obj + break + assert tensor is not None, "at least one argument must be a Tensor" + + # Force variances to be Tensors. Broadcasting helps convert scalars to + # Tensors, but it does not work for torch.exp(). + logvar1, logvar2 = [ + x if isinstance(x, torch.Tensor) else torch.tensor(x).to(tensor) + for x in (logvar1, logvar2) + ] + + return 0.5 * ( + -1.0 + + logvar2 + - logvar1 + + torch.exp(logvar1 - logvar2) + + ((mean1 - mean2) ** 2) * torch.exp(-logvar2) + ) diff --git a/ldm/modules/ema.py b/ldm/modules/ema.py new file mode 100644 index 0000000000000000000000000000000000000000..bded25019b9bcbcd0260f0b8185f8c7859ca58c4 --- /dev/null +++ b/ldm/modules/ema.py @@ -0,0 +1,80 @@ +import torch +from torch import nn + + +class LitEma(nn.Module): + def __init__(self, model, decay=0.9999, use_num_upates=True): + super().__init__() + if decay < 0.0 or decay > 1.0: + raise ValueError('Decay must be between 0 and 1') + + self.m_name2s_name = {} + self.register_buffer('decay', torch.tensor(decay, dtype=torch.float32)) + self.register_buffer('num_updates', torch.tensor(0, dtype=torch.int) if use_num_upates + else torch.tensor(-1, dtype=torch.int)) + + for name, p in model.named_parameters(): + if p.requires_grad: + # remove as '.'-character is not allowed in buffers + s_name = name.replace('.', '') + self.m_name2s_name.update({name: s_name}) + self.register_buffer(s_name, p.clone().detach().data) + + self.collected_params = [] + + def reset_num_updates(self): + del self.num_updates + self.register_buffer('num_updates', torch.tensor(0, dtype=torch.int)) + + def forward(self, model): + decay = self.decay + + if self.num_updates >= 0: + self.num_updates += 1 + decay = min(self.decay, (1 + self.num_updates) / (10 + self.num_updates)) + + one_minus_decay = 1.0 - decay + + with torch.no_grad(): + m_param = dict(model.named_parameters()) + shadow_params = dict(self.named_buffers()) + + for key in m_param: + if m_param[key].requires_grad: + sname = self.m_name2s_name[key] + shadow_params[sname] = shadow_params[sname].type_as(m_param[key]) + shadow_params[sname].sub_(one_minus_decay * (shadow_params[sname] - m_param[key])) + else: + assert not key in self.m_name2s_name + + def copy_to(self, model): + m_param = dict(model.named_parameters()) + shadow_params = dict(self.named_buffers()) + for key in m_param: + if m_param[key].requires_grad: + m_param[key].data.copy_(shadow_params[self.m_name2s_name[key]].data) + else: + assert not key in self.m_name2s_name + + def store(self, parameters): + """ + Save the current parameters for restoring later. + Args: + parameters: Iterable of `torch.nn.Parameter`; the parameters to be + temporarily stored. + """ + self.collected_params = [param.clone() for param in parameters] + + def restore(self, parameters): + """ + Restore the parameters stored with the `store` method. + Useful to validate the model with EMA parameters without affecting the + original optimization process. Store the parameters before the + `copy_to` method. After validation (or model saving), use this to + restore the former parameters. + Args: + parameters: Iterable of `torch.nn.Parameter`; the parameters to be + updated with the stored parameters. + """ + for c_param, param in zip(self.collected_params, parameters): + param.data.copy_(c_param.data) diff --git a/ldm/modules/encoders/__init__.py b/ldm/modules/encoders/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/ldm/modules/encoders/modules.py b/ldm/modules/encoders/modules.py new file mode 100644 index 0000000000000000000000000000000000000000..523a7d8535b7a1741bc3d8bbadf59ac63b0de8ab --- /dev/null +++ b/ldm/modules/encoders/modules.py @@ -0,0 +1,350 @@ +import torch +import torch.nn as nn +import kornia +from torch.utils.checkpoint import checkpoint + +from transformers import T5Tokenizer, T5EncoderModel, CLIPTokenizer, CLIPTextModel + +import open_clip +from ldm.util import default, count_params, autocast + + +class AbstractEncoder(nn.Module): + def __init__(self): + super().__init__() + + def encode(self, *args, **kwargs): + raise NotImplementedError + + +class IdentityEncoder(AbstractEncoder): + + def encode(self, x): + return x + + +class ClassEmbedder(nn.Module): + def __init__(self, embed_dim, n_classes=1000, key='class', ucg_rate=0.1): + super().__init__() + self.key = key + self.embedding = nn.Embedding(n_classes, embed_dim) + self.n_classes = n_classes + self.ucg_rate = ucg_rate + + def forward(self, batch, key=None, disable_dropout=False): + if key is None: + key = self.key + # this is for use in crossattn + c = batch[key][:, None] + if self.ucg_rate > 0. and not disable_dropout: + mask = 1. - torch.bernoulli(torch.ones_like(c) * self.ucg_rate) + c = mask * c + (1 - mask) * torch.ones_like(c) * (self.n_classes - 1) + c = c.long() + c = self.embedding(c) + return c + + def get_unconditional_conditioning(self, bs, device="cuda"): + uc_class = self.n_classes - 1 # 1000 classes --> 0 ... 999, one extra class for ucg (class 1000) + uc = torch.ones((bs,), device=device) * uc_class + uc = {self.key: uc} + return uc + + +def disabled_train(self, mode=True): + """Overwrite model.train with this function to make sure train/eval mode + does not change anymore.""" + return self + + +class FrozenT5Embedder(AbstractEncoder): + """Uses the T5 transformer encoder for text""" + + def __init__(self, version="google/t5-v1_1-large", device="cuda", max_length=77, + freeze=True): # others are google/t5-v1_1-xl and google/t5-v1_1-xxl + super().__init__() + self.tokenizer = T5Tokenizer.from_pretrained(version) + self.transformer = T5EncoderModel.from_pretrained(version) + self.device = device + self.max_length = max_length # TODO: typical value? + if freeze: + self.freeze() + + def freeze(self): + self.transformer = self.transformer.eval() + # self.train = disabled_train + for param in self.parameters(): + param.requires_grad = False + + def forward(self, text): + batch_encoding = self.tokenizer(text, truncation=True, max_length=self.max_length, return_length=True, + return_overflowing_tokens=False, padding="max_length", return_tensors="pt") + tokens = batch_encoding["input_ids"].to(self.device) + outputs = self.transformer(input_ids=tokens) + + z = outputs.last_hidden_state + return z + + def encode(self, text): + return self(text) + + +class FrozenCLIPEmbedder(AbstractEncoder): + """Uses the CLIP transformer encoder for text (from huggingface)""" + LAYERS = [ + "last", + "pooled", + "hidden" + ] + + def __init__(self, version="openai/clip-vit-large-patch14", device="cuda", max_length=77, + freeze=True, layer="last", layer_idx=None): # clip-vit-base-patch32 + super().__init__() + assert layer in self.LAYERS + self.tokenizer = CLIPTokenizer.from_pretrained(version) + self.transformer = CLIPTextModel.from_pretrained(version) + self.device = device + self.max_length = max_length + if freeze: + self.freeze() + self.layer = layer + self.layer_idx = layer_idx + if layer == "hidden": + assert layer_idx is not None + assert 0 <= abs(layer_idx) <= 12 + + def freeze(self): + self.transformer = self.transformer.eval() + # self.train = disabled_train + for param in self.parameters(): + param.requires_grad = False + + def forward(self, text): + batch_encoding = self.tokenizer(text, truncation=True, max_length=self.max_length, return_length=True, + return_overflowing_tokens=False, padding="max_length", return_tensors="pt") + tokens = batch_encoding["input_ids"].to(self.device) + outputs = self.transformer(input_ids=tokens, output_hidden_states=self.layer == "hidden") + if self.layer == "last": + z = outputs.last_hidden_state + elif self.layer == "pooled": + z = outputs.pooler_output[:, None, :] + else: + z = outputs.hidden_states[self.layer_idx] + return z + + def encode(self, text): + return self(text) + + +class ClipImageEmbedder(nn.Module): + def __init__( + self, + model, + jit=False, + device='cuda' if torch.cuda.is_available() else 'cpu', + antialias=True, + ucg_rate=0. + ): + super().__init__() + from clip import load as load_clip + self.model, _ = load_clip(name=model, device=device, jit=jit) + + self.antialias = antialias + + self.register_buffer('mean', torch.Tensor([0.48145466, 0.4578275, 0.40821073]), persistent=False) + self.register_buffer('std', torch.Tensor([0.26862954, 0.26130258, 0.27577711]), persistent=False) + self.ucg_rate = ucg_rate + + def preprocess(self, x): + # normalize to [0,1] + x = kornia.geometry.resize(x, (224, 224), + interpolation='bicubic', align_corners=True, + antialias=self.antialias) + x = (x + 1.) / 2. + # re-normalize according to clip + x = kornia.enhance.normalize(x, self.mean, self.std) + return x + + def forward(self, x, no_dropout=False): + # x is assumed to be in range [-1,1] + out = self.model.encode_image(self.preprocess(x)) + out = out.to(x.dtype) + if self.ucg_rate > 0. and not no_dropout: + out = torch.bernoulli((1. - self.ucg_rate) * torch.ones(out.shape[0], device=out.device))[:, None] * out + return out + + +class FrozenOpenCLIPEmbedder(AbstractEncoder): + """ + Uses the OpenCLIP transformer encoder for text + """ + LAYERS = [ + # "pooled", + "last", + "penultimate" + ] + + def __init__(self, arch="ViT-H-14", version="laion2b_s32b_b79k", device="cuda", max_length=77, + freeze=True, layer="last"): + super().__init__() + assert layer in self.LAYERS + model, _, _ = open_clip.create_model_and_transforms(arch, device=torch.device('cpu'), pretrained=version) + del model.visual + self.model = model + + self.device = device + self.max_length = max_length + if freeze: + self.freeze() + self.layer = layer + if self.layer == "last": + self.layer_idx = 0 + elif self.layer == "penultimate": + self.layer_idx = 1 + else: + raise NotImplementedError() + + def freeze(self): + self.model = self.model.eval() + for param in self.parameters(): + param.requires_grad = False + + def forward(self, text): + tokens = open_clip.tokenize(text) + z = self.encode_with_transformer(tokens.to(self.device)) + return z + + def encode_with_transformer(self, text): + x = self.model.token_embedding(text) # [batch_size, n_ctx, d_model] + x = x + self.model.positional_embedding + x = x.permute(1, 0, 2) # NLD -> LND + x = self.text_transformer_forward(x, attn_mask=self.model.attn_mask) + x = x.permute(1, 0, 2) # LND -> NLD + x = self.model.ln_final(x) + return x + + def text_transformer_forward(self, x: torch.Tensor, attn_mask=None): + for i, r in enumerate(self.model.transformer.resblocks): + if i == len(self.model.transformer.resblocks) - self.layer_idx: + break + if self.model.transformer.grad_checkpointing and not torch.jit.is_scripting(): + x = checkpoint(r, x, attn_mask) + else: + x = r(x, attn_mask=attn_mask) + return x + + def encode(self, text): + return self(text) + + +class FrozenOpenCLIPImageEmbedder(AbstractEncoder): + """ + Uses the OpenCLIP vision transformer encoder for images + """ + + def __init__(self, arch="ViT-H-14", version="laion2b_s32b_b79k", device="cuda", max_length=77, + freeze=True, layer="pooled", antialias=True, ucg_rate=0.): + super().__init__() + model, _, _ = open_clip.create_model_and_transforms(arch, device=torch.device('cpu'), + pretrained=version, ) + del model.transformer + self.model = model + + self.device = device + self.max_length = max_length + if freeze: + self.freeze() + self.layer = layer + if self.layer == "penultimate": + raise NotImplementedError() + self.layer_idx = 1 + + self.antialias = antialias + + self.register_buffer('mean', torch.Tensor([0.48145466, 0.4578275, 0.40821073]), persistent=False) + self.register_buffer('std', torch.Tensor([0.26862954, 0.26130258, 0.27577711]), persistent=False) + self.ucg_rate = ucg_rate + + def preprocess(self, x): + # normalize to [0,1] + x = kornia.geometry.resize(x, (224, 224), + interpolation='bicubic', align_corners=True, + antialias=self.antialias) + x = (x + 1.) / 2. + # renormalize according to clip + x = kornia.enhance.normalize(x, self.mean, self.std) + return x + + def freeze(self): + self.model = self.model.eval() + for param in self.parameters(): + param.requires_grad = False + + @autocast + def forward(self, image, no_dropout=False): + z = self.encode_with_vision_transformer(image) + if self.ucg_rate > 0. and not no_dropout: + z = torch.bernoulli((1. - self.ucg_rate) * torch.ones(z.shape[0], device=z.device))[:, None] * z + return z + + def encode_with_vision_transformer(self, img): + img = self.preprocess(img) + x = self.model.visual(img) + return x + + def encode(self, text): + return self(text) + + +class FrozenCLIPT5Encoder(AbstractEncoder): + def __init__(self, clip_version="openai/clip-vit-large-patch14", t5_version="google/t5-v1_1-xl", device="cuda", + clip_max_length=77, t5_max_length=77): + super().__init__() + self.clip_encoder = FrozenCLIPEmbedder(clip_version, device, max_length=clip_max_length) + self.t5_encoder = FrozenT5Embedder(t5_version, device, max_length=t5_max_length) + print(f"{self.clip_encoder.__class__.__name__} has {count_params(self.clip_encoder) * 1.e-6:.2f} M parameters, " + f"{self.t5_encoder.__class__.__name__} comes with {count_params(self.t5_encoder) * 1.e-6:.2f} M params.") + + def encode(self, text): + return self(text) + + def forward(self, text): + clip_z = self.clip_encoder.encode(text) + t5_z = self.t5_encoder.encode(text) + return [clip_z, t5_z] + + +from ldm.modules.diffusionmodules.upscaling import ImageConcatWithNoiseAugmentation +from ldm.modules.diffusionmodules.openaimodel import Timestep + + +class CLIPEmbeddingNoiseAugmentation(ImageConcatWithNoiseAugmentation): + def __init__(self, *args, clip_stats_path=None, timestep_dim=256, **kwargs): + super().__init__(*args, **kwargs) + if clip_stats_path is None: + clip_mean, clip_std = torch.zeros(timestep_dim), torch.ones(timestep_dim) + else: + clip_mean, clip_std = torch.load(clip_stats_path, map_location="cpu") + self.register_buffer("data_mean", clip_mean[None, :], persistent=False) + self.register_buffer("data_std", clip_std[None, :], persistent=False) + self.time_embed = Timestep(timestep_dim) + + def scale(self, x): + # re-normalize to centered mean and unit variance + x = (x - self.data_mean) * 1. / self.data_std + return x + + def unscale(self, x): + # back to original data stats + x = (x * self.data_std) + self.data_mean + return x + + def forward(self, x, noise_level=None): + if noise_level is None: + noise_level = torch.randint(0, self.max_noise_level, (x.shape[0],), device=x.device).long() + else: + assert isinstance(noise_level, torch.Tensor) + x = self.scale(x) + z = self.q_sample(x, noise_level) + z = self.unscale(z) + noise_level = self.time_embed(noise_level) + return z, noise_level diff --git a/ldm/modules/image_degradation/__init__.py b/ldm/modules/image_degradation/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..7836cada81f90ded99c58d5942eea4c3477f58fc --- /dev/null +++ b/ldm/modules/image_degradation/__init__.py @@ -0,0 +1,2 @@ +from ldm.modules.image_degradation.bsrgan import degradation_bsrgan_variant as degradation_fn_bsr +from ldm.modules.image_degradation.bsrgan_light import degradation_bsrgan_variant as degradation_fn_bsr_light diff --git a/ldm/modules/image_degradation/bsrgan.py b/ldm/modules/image_degradation/bsrgan.py new file mode 100644 index 0000000000000000000000000000000000000000..32ef56169978e550090261cddbcf5eb611a6173b --- /dev/null +++ b/ldm/modules/image_degradation/bsrgan.py @@ -0,0 +1,730 @@ +# -*- coding: utf-8 -*- +""" +# -------------------------------------------- +# Super-Resolution +# -------------------------------------------- +# +# Kai Zhang (cskaizhang@gmail.com) +# https://github.com/cszn +# From 2019/03--2021/08 +# -------------------------------------------- +""" + +import numpy as np +import cv2 +import torch + +from functools import partial +import random +from scipy import ndimage +import scipy +import scipy.stats as ss +from scipy.interpolate import interp2d +from scipy.linalg import orth +import albumentations + +import ldm.modules.image_degradation.utils_image as util + + +def modcrop_np(img, sf): + ''' + Args: + img: numpy image, WxH or WxHxC + sf: scale factor + Return: + cropped image + ''' + w, h = img.shape[:2] + im = np.copy(img) + return im[:w - w % sf, :h - h % sf, ...] + + +""" +# -------------------------------------------- +# anisotropic Gaussian kernels +# -------------------------------------------- +""" + + +def analytic_kernel(k): + """Calculate the X4 kernel from the X2 kernel (for proof see appendix in paper)""" + k_size = k.shape[0] + # Calculate the big kernels size + big_k = np.zeros((3 * k_size - 2, 3 * k_size - 2)) + # Loop over the small kernel to fill the big one + for r in range(k_size): + for c in range(k_size): + big_k[2 * r:2 * r + k_size, 2 * c:2 * c + k_size] += k[r, c] * k + # Crop the edges of the big kernel to ignore very small values and increase run time of SR + crop = k_size // 2 + cropped_big_k = big_k[crop:-crop, crop:-crop] + # Normalize to 1 + return cropped_big_k / cropped_big_k.sum() + + +def anisotropic_Gaussian(ksize=15, theta=np.pi, l1=6, l2=6): + """ generate an anisotropic Gaussian kernel + Args: + ksize : e.g., 15, kernel size + theta : [0, pi], rotation angle range + l1 : [0.1,50], scaling of eigenvalues + l2 : [0.1,l1], scaling of eigenvalues + If l1 = l2, will get an isotropic Gaussian kernel. + Returns: + k : kernel + """ + + v = np.dot(np.array([[np.cos(theta), -np.sin(theta)], [np.sin(theta), np.cos(theta)]]), np.array([1., 0.])) + V = np.array([[v[0], v[1]], [v[1], -v[0]]]) + D = np.array([[l1, 0], [0, l2]]) + Sigma = np.dot(np.dot(V, D), np.linalg.inv(V)) + k = gm_blur_kernel(mean=[0, 0], cov=Sigma, size=ksize) + + return k + + +def gm_blur_kernel(mean, cov, size=15): + center = size / 2.0 + 0.5 + k = np.zeros([size, size]) + for y in range(size): + for x in range(size): + cy = y - center + 1 + cx = x - center + 1 + k[y, x] = ss.multivariate_normal.pdf([cx, cy], mean=mean, cov=cov) + + k = k / np.sum(k) + return k + + +def shift_pixel(x, sf, upper_left=True): + """shift pixel for super-resolution with different scale factors + Args: + x: WxHxC or WxH + sf: scale factor + upper_left: shift direction + """ + h, w = x.shape[:2] + shift = (sf - 1) * 0.5 + xv, yv = np.arange(0, w, 1.0), np.arange(0, h, 1.0) + if upper_left: + x1 = xv + shift + y1 = yv + shift + else: + x1 = xv - shift + y1 = yv - shift + + x1 = np.clip(x1, 0, w - 1) + y1 = np.clip(y1, 0, h - 1) + + if x.ndim == 2: + x = interp2d(xv, yv, x)(x1, y1) + if x.ndim == 3: + for i in range(x.shape[-1]): + x[:, :, i] = interp2d(xv, yv, x[:, :, i])(x1, y1) + + return x + + +def blur(x, k): + ''' + x: image, NxcxHxW + k: kernel, Nx1xhxw + ''' + n, c = x.shape[:2] + p1, p2 = (k.shape[-2] - 1) // 2, (k.shape[-1] - 1) // 2 + x = torch.nn.functional.pad(x, pad=(p1, p2, p1, p2), mode='replicate') + k = k.repeat(1, c, 1, 1) + k = k.view(-1, 1, k.shape[2], k.shape[3]) + x = x.view(1, -1, x.shape[2], x.shape[3]) + x = torch.nn.functional.conv2d(x, k, bias=None, stride=1, padding=0, groups=n * c) + x = x.view(n, c, x.shape[2], x.shape[3]) + + return x + + +def gen_kernel(k_size=np.array([15, 15]), scale_factor=np.array([4, 4]), min_var=0.6, max_var=10., noise_level=0): + """" + # modified version of https://github.com/assafshocher/BlindSR_dataset_generator + # Kai Zhang + # min_var = 0.175 * sf # variance of the gaussian kernel will be sampled between min_var and max_var + # max_var = 2.5 * sf + """ + # Set random eigen-vals (lambdas) and angle (theta) for COV matrix + lambda_1 = min_var + np.random.rand() * (max_var - min_var) + lambda_2 = min_var + np.random.rand() * (max_var - min_var) + theta = np.random.rand() * np.pi # random theta + noise = -noise_level + np.random.rand(*k_size) * noise_level * 2 + + # Set COV matrix using Lambdas and Theta + LAMBDA = np.diag([lambda_1, lambda_2]) + Q = np.array([[np.cos(theta), -np.sin(theta)], + [np.sin(theta), np.cos(theta)]]) + SIGMA = Q @ LAMBDA @ Q.T + INV_SIGMA = np.linalg.inv(SIGMA)[None, None, :, :] + + # Set expectation position (shifting kernel for aligned image) + MU = k_size // 2 - 0.5 * (scale_factor - 1) # - 0.5 * (scale_factor - k_size % 2) + MU = MU[None, None, :, None] + + # Create meshgrid for Gaussian + [X, Y] = np.meshgrid(range(k_size[0]), range(k_size[1])) + Z = np.stack([X, Y], 2)[:, :, :, None] + + # Calcualte Gaussian for every pixel of the kernel + ZZ = Z - MU + ZZ_t = ZZ.transpose(0, 1, 3, 2) + raw_kernel = np.exp(-0.5 * np.squeeze(ZZ_t @ INV_SIGMA @ ZZ)) * (1 + noise) + + # shift the kernel so it will be centered + # raw_kernel_centered = kernel_shift(raw_kernel, scale_factor) + + # Normalize the kernel and return + # kernel = raw_kernel_centered / np.sum(raw_kernel_centered) + kernel = raw_kernel / np.sum(raw_kernel) + return kernel + + +def fspecial_gaussian(hsize, sigma): + hsize = [hsize, hsize] + siz = [(hsize[0] - 1.0) / 2.0, (hsize[1] - 1.0) / 2.0] + std = sigma + [x, y] = np.meshgrid(np.arange(-siz[1], siz[1] + 1), np.arange(-siz[0], siz[0] + 1)) + arg = -(x * x + y * y) / (2 * std * std) + h = np.exp(arg) + h[h < scipy.finfo(float).eps * h.max()] = 0 + sumh = h.sum() + if sumh != 0: + h = h / sumh + return h + + +def fspecial_laplacian(alpha): + alpha = max([0, min([alpha, 1])]) + h1 = alpha / (alpha + 1) + h2 = (1 - alpha) / (alpha + 1) + h = [[h1, h2, h1], [h2, -4 / (alpha + 1), h2], [h1, h2, h1]] + h = np.array(h) + return h + + +def fspecial(filter_type, *args, **kwargs): + ''' + python code from: + https://github.com/ronaldosena/imagens-medicas-2/blob/40171a6c259edec7827a6693a93955de2bd39e76/Aulas/aula_2_-_uniform_filter/matlab_fspecial.py + ''' + if filter_type == 'gaussian': + return fspecial_gaussian(*args, **kwargs) + if filter_type == 'laplacian': + return fspecial_laplacian(*args, **kwargs) + + +""" +# -------------------------------------------- +# degradation models +# -------------------------------------------- +""" + + +def bicubic_degradation(x, sf=3): + ''' + Args: + x: HxWxC image, [0, 1] + sf: down-scale factor + Return: + bicubicly downsampled LR image + ''' + x = util.imresize_np(x, scale=1 / sf) + return x + + +def srmd_degradation(x, k, sf=3): + ''' blur + bicubic downsampling + Args: + x: HxWxC image, [0, 1] + k: hxw, double + sf: down-scale factor + Return: + downsampled LR image + Reference: + @inproceedings{zhang2018learning, + title={Learning a single convolutional super-resolution network for multiple degradations}, + author={Zhang, Kai and Zuo, Wangmeng and Zhang, Lei}, + booktitle={IEEE Conference on Computer Vision and Pattern Recognition}, + pages={3262--3271}, + year={2018} + } + ''' + x = ndimage.filters.convolve(x, np.expand_dims(k, axis=2), mode='wrap') # 'nearest' | 'mirror' + x = bicubic_degradation(x, sf=sf) + return x + + +def dpsr_degradation(x, k, sf=3): + ''' bicubic downsampling + blur + Args: + x: HxWxC image, [0, 1] + k: hxw, double + sf: down-scale factor + Return: + downsampled LR image + Reference: + @inproceedings{zhang2019deep, + title={Deep Plug-and-Play Super-Resolution for Arbitrary Blur Kernels}, + author={Zhang, Kai and Zuo, Wangmeng and Zhang, Lei}, + booktitle={IEEE Conference on Computer Vision and Pattern Recognition}, + pages={1671--1681}, + year={2019} + } + ''' + x = bicubic_degradation(x, sf=sf) + x = ndimage.filters.convolve(x, np.expand_dims(k, axis=2), mode='wrap') + return x + + +def classical_degradation(x, k, sf=3): + ''' blur + downsampling + Args: + x: HxWxC image, [0, 1]/[0, 255] + k: hxw, double + sf: down-scale factor + Return: + downsampled LR image + ''' + x = ndimage.filters.convolve(x, np.expand_dims(k, axis=2), mode='wrap') + # x = filters.correlate(x, np.expand_dims(np.flip(k), axis=2)) + st = 0 + return x[st::sf, st::sf, ...] + + +def add_sharpening(img, weight=0.5, radius=50, threshold=10): + """USM sharpening. borrowed from real-ESRGAN + Input image: I; Blurry image: B. + 1. K = I + weight * (I - B) + 2. Mask = 1 if abs(I - B) > threshold, else: 0 + 3. Blur mask: + 4. Out = Mask * K + (1 - Mask) * I + Args: + img (Numpy array): Input image, HWC, BGR; float32, [0, 1]. + weight (float): Sharp weight. Default: 1. + radius (float): Kernel size of Gaussian blur. Default: 50. + threshold (int): + """ + if radius % 2 == 0: + radius += 1 + blur = cv2.GaussianBlur(img, (radius, radius), 0) + residual = img - blur + mask = np.abs(residual) * 255 > threshold + mask = mask.astype('float32') + soft_mask = cv2.GaussianBlur(mask, (radius, radius), 0) + + K = img + weight * residual + K = np.clip(K, 0, 1) + return soft_mask * K + (1 - soft_mask) * img + + +def add_blur(img, sf=4): + wd2 = 4.0 + sf + wd = 2.0 + 0.2 * sf + if random.random() < 0.5: + l1 = wd2 * random.random() + l2 = wd2 * random.random() + k = anisotropic_Gaussian(ksize=2 * random.randint(2, 11) + 3, theta=random.random() * np.pi, l1=l1, l2=l2) + else: + k = fspecial('gaussian', 2 * random.randint(2, 11) + 3, wd * random.random()) + img = ndimage.filters.convolve(img, np.expand_dims(k, axis=2), mode='mirror') + + return img + + +def add_resize(img, sf=4): + rnum = np.random.rand() + if rnum > 0.8: # up + sf1 = random.uniform(1, 2) + elif rnum < 0.7: # down + sf1 = random.uniform(0.5 / sf, 1) + else: + sf1 = 1.0 + img = cv2.resize(img, (int(sf1 * img.shape[1]), int(sf1 * img.shape[0])), interpolation=random.choice([1, 2, 3])) + img = np.clip(img, 0.0, 1.0) + + return img + + +# def add_Gaussian_noise(img, noise_level1=2, noise_level2=25): +# noise_level = random.randint(noise_level1, noise_level2) +# rnum = np.random.rand() +# if rnum > 0.6: # add color Gaussian noise +# img += np.random.normal(0, noise_level / 255.0, img.shape).astype(np.float32) +# elif rnum < 0.4: # add grayscale Gaussian noise +# img += np.random.normal(0, noise_level / 255.0, (*img.shape[:2], 1)).astype(np.float32) +# else: # add noise +# L = noise_level2 / 255. +# D = np.diag(np.random.rand(3)) +# U = orth(np.random.rand(3, 3)) +# conv = np.dot(np.dot(np.transpose(U), D), U) +# img += np.random.multivariate_normal([0, 0, 0], np.abs(L ** 2 * conv), img.shape[:2]).astype(np.float32) +# img = np.clip(img, 0.0, 1.0) +# return img + +def add_Gaussian_noise(img, noise_level1=2, noise_level2=25): + noise_level = random.randint(noise_level1, noise_level2) + rnum = np.random.rand() + if rnum > 0.6: # add color Gaussian noise + img = img + np.random.normal(0, noise_level / 255.0, img.shape).astype(np.float32) + elif rnum < 0.4: # add grayscale Gaussian noise + img = img + np.random.normal(0, noise_level / 255.0, (*img.shape[:2], 1)).astype(np.float32) + else: # add noise + L = noise_level2 / 255. + D = np.diag(np.random.rand(3)) + U = orth(np.random.rand(3, 3)) + conv = np.dot(np.dot(np.transpose(U), D), U) + img = img + np.random.multivariate_normal([0, 0, 0], np.abs(L ** 2 * conv), img.shape[:2]).astype(np.float32) + img = np.clip(img, 0.0, 1.0) + return img + + +def add_speckle_noise(img, noise_level1=2, noise_level2=25): + noise_level = random.randint(noise_level1, noise_level2) + img = np.clip(img, 0.0, 1.0) + rnum = random.random() + if rnum > 0.6: + img += img * np.random.normal(0, noise_level / 255.0, img.shape).astype(np.float32) + elif rnum < 0.4: + img += img * np.random.normal(0, noise_level / 255.0, (*img.shape[:2], 1)).astype(np.float32) + else: + L = noise_level2 / 255. + D = np.diag(np.random.rand(3)) + U = orth(np.random.rand(3, 3)) + conv = np.dot(np.dot(np.transpose(U), D), U) + img += img * np.random.multivariate_normal([0, 0, 0], np.abs(L ** 2 * conv), img.shape[:2]).astype(np.float32) + img = np.clip(img, 0.0, 1.0) + return img + + +def add_Poisson_noise(img): + img = np.clip((img * 255.0).round(), 0, 255) / 255. + vals = 10 ** (2 * random.random() + 2.0) # [2, 4] + if random.random() < 0.5: + img = np.random.poisson(img * vals).astype(np.float32) / vals + else: + img_gray = np.dot(img[..., :3], [0.299, 0.587, 0.114]) + img_gray = np.clip((img_gray * 255.0).round(), 0, 255) / 255. + noise_gray = np.random.poisson(img_gray * vals).astype(np.float32) / vals - img_gray + img += noise_gray[:, :, np.newaxis] + img = np.clip(img, 0.0, 1.0) + return img + + +def add_JPEG_noise(img): + quality_factor = random.randint(30, 95) + img = cv2.cvtColor(util.single2uint(img), cv2.COLOR_RGB2BGR) + result, encimg = cv2.imencode('.jpg', img, [int(cv2.IMWRITE_JPEG_QUALITY), quality_factor]) + img = cv2.imdecode(encimg, 1) + img = cv2.cvtColor(util.uint2single(img), cv2.COLOR_BGR2RGB) + return img + + +def random_crop(lq, hq, sf=4, lq_patchsize=64): + h, w = lq.shape[:2] + rnd_h = random.randint(0, h - lq_patchsize) + rnd_w = random.randint(0, w - lq_patchsize) + lq = lq[rnd_h:rnd_h + lq_patchsize, rnd_w:rnd_w + lq_patchsize, :] + + rnd_h_H, rnd_w_H = int(rnd_h * sf), int(rnd_w * sf) + hq = hq[rnd_h_H:rnd_h_H + lq_patchsize * sf, rnd_w_H:rnd_w_H + lq_patchsize * sf, :] + return lq, hq + + +def degradation_bsrgan(img, sf=4, lq_patchsize=72, isp_model=None): + """ + This is the degradation model of BSRGAN from the paper + "Designing a Practical Degradation Model for Deep Blind Image Super-Resolution" + ---------- + img: HXWXC, [0, 1], its size should be large than (lq_patchsizexsf)x(lq_patchsizexsf) + sf: scale factor + isp_model: camera ISP model + Returns + ------- + img: low-quality patch, size: lq_patchsizeXlq_patchsizeXC, range: [0, 1] + hq: corresponding high-quality patch, size: (lq_patchsizexsf)X(lq_patchsizexsf)XC, range: [0, 1] + """ + isp_prob, jpeg_prob, scale2_prob = 0.25, 0.9, 0.25 + sf_ori = sf + + h1, w1 = img.shape[:2] + img = img.copy()[:w1 - w1 % sf, :h1 - h1 % sf, ...] # mod crop + h, w = img.shape[:2] + + if h < lq_patchsize * sf or w < lq_patchsize * sf: + raise ValueError(f'img size ({h1}X{w1}) is too small!') + + hq = img.copy() + + if sf == 4 and random.random() < scale2_prob: # downsample1 + if np.random.rand() < 0.5: + img = cv2.resize(img, (int(1 / 2 * img.shape[1]), int(1 / 2 * img.shape[0])), + interpolation=random.choice([1, 2, 3])) + else: + img = util.imresize_np(img, 1 / 2, True) + img = np.clip(img, 0.0, 1.0) + sf = 2 + + shuffle_order = random.sample(range(7), 7) + idx1, idx2 = shuffle_order.index(2), shuffle_order.index(3) + if idx1 > idx2: # keep downsample3 last + shuffle_order[idx1], shuffle_order[idx2] = shuffle_order[idx2], shuffle_order[idx1] + + for i in shuffle_order: + + if i == 0: + img = add_blur(img, sf=sf) + + elif i == 1: + img = add_blur(img, sf=sf) + + elif i == 2: + a, b = img.shape[1], img.shape[0] + # downsample2 + if random.random() < 0.75: + sf1 = random.uniform(1, 2 * sf) + img = cv2.resize(img, (int(1 / sf1 * img.shape[1]), int(1 / sf1 * img.shape[0])), + interpolation=random.choice([1, 2, 3])) + else: + k = fspecial('gaussian', 25, random.uniform(0.1, 0.6 * sf)) + k_shifted = shift_pixel(k, sf) + k_shifted = k_shifted / k_shifted.sum() # blur with shifted kernel + img = ndimage.filters.convolve(img, np.expand_dims(k_shifted, axis=2), mode='mirror') + img = img[0::sf, 0::sf, ...] # nearest downsampling + img = np.clip(img, 0.0, 1.0) + + elif i == 3: + # downsample3 + img = cv2.resize(img, (int(1 / sf * a), int(1 / sf * b)), interpolation=random.choice([1, 2, 3])) + img = np.clip(img, 0.0, 1.0) + + elif i == 4: + # add Gaussian noise + img = add_Gaussian_noise(img, noise_level1=2, noise_level2=25) + + elif i == 5: + # add JPEG noise + if random.random() < jpeg_prob: + img = add_JPEG_noise(img) + + elif i == 6: + # add processed camera sensor noise + if random.random() < isp_prob and isp_model is not None: + with torch.no_grad(): + img, hq = isp_model.forward(img.copy(), hq) + + # add final JPEG compression noise + img = add_JPEG_noise(img) + + # random crop + img, hq = random_crop(img, hq, sf_ori, lq_patchsize) + + return img, hq + + +# todo no isp_model? +def degradation_bsrgan_variant(image, sf=4, isp_model=None): + """ + This is the degradation model of BSRGAN from the paper + "Designing a Practical Degradation Model for Deep Blind Image Super-Resolution" + ---------- + sf: scale factor + isp_model: camera ISP model + Returns + ------- + img: low-quality patch, size: lq_patchsizeXlq_patchsizeXC, range: [0, 1] + hq: corresponding high-quality patch, size: (lq_patchsizexsf)X(lq_patchsizexsf)XC, range: [0, 1] + """ + image = util.uint2single(image) + isp_prob, jpeg_prob, scale2_prob = 0.25, 0.9, 0.25 + sf_ori = sf + + h1, w1 = image.shape[:2] + image = image.copy()[:w1 - w1 % sf, :h1 - h1 % sf, ...] # mod crop + h, w = image.shape[:2] + + hq = image.copy() + + if sf == 4 and random.random() < scale2_prob: # downsample1 + if np.random.rand() < 0.5: + image = cv2.resize(image, (int(1 / 2 * image.shape[1]), int(1 / 2 * image.shape[0])), + interpolation=random.choice([1, 2, 3])) + else: + image = util.imresize_np(image, 1 / 2, True) + image = np.clip(image, 0.0, 1.0) + sf = 2 + + shuffle_order = random.sample(range(7), 7) + idx1, idx2 = shuffle_order.index(2), shuffle_order.index(3) + if idx1 > idx2: # keep downsample3 last + shuffle_order[idx1], shuffle_order[idx2] = shuffle_order[idx2], shuffle_order[idx1] + + for i in shuffle_order: + + if i == 0: + image = add_blur(image, sf=sf) + + elif i == 1: + image = add_blur(image, sf=sf) + + elif i == 2: + a, b = image.shape[1], image.shape[0] + # downsample2 + if random.random() < 0.75: + sf1 = random.uniform(1, 2 * sf) + image = cv2.resize(image, (int(1 / sf1 * image.shape[1]), int(1 / sf1 * image.shape[0])), + interpolation=random.choice([1, 2, 3])) + else: + k = fspecial('gaussian', 25, random.uniform(0.1, 0.6 * sf)) + k_shifted = shift_pixel(k, sf) + k_shifted = k_shifted / k_shifted.sum() # blur with shifted kernel + image = ndimage.filters.convolve(image, np.expand_dims(k_shifted, axis=2), mode='mirror') + image = image[0::sf, 0::sf, ...] # nearest downsampling + image = np.clip(image, 0.0, 1.0) + + elif i == 3: + # downsample3 + image = cv2.resize(image, (int(1 / sf * a), int(1 / sf * b)), interpolation=random.choice([1, 2, 3])) + image = np.clip(image, 0.0, 1.0) + + elif i == 4: + # add Gaussian noise + image = add_Gaussian_noise(image, noise_level1=2, noise_level2=25) + + elif i == 5: + # add JPEG noise + if random.random() < jpeg_prob: + image = add_JPEG_noise(image) + + # elif i == 6: + # # add processed camera sensor noise + # if random.random() < isp_prob and isp_model is not None: + # with torch.no_grad(): + # img, hq = isp_model.forward(img.copy(), hq) + + # add final JPEG compression noise + image = add_JPEG_noise(image) + image = util.single2uint(image) + example = {"image":image} + return example + + +# TODO incase there is a pickle error one needs to replace a += x with a = a + x in add_speckle_noise etc... +def degradation_bsrgan_plus(img, sf=4, shuffle_prob=0.5, use_sharp=True, lq_patchsize=64, isp_model=None): + """ + This is an extended degradation model by combining + the degradation models of BSRGAN and Real-ESRGAN + ---------- + img: HXWXC, [0, 1], its size should be large than (lq_patchsizexsf)x(lq_patchsizexsf) + sf: scale factor + use_shuffle: the degradation shuffle + use_sharp: sharpening the img + Returns + ------- + img: low-quality patch, size: lq_patchsizeXlq_patchsizeXC, range: [0, 1] + hq: corresponding high-quality patch, size: (lq_patchsizexsf)X(lq_patchsizexsf)XC, range: [0, 1] + """ + + h1, w1 = img.shape[:2] + img = img.copy()[:w1 - w1 % sf, :h1 - h1 % sf, ...] # mod crop + h, w = img.shape[:2] + + if h < lq_patchsize * sf or w < lq_patchsize * sf: + raise ValueError(f'img size ({h1}X{w1}) is too small!') + + if use_sharp: + img = add_sharpening(img) + hq = img.copy() + + if random.random() < shuffle_prob: + shuffle_order = random.sample(range(13), 13) + else: + shuffle_order = list(range(13)) + # local shuffle for noise, JPEG is always the last one + shuffle_order[2:6] = random.sample(shuffle_order[2:6], len(range(2, 6))) + shuffle_order[9:13] = random.sample(shuffle_order[9:13], len(range(9, 13))) + + poisson_prob, speckle_prob, isp_prob = 0.1, 0.1, 0.1 + + for i in shuffle_order: + if i == 0: + img = add_blur(img, sf=sf) + elif i == 1: + img = add_resize(img, sf=sf) + elif i == 2: + img = add_Gaussian_noise(img, noise_level1=2, noise_level2=25) + elif i == 3: + if random.random() < poisson_prob: + img = add_Poisson_noise(img) + elif i == 4: + if random.random() < speckle_prob: + img = add_speckle_noise(img) + elif i == 5: + if random.random() < isp_prob and isp_model is not None: + with torch.no_grad(): + img, hq = isp_model.forward(img.copy(), hq) + elif i == 6: + img = add_JPEG_noise(img) + elif i == 7: + img = add_blur(img, sf=sf) + elif i == 8: + img = add_resize(img, sf=sf) + elif i == 9: + img = add_Gaussian_noise(img, noise_level1=2, noise_level2=25) + elif i == 10: + if random.random() < poisson_prob: + img = add_Poisson_noise(img) + elif i == 11: + if random.random() < speckle_prob: + img = add_speckle_noise(img) + elif i == 12: + if random.random() < isp_prob and isp_model is not None: + with torch.no_grad(): + img, hq = isp_model.forward(img.copy(), hq) + else: + print('check the shuffle!') + + # resize to desired size + img = cv2.resize(img, (int(1 / sf * hq.shape[1]), int(1 / sf * hq.shape[0])), + interpolation=random.choice([1, 2, 3])) + + # add final JPEG compression noise + img = add_JPEG_noise(img) + + # random crop + img, hq = random_crop(img, hq, sf, lq_patchsize) + + return img, hq + + +if __name__ == '__main__': + print("hey") + img = util.imread_uint('utils/test.png', 3) + print(img) + img = util.uint2single(img) + print(img) + img = img[:448, :448] + h = img.shape[0] // 4 + print("resizing to", h) + sf = 4 + deg_fn = partial(degradation_bsrgan_variant, sf=sf) + for i in range(20): + print(i) + img_lq = deg_fn(img) + print(img_lq) + img_lq_bicubic = albumentations.SmallestMaxSize(max_size=h, interpolation=cv2.INTER_CUBIC)(image=img)["image"] + print(img_lq.shape) + print("bicubic", img_lq_bicubic.shape) + print(img_hq.shape) + lq_nearest = cv2.resize(util.single2uint(img_lq), (int(sf * img_lq.shape[1]), int(sf * img_lq.shape[0])), + interpolation=0) + lq_bicubic_nearest = cv2.resize(util.single2uint(img_lq_bicubic), (int(sf * img_lq.shape[1]), int(sf * img_lq.shape[0])), + interpolation=0) + img_concat = np.concatenate([lq_bicubic_nearest, lq_nearest, util.single2uint(img_hq)], axis=1) + util.imsave(img_concat, str(i) + '.png') + + diff --git a/ldm/modules/image_degradation/bsrgan_light.py b/ldm/modules/image_degradation/bsrgan_light.py new file mode 100644 index 0000000000000000000000000000000000000000..808c7f882cb75e2ba2340d5b55881d11927351f0 --- /dev/null +++ b/ldm/modules/image_degradation/bsrgan_light.py @@ -0,0 +1,651 @@ +# -*- coding: utf-8 -*- +import numpy as np +import cv2 +import torch + +from functools import partial +import random +from scipy import ndimage +import scipy +import scipy.stats as ss +from scipy.interpolate import interp2d +from scipy.linalg import orth +import albumentations + +import ldm.modules.image_degradation.utils_image as util + +""" +# -------------------------------------------- +# Super-Resolution +# -------------------------------------------- +# +# Kai Zhang (cskaizhang@gmail.com) +# https://github.com/cszn +# From 2019/03--2021/08 +# -------------------------------------------- +""" + +def modcrop_np(img, sf): + ''' + Args: + img: numpy image, WxH or WxHxC + sf: scale factor + Return: + cropped image + ''' + w, h = img.shape[:2] + im = np.copy(img) + return im[:w - w % sf, :h - h % sf, ...] + + +""" +# -------------------------------------------- +# anisotropic Gaussian kernels +# -------------------------------------------- +""" + + +def analytic_kernel(k): + """Calculate the X4 kernel from the X2 kernel (for proof see appendix in paper)""" + k_size = k.shape[0] + # Calculate the big kernels size + big_k = np.zeros((3 * k_size - 2, 3 * k_size - 2)) + # Loop over the small kernel to fill the big one + for r in range(k_size): + for c in range(k_size): + big_k[2 * r:2 * r + k_size, 2 * c:2 * c + k_size] += k[r, c] * k + # Crop the edges of the big kernel to ignore very small values and increase run time of SR + crop = k_size // 2 + cropped_big_k = big_k[crop:-crop, crop:-crop] + # Normalize to 1 + return cropped_big_k / cropped_big_k.sum() + + +def anisotropic_Gaussian(ksize=15, theta=np.pi, l1=6, l2=6): + """ generate an anisotropic Gaussian kernel + Args: + ksize : e.g., 15, kernel size + theta : [0, pi], rotation angle range + l1 : [0.1,50], scaling of eigenvalues + l2 : [0.1,l1], scaling of eigenvalues + If l1 = l2, will get an isotropic Gaussian kernel. + Returns: + k : kernel + """ + + v = np.dot(np.array([[np.cos(theta), -np.sin(theta)], [np.sin(theta), np.cos(theta)]]), np.array([1., 0.])) + V = np.array([[v[0], v[1]], [v[1], -v[0]]]) + D = np.array([[l1, 0], [0, l2]]) + Sigma = np.dot(np.dot(V, D), np.linalg.inv(V)) + k = gm_blur_kernel(mean=[0, 0], cov=Sigma, size=ksize) + + return k + + +def gm_blur_kernel(mean, cov, size=15): + center = size / 2.0 + 0.5 + k = np.zeros([size, size]) + for y in range(size): + for x in range(size): + cy = y - center + 1 + cx = x - center + 1 + k[y, x] = ss.multivariate_normal.pdf([cx, cy], mean=mean, cov=cov) + + k = k / np.sum(k) + return k + + +def shift_pixel(x, sf, upper_left=True): + """shift pixel for super-resolution with different scale factors + Args: + x: WxHxC or WxH + sf: scale factor + upper_left: shift direction + """ + h, w = x.shape[:2] + shift = (sf - 1) * 0.5 + xv, yv = np.arange(0, w, 1.0), np.arange(0, h, 1.0) + if upper_left: + x1 = xv + shift + y1 = yv + shift + else: + x1 = xv - shift + y1 = yv - shift + + x1 = np.clip(x1, 0, w - 1) + y1 = np.clip(y1, 0, h - 1) + + if x.ndim == 2: + x = interp2d(xv, yv, x)(x1, y1) + if x.ndim == 3: + for i in range(x.shape[-1]): + x[:, :, i] = interp2d(xv, yv, x[:, :, i])(x1, y1) + + return x + + +def blur(x, k): + ''' + x: image, NxcxHxW + k: kernel, Nx1xhxw + ''' + n, c = x.shape[:2] + p1, p2 = (k.shape[-2] - 1) // 2, (k.shape[-1] - 1) // 2 + x = torch.nn.functional.pad(x, pad=(p1, p2, p1, p2), mode='replicate') + k = k.repeat(1, c, 1, 1) + k = k.view(-1, 1, k.shape[2], k.shape[3]) + x = x.view(1, -1, x.shape[2], x.shape[3]) + x = torch.nn.functional.conv2d(x, k, bias=None, stride=1, padding=0, groups=n * c) + x = x.view(n, c, x.shape[2], x.shape[3]) + + return x + + +def gen_kernel(k_size=np.array([15, 15]), scale_factor=np.array([4, 4]), min_var=0.6, max_var=10., noise_level=0): + """" + # modified version of https://github.com/assafshocher/BlindSR_dataset_generator + # Kai Zhang + # min_var = 0.175 * sf # variance of the gaussian kernel will be sampled between min_var and max_var + # max_var = 2.5 * sf + """ + # Set random eigen-vals (lambdas) and angle (theta) for COV matrix + lambda_1 = min_var + np.random.rand() * (max_var - min_var) + lambda_2 = min_var + np.random.rand() * (max_var - min_var) + theta = np.random.rand() * np.pi # random theta + noise = -noise_level + np.random.rand(*k_size) * noise_level * 2 + + # Set COV matrix using Lambdas and Theta + LAMBDA = np.diag([lambda_1, lambda_2]) + Q = np.array([[np.cos(theta), -np.sin(theta)], + [np.sin(theta), np.cos(theta)]]) + SIGMA = Q @ LAMBDA @ Q.T + INV_SIGMA = np.linalg.inv(SIGMA)[None, None, :, :] + + # Set expectation position (shifting kernel for aligned image) + MU = k_size // 2 - 0.5 * (scale_factor - 1) # - 0.5 * (scale_factor - k_size % 2) + MU = MU[None, None, :, None] + + # Create meshgrid for Gaussian + [X, Y] = np.meshgrid(range(k_size[0]), range(k_size[1])) + Z = np.stack([X, Y], 2)[:, :, :, None] + + # Calcualte Gaussian for every pixel of the kernel + ZZ = Z - MU + ZZ_t = ZZ.transpose(0, 1, 3, 2) + raw_kernel = np.exp(-0.5 * np.squeeze(ZZ_t @ INV_SIGMA @ ZZ)) * (1 + noise) + + # shift the kernel so it will be centered + # raw_kernel_centered = kernel_shift(raw_kernel, scale_factor) + + # Normalize the kernel and return + # kernel = raw_kernel_centered / np.sum(raw_kernel_centered) + kernel = raw_kernel / np.sum(raw_kernel) + return kernel + + +def fspecial_gaussian(hsize, sigma): + hsize = [hsize, hsize] + siz = [(hsize[0] - 1.0) / 2.0, (hsize[1] - 1.0) / 2.0] + std = sigma + [x, y] = np.meshgrid(np.arange(-siz[1], siz[1] + 1), np.arange(-siz[0], siz[0] + 1)) + arg = -(x * x + y * y) / (2 * std * std) + h = np.exp(arg) + h[h < scipy.finfo(float).eps * h.max()] = 0 + sumh = h.sum() + if sumh != 0: + h = h / sumh + return h + + +def fspecial_laplacian(alpha): + alpha = max([0, min([alpha, 1])]) + h1 = alpha / (alpha + 1) + h2 = (1 - alpha) / (alpha + 1) + h = [[h1, h2, h1], [h2, -4 / (alpha + 1), h2], [h1, h2, h1]] + h = np.array(h) + return h + + +def fspecial(filter_type, *args, **kwargs): + ''' + python code from: + https://github.com/ronaldosena/imagens-medicas-2/blob/40171a6c259edec7827a6693a93955de2bd39e76/Aulas/aula_2_-_uniform_filter/matlab_fspecial.py + ''' + if filter_type == 'gaussian': + return fspecial_gaussian(*args, **kwargs) + if filter_type == 'laplacian': + return fspecial_laplacian(*args, **kwargs) + + +""" +# -------------------------------------------- +# degradation models +# -------------------------------------------- +""" + + +def bicubic_degradation(x, sf=3): + ''' + Args: + x: HxWxC image, [0, 1] + sf: down-scale factor + Return: + bicubicly downsampled LR image + ''' + x = util.imresize_np(x, scale=1 / sf) + return x + + +def srmd_degradation(x, k, sf=3): + ''' blur + bicubic downsampling + Args: + x: HxWxC image, [0, 1] + k: hxw, double + sf: down-scale factor + Return: + downsampled LR image + Reference: + @inproceedings{zhang2018learning, + title={Learning a single convolutional super-resolution network for multiple degradations}, + author={Zhang, Kai and Zuo, Wangmeng and Zhang, Lei}, + booktitle={IEEE Conference on Computer Vision and Pattern Recognition}, + pages={3262--3271}, + year={2018} + } + ''' + x = ndimage.convolve(x, np.expand_dims(k, axis=2), mode='wrap') # 'nearest' | 'mirror' + x = bicubic_degradation(x, sf=sf) + return x + + +def dpsr_degradation(x, k, sf=3): + ''' bicubic downsampling + blur + Args: + x: HxWxC image, [0, 1] + k: hxw, double + sf: down-scale factor + Return: + downsampled LR image + Reference: + @inproceedings{zhang2019deep, + title={Deep Plug-and-Play Super-Resolution for Arbitrary Blur Kernels}, + author={Zhang, Kai and Zuo, Wangmeng and Zhang, Lei}, + booktitle={IEEE Conference on Computer Vision and Pattern Recognition}, + pages={1671--1681}, + year={2019} + } + ''' + x = bicubic_degradation(x, sf=sf) + x = ndimage.convolve(x, np.expand_dims(k, axis=2), mode='wrap') + return x + + +def classical_degradation(x, k, sf=3): + ''' blur + downsampling + Args: + x: HxWxC image, [0, 1]/[0, 255] + k: hxw, double + sf: down-scale factor + Return: + downsampled LR image + ''' + x = ndimage.convolve(x, np.expand_dims(k, axis=2), mode='wrap') + # x = filters.correlate(x, np.expand_dims(np.flip(k), axis=2)) + st = 0 + return x[st::sf, st::sf, ...] + + +def add_sharpening(img, weight=0.5, radius=50, threshold=10): + """USM sharpening. borrowed from real-ESRGAN + Input image: I; Blurry image: B. + 1. K = I + weight * (I - B) + 2. Mask = 1 if abs(I - B) > threshold, else: 0 + 3. Blur mask: + 4. Out = Mask * K + (1 - Mask) * I + Args: + img (Numpy array): Input image, HWC, BGR; float32, [0, 1]. + weight (float): Sharp weight. Default: 1. + radius (float): Kernel size of Gaussian blur. Default: 50. + threshold (int): + """ + if radius % 2 == 0: + radius += 1 + blur = cv2.GaussianBlur(img, (radius, radius), 0) + residual = img - blur + mask = np.abs(residual) * 255 > threshold + mask = mask.astype('float32') + soft_mask = cv2.GaussianBlur(mask, (radius, radius), 0) + + K = img + weight * residual + K = np.clip(K, 0, 1) + return soft_mask * K + (1 - soft_mask) * img + + +def add_blur(img, sf=4): + wd2 = 4.0 + sf + wd = 2.0 + 0.2 * sf + + wd2 = wd2/4 + wd = wd/4 + + if random.random() < 0.5: + l1 = wd2 * random.random() + l2 = wd2 * random.random() + k = anisotropic_Gaussian(ksize=random.randint(2, 11) + 3, theta=random.random() * np.pi, l1=l1, l2=l2) + else: + k = fspecial('gaussian', random.randint(2, 4) + 3, wd * random.random()) + img = ndimage.convolve(img, np.expand_dims(k, axis=2), mode='mirror') + + return img + + +def add_resize(img, sf=4): + rnum = np.random.rand() + if rnum > 0.8: # up + sf1 = random.uniform(1, 2) + elif rnum < 0.7: # down + sf1 = random.uniform(0.5 / sf, 1) + else: + sf1 = 1.0 + img = cv2.resize(img, (int(sf1 * img.shape[1]), int(sf1 * img.shape[0])), interpolation=random.choice([1, 2, 3])) + img = np.clip(img, 0.0, 1.0) + + return img + + +# def add_Gaussian_noise(img, noise_level1=2, noise_level2=25): +# noise_level = random.randint(noise_level1, noise_level2) +# rnum = np.random.rand() +# if rnum > 0.6: # add color Gaussian noise +# img += np.random.normal(0, noise_level / 255.0, img.shape).astype(np.float32) +# elif rnum < 0.4: # add grayscale Gaussian noise +# img += np.random.normal(0, noise_level / 255.0, (*img.shape[:2], 1)).astype(np.float32) +# else: # add noise +# L = noise_level2 / 255. +# D = np.diag(np.random.rand(3)) +# U = orth(np.random.rand(3, 3)) +# conv = np.dot(np.dot(np.transpose(U), D), U) +# img += np.random.multivariate_normal([0, 0, 0], np.abs(L ** 2 * conv), img.shape[:2]).astype(np.float32) +# img = np.clip(img, 0.0, 1.0) +# return img + +def add_Gaussian_noise(img, noise_level1=2, noise_level2=25): + noise_level = random.randint(noise_level1, noise_level2) + rnum = np.random.rand() + if rnum > 0.6: # add color Gaussian noise + img = img + np.random.normal(0, noise_level / 255.0, img.shape).astype(np.float32) + elif rnum < 0.4: # add grayscale Gaussian noise + img = img + np.random.normal(0, noise_level / 255.0, (*img.shape[:2], 1)).astype(np.float32) + else: # add noise + L = noise_level2 / 255. + D = np.diag(np.random.rand(3)) + U = orth(np.random.rand(3, 3)) + conv = np.dot(np.dot(np.transpose(U), D), U) + img = img + np.random.multivariate_normal([0, 0, 0], np.abs(L ** 2 * conv), img.shape[:2]).astype(np.float32) + img = np.clip(img, 0.0, 1.0) + return img + + +def add_speckle_noise(img, noise_level1=2, noise_level2=25): + noise_level = random.randint(noise_level1, noise_level2) + img = np.clip(img, 0.0, 1.0) + rnum = random.random() + if rnum > 0.6: + img += img * np.random.normal(0, noise_level / 255.0, img.shape).astype(np.float32) + elif rnum < 0.4: + img += img * np.random.normal(0, noise_level / 255.0, (*img.shape[:2], 1)).astype(np.float32) + else: + L = noise_level2 / 255. + D = np.diag(np.random.rand(3)) + U = orth(np.random.rand(3, 3)) + conv = np.dot(np.dot(np.transpose(U), D), U) + img += img * np.random.multivariate_normal([0, 0, 0], np.abs(L ** 2 * conv), img.shape[:2]).astype(np.float32) + img = np.clip(img, 0.0, 1.0) + return img + + +def add_Poisson_noise(img): + img = np.clip((img * 255.0).round(), 0, 255) / 255. + vals = 10 ** (2 * random.random() + 2.0) # [2, 4] + if random.random() < 0.5: + img = np.random.poisson(img * vals).astype(np.float32) / vals + else: + img_gray = np.dot(img[..., :3], [0.299, 0.587, 0.114]) + img_gray = np.clip((img_gray * 255.0).round(), 0, 255) / 255. + noise_gray = np.random.poisson(img_gray * vals).astype(np.float32) / vals - img_gray + img += noise_gray[:, :, np.newaxis] + img = np.clip(img, 0.0, 1.0) + return img + + +def add_JPEG_noise(img): + quality_factor = random.randint(80, 95) + img = cv2.cvtColor(util.single2uint(img), cv2.COLOR_RGB2BGR) + result, encimg = cv2.imencode('.jpg', img, [int(cv2.IMWRITE_JPEG_QUALITY), quality_factor]) + img = cv2.imdecode(encimg, 1) + img = cv2.cvtColor(util.uint2single(img), cv2.COLOR_BGR2RGB) + return img + + +def random_crop(lq, hq, sf=4, lq_patchsize=64): + h, w = lq.shape[:2] + rnd_h = random.randint(0, h - lq_patchsize) + rnd_w = random.randint(0, w - lq_patchsize) + lq = lq[rnd_h:rnd_h + lq_patchsize, rnd_w:rnd_w + lq_patchsize, :] + + rnd_h_H, rnd_w_H = int(rnd_h * sf), int(rnd_w * sf) + hq = hq[rnd_h_H:rnd_h_H + lq_patchsize * sf, rnd_w_H:rnd_w_H + lq_patchsize * sf, :] + return lq, hq + + +def degradation_bsrgan(img, sf=4, lq_patchsize=72, isp_model=None): + """ + This is the degradation model of BSRGAN from the paper + "Designing a Practical Degradation Model for Deep Blind Image Super-Resolution" + ---------- + img: HXWXC, [0, 1], its size should be large than (lq_patchsizexsf)x(lq_patchsizexsf) + sf: scale factor + isp_model: camera ISP model + Returns + ------- + img: low-quality patch, size: lq_patchsizeXlq_patchsizeXC, range: [0, 1] + hq: corresponding high-quality patch, size: (lq_patchsizexsf)X(lq_patchsizexsf)XC, range: [0, 1] + """ + isp_prob, jpeg_prob, scale2_prob = 0.25, 0.9, 0.25 + sf_ori = sf + + h1, w1 = img.shape[:2] + img = img.copy()[:w1 - w1 % sf, :h1 - h1 % sf, ...] # mod crop + h, w = img.shape[:2] + + if h < lq_patchsize * sf or w < lq_patchsize * sf: + raise ValueError(f'img size ({h1}X{w1}) is too small!') + + hq = img.copy() + + if sf == 4 and random.random() < scale2_prob: # downsample1 + if np.random.rand() < 0.5: + img = cv2.resize(img, (int(1 / 2 * img.shape[1]), int(1 / 2 * img.shape[0])), + interpolation=random.choice([1, 2, 3])) + else: + img = util.imresize_np(img, 1 / 2, True) + img = np.clip(img, 0.0, 1.0) + sf = 2 + + shuffle_order = random.sample(range(7), 7) + idx1, idx2 = shuffle_order.index(2), shuffle_order.index(3) + if idx1 > idx2: # keep downsample3 last + shuffle_order[idx1], shuffle_order[idx2] = shuffle_order[idx2], shuffle_order[idx1] + + for i in shuffle_order: + + if i == 0: + img = add_blur(img, sf=sf) + + elif i == 1: + img = add_blur(img, sf=sf) + + elif i == 2: + a, b = img.shape[1], img.shape[0] + # downsample2 + if random.random() < 0.75: + sf1 = random.uniform(1, 2 * sf) + img = cv2.resize(img, (int(1 / sf1 * img.shape[1]), int(1 / sf1 * img.shape[0])), + interpolation=random.choice([1, 2, 3])) + else: + k = fspecial('gaussian', 25, random.uniform(0.1, 0.6 * sf)) + k_shifted = shift_pixel(k, sf) + k_shifted = k_shifted / k_shifted.sum() # blur with shifted kernel + img = ndimage.convolve(img, np.expand_dims(k_shifted, axis=2), mode='mirror') + img = img[0::sf, 0::sf, ...] # nearest downsampling + img = np.clip(img, 0.0, 1.0) + + elif i == 3: + # downsample3 + img = cv2.resize(img, (int(1 / sf * a), int(1 / sf * b)), interpolation=random.choice([1, 2, 3])) + img = np.clip(img, 0.0, 1.0) + + elif i == 4: + # add Gaussian noise + img = add_Gaussian_noise(img, noise_level1=2, noise_level2=8) + + elif i == 5: + # add JPEG noise + if random.random() < jpeg_prob: + img = add_JPEG_noise(img) + + elif i == 6: + # add processed camera sensor noise + if random.random() < isp_prob and isp_model is not None: + with torch.no_grad(): + img, hq = isp_model.forward(img.copy(), hq) + + # add final JPEG compression noise + img = add_JPEG_noise(img) + + # random crop + img, hq = random_crop(img, hq, sf_ori, lq_patchsize) + + return img, hq + + +# todo no isp_model? +def degradation_bsrgan_variant(image, sf=4, isp_model=None, up=False): + """ + This is the degradation model of BSRGAN from the paper + "Designing a Practical Degradation Model for Deep Blind Image Super-Resolution" + ---------- + sf: scale factor + isp_model: camera ISP model + Returns + ------- + img: low-quality patch, size: lq_patchsizeXlq_patchsizeXC, range: [0, 1] + hq: corresponding high-quality patch, size: (lq_patchsizexsf)X(lq_patchsizexsf)XC, range: [0, 1] + """ + image = util.uint2single(image) + isp_prob, jpeg_prob, scale2_prob = 0.25, 0.9, 0.25 + sf_ori = sf + + h1, w1 = image.shape[:2] + image = image.copy()[:w1 - w1 % sf, :h1 - h1 % sf, ...] # mod crop + h, w = image.shape[:2] + + hq = image.copy() + + if sf == 4 and random.random() < scale2_prob: # downsample1 + if np.random.rand() < 0.5: + image = cv2.resize(image, (int(1 / 2 * image.shape[1]), int(1 / 2 * image.shape[0])), + interpolation=random.choice([1, 2, 3])) + else: + image = util.imresize_np(image, 1 / 2, True) + image = np.clip(image, 0.0, 1.0) + sf = 2 + + shuffle_order = random.sample(range(7), 7) + idx1, idx2 = shuffle_order.index(2), shuffle_order.index(3) + if idx1 > idx2: # keep downsample3 last + shuffle_order[idx1], shuffle_order[idx2] = shuffle_order[idx2], shuffle_order[idx1] + + for i in shuffle_order: + + if i == 0: + image = add_blur(image, sf=sf) + + # elif i == 1: + # image = add_blur(image, sf=sf) + + if i == 0: + pass + + elif i == 2: + a, b = image.shape[1], image.shape[0] + # downsample2 + if random.random() < 0.8: + sf1 = random.uniform(1, 2 * sf) + image = cv2.resize(image, (int(1 / sf1 * image.shape[1]), int(1 / sf1 * image.shape[0])), + interpolation=random.choice([1, 2, 3])) + else: + k = fspecial('gaussian', 25, random.uniform(0.1, 0.6 * sf)) + k_shifted = shift_pixel(k, sf) + k_shifted = k_shifted / k_shifted.sum() # blur with shifted kernel + image = ndimage.convolve(image, np.expand_dims(k_shifted, axis=2), mode='mirror') + image = image[0::sf, 0::sf, ...] # nearest downsampling + + image = np.clip(image, 0.0, 1.0) + + elif i == 3: + # downsample3 + image = cv2.resize(image, (int(1 / sf * a), int(1 / sf * b)), interpolation=random.choice([1, 2, 3])) + image = np.clip(image, 0.0, 1.0) + + elif i == 4: + # add Gaussian noise + image = add_Gaussian_noise(image, noise_level1=1, noise_level2=2) + + elif i == 5: + # add JPEG noise + if random.random() < jpeg_prob: + image = add_JPEG_noise(image) + # + # elif i == 6: + # # add processed camera sensor noise + # if random.random() < isp_prob and isp_model is not None: + # with torch.no_grad(): + # img, hq = isp_model.forward(img.copy(), hq) + + # add final JPEG compression noise + image = add_JPEG_noise(image) + image = util.single2uint(image) + if up: + image = cv2.resize(image, (w1, h1), interpolation=cv2.INTER_CUBIC) # todo: random, as above? want to condition on it then + example = {"image": image} + return example + + + + +if __name__ == '__main__': + print("hey") + img = util.imread_uint('utils/test.png', 3) + img = img[:448, :448] + h = img.shape[0] // 4 + print("resizing to", h) + sf = 4 + deg_fn = partial(degradation_bsrgan_variant, sf=sf) + for i in range(20): + print(i) + img_hq = img + img_lq = deg_fn(img)["image"] + img_hq, img_lq = util.uint2single(img_hq), util.uint2single(img_lq) + print(img_lq) + img_lq_bicubic = albumentations.SmallestMaxSize(max_size=h, interpolation=cv2.INTER_CUBIC)(image=img_hq)["image"] + print(img_lq.shape) + print("bicubic", img_lq_bicubic.shape) + print(img_hq.shape) + lq_nearest = cv2.resize(util.single2uint(img_lq), (int(sf * img_lq.shape[1]), int(sf * img_lq.shape[0])), + interpolation=0) + lq_bicubic_nearest = cv2.resize(util.single2uint(img_lq_bicubic), + (int(sf * img_lq.shape[1]), int(sf * img_lq.shape[0])), + interpolation=0) + img_concat = np.concatenate([lq_bicubic_nearest, lq_nearest, util.single2uint(img_hq)], axis=1) + util.imsave(img_concat, str(i) + '.png') diff --git a/ldm/modules/image_degradation/utils/test.png b/ldm/modules/image_degradation/utils/test.png new file mode 100644 index 0000000000000000000000000000000000000000..4249b43de0f22707758d13c240268a401642f6e6 Binary files /dev/null and b/ldm/modules/image_degradation/utils/test.png differ diff --git a/ldm/modules/image_degradation/utils_image.py b/ldm/modules/image_degradation/utils_image.py new file mode 100644 index 0000000000000000000000000000000000000000..0175f155ad900ae33c3c46ed87f49b352e3faf98 --- /dev/null +++ b/ldm/modules/image_degradation/utils_image.py @@ -0,0 +1,916 @@ +import os +import math +import random +import numpy as np +import torch +import cv2 +from torchvision.utils import make_grid +from datetime import datetime +#import matplotlib.pyplot as plt # TODO: check with Dominik, also bsrgan.py vs bsrgan_light.py + + +os.environ["KMP_DUPLICATE_LIB_OK"]="TRUE" + + +''' +# -------------------------------------------- +# Kai Zhang (github: https://github.com/cszn) +# 03/Mar/2019 +# -------------------------------------------- +# https://github.com/twhui/SRGAN-pyTorch +# https://github.com/xinntao/BasicSR +# -------------------------------------------- +''' + + +IMG_EXTENSIONS = ['.jpg', '.JPG', '.jpeg', '.JPEG', '.png', '.PNG', '.ppm', '.PPM', '.bmp', '.BMP', '.tif'] + + +def is_image_file(filename): + return any(filename.endswith(extension) for extension in IMG_EXTENSIONS) + + +def get_timestamp(): + return datetime.now().strftime('%y%m%d-%H%M%S') + + +def imshow(x, title=None, cbar=False, figsize=None): + plt.figure(figsize=figsize) + plt.imshow(np.squeeze(x), interpolation='nearest', cmap='gray') + if title: + plt.title(title) + if cbar: + plt.colorbar() + plt.show() + + +def surf(Z, cmap='rainbow', figsize=None): + plt.figure(figsize=figsize) + ax3 = plt.axes(projection='3d') + + w, h = Z.shape[:2] + xx = np.arange(0,w,1) + yy = np.arange(0,h,1) + X, Y = np.meshgrid(xx, yy) + ax3.plot_surface(X,Y,Z,cmap=cmap) + #ax3.contour(X,Y,Z, zdim='z',offset=-2，cmap=cmap) + plt.show() + + +''' +# -------------------------------------------- +# get image pathes +# -------------------------------------------- +''' + + +def get_image_paths(dataroot): + paths = None # return None if dataroot is None + if dataroot is not None: + paths = sorted(_get_paths_from_images(dataroot)) + return paths + + +def _get_paths_from_images(path): + assert os.path.isdir(path), '{:s} is not a valid directory'.format(path) + images = [] + for dirpath, _, fnames in sorted(os.walk(path)): + for fname in sorted(fnames): + if is_image_file(fname): + img_path = os.path.join(dirpath, fname) + images.append(img_path) + assert images, '{:s} has no valid image file'.format(path) + return images + + +''' +# -------------------------------------------- +# split large images into small images +# -------------------------------------------- +''' + + +def patches_from_image(img, p_size=512, p_overlap=64, p_max=800): + w, h = img.shape[:2] + patches = [] + if w > p_max and h > p_max: + w1 = list(np.arange(0, w-p_size, p_size-p_overlap, dtype=np.int)) + h1 = list(np.arange(0, h-p_size, p_size-p_overlap, dtype=np.int)) + w1.append(w-p_size) + h1.append(h-p_size) +# print(w1) +# print(h1) + for i in w1: + for j in h1: + patches.append(img[i:i+p_size, j:j+p_size,:]) + else: + patches.append(img) + + return patches + + +def imssave(imgs, img_path): + """ + imgs: list, N images of size WxHxC + """ + img_name, ext = os.path.splitext(os.path.basename(img_path)) + + for i, img in enumerate(imgs): + if img.ndim == 3: + img = img[:, :, [2, 1, 0]] + new_path = os.path.join(os.path.dirname(img_path), img_name+str('_s{:04d}'.format(i))+'.png') + cv2.imwrite(new_path, img) + + +def split_imageset(original_dataroot, taget_dataroot, n_channels=3, p_size=800, p_overlap=96, p_max=1000): + """ + split the large images from original_dataroot into small overlapped images with size (p_size)x(p_size), + and save them into taget_dataroot; only the images with larger size than (p_max)x(p_max) + will be splitted. + Args: + original_dataroot: + taget_dataroot: + p_size: size of small images + p_overlap: patch size in training is a good choice + p_max: images with smaller size than (p_max)x(p_max) keep unchanged. + """ + paths = get_image_paths(original_dataroot) + for img_path in paths: + # img_name, ext = os.path.splitext(os.path.basename(img_path)) + img = imread_uint(img_path, n_channels=n_channels) + patches = patches_from_image(img, p_size, p_overlap, p_max) + imssave(patches, os.path.join(taget_dataroot,os.path.basename(img_path))) + #if original_dataroot == taget_dataroot: + #del img_path + +''' +# -------------------------------------------- +# makedir +# -------------------------------------------- +''' + + +def mkdir(path): + if not os.path.exists(path): + os.makedirs(path) + + +def mkdirs(paths): + if isinstance(paths, str): + mkdir(paths) + else: + for path in paths: + mkdir(path) + + +def mkdir_and_rename(path): + if os.path.exists(path): + new_name = path + '_archived_' + get_timestamp() + print('Path already exists. Rename it to [{:s}]'.format(new_name)) + os.rename(path, new_name) + os.makedirs(path) + + +''' +# -------------------------------------------- +# read image from path +# opencv is fast, but read BGR numpy image +# -------------------------------------------- +''' + + +# -------------------------------------------- +# get uint8 image of size HxWxn_channles (RGB) +# -------------------------------------------- +def imread_uint(path, n_channels=3): + # input: path + # output: HxWx3(RGB or GGG), or HxWx1 (G) + if n_channels == 1: + img = cv2.imread(path, 0) # cv2.IMREAD_GRAYSCALE + img = np.expand_dims(img, axis=2) # HxWx1 + elif n_channels == 3: + img = cv2.imread(path, cv2.IMREAD_UNCHANGED) # BGR or G + if img.ndim == 2: + img = cv2.cvtColor(img, cv2.COLOR_GRAY2RGB) # GGG + else: + img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB) # RGB + return img + + +# -------------------------------------------- +# matlab's imwrite +# -------------------------------------------- +def imsave(img, img_path): + img = np.squeeze(img) + if img.ndim == 3: + img = img[:, :, [2, 1, 0]] + cv2.imwrite(img_path, img) + +def imwrite(img, img_path): + img = np.squeeze(img) + if img.ndim == 3: + img = img[:, :, [2, 1, 0]] + cv2.imwrite(img_path, img) + + + +# -------------------------------------------- +# get single image of size HxWxn_channles (BGR) +# -------------------------------------------- +def read_img(path): + # read image by cv2 + # return: Numpy float32, HWC, BGR, [0,1] + img = cv2.imread(path, cv2.IMREAD_UNCHANGED) # cv2.IMREAD_GRAYSCALE + img = img.astype(np.float32) / 255. + if img.ndim == 2: + img = np.expand_dims(img, axis=2) + # some images have 4 channels + if img.shape[2] > 3: + img = img[:, :, :3] + return img + + +''' +# -------------------------------------------- +# image format conversion +# -------------------------------------------- +# numpy(single) <---> numpy(unit) +# numpy(single) <---> tensor +# numpy(unit) <---> tensor +# -------------------------------------------- +''' + + +# -------------------------------------------- +# numpy(single) [0, 1] <---> numpy(unit) +# -------------------------------------------- + + +def uint2single(img): + + return np.float32(img/255.) + + +def single2uint(img): + + return np.uint8((img.clip(0, 1)*255.).round()) + + +def uint162single(img): + + return np.float32(img/65535.) + + +def single2uint16(img): + + return np.uint16((img.clip(0, 1)*65535.).round()) + + +# -------------------------------------------- +# numpy(unit) (HxWxC or HxW) <---> tensor +# -------------------------------------------- + + +# convert uint to 4-dimensional torch tensor +def uint2tensor4(img): + if img.ndim == 2: + img = np.expand_dims(img, axis=2) + return torch.from_numpy(np.ascontiguousarray(img)).permute(2, 0, 1).float().div(255.).unsqueeze(0) + + +# convert uint to 3-dimensional torch tensor +def uint2tensor3(img): + if img.ndim == 2: + img = np.expand_dims(img, axis=2) + return torch.from_numpy(np.ascontiguousarray(img)).permute(2, 0, 1).float().div(255.) + + +# convert 2/3/4-dimensional torch tensor to uint +def tensor2uint(img): + img = img.data.squeeze().float().clamp_(0, 1).cpu().numpy() + if img.ndim == 3: + img = np.transpose(img, (1, 2, 0)) + return np.uint8((img*255.0).round()) + + +# -------------------------------------------- +# numpy(single) (HxWxC) <---> tensor +# -------------------------------------------- + + +# convert single (HxWxC) to 3-dimensional torch tensor +def single2tensor3(img): + return torch.from_numpy(np.ascontiguousarray(img)).permute(2, 0, 1).float() + + +# convert single (HxWxC) to 4-dimensional torch tensor +def single2tensor4(img): + return torch.from_numpy(np.ascontiguousarray(img)).permute(2, 0, 1).float().unsqueeze(0) + + +# convert torch tensor to single +def tensor2single(img): + img = img.data.squeeze().float().cpu().numpy() + if img.ndim == 3: + img = np.transpose(img, (1, 2, 0)) + + return img + +# convert torch tensor to single +def tensor2single3(img): + img = img.data.squeeze().float().cpu().numpy() + if img.ndim == 3: + img = np.transpose(img, (1, 2, 0)) + elif img.ndim == 2: + img = np.expand_dims(img, axis=2) + return img + + +def single2tensor5(img): + return torch.from_numpy(np.ascontiguousarray(img)).permute(2, 0, 1, 3).float().unsqueeze(0) + + +def single32tensor5(img): + return torch.from_numpy(np.ascontiguousarray(img)).float().unsqueeze(0).unsqueeze(0) + + +def single42tensor4(img): + return torch.from_numpy(np.ascontiguousarray(img)).permute(2, 0, 1, 3).float() + + +# from skimage.io import imread, imsave +def tensor2img(tensor, out_type=np.uint8, min_max=(0, 1)): + ''' + Converts a torch Tensor into an image Numpy array of BGR channel order + Input: 4D(B,(3/1),H,W), 3D(C,H,W), or 2D(H,W), any range, RGB channel order + Output: 3D(H,W,C) or 2D(H,W), [0,255], np.uint8 (default) + ''' + tensor = tensor.squeeze().float().cpu().clamp_(*min_max) # squeeze first, then clamp + tensor = (tensor - min_max[0]) / (min_max[1] - min_max[0]) # to range [0,1] + n_dim = tensor.dim() + if n_dim == 4: + n_img = len(tensor) + img_np = make_grid(tensor, nrow=int(math.sqrt(n_img)), normalize=False).numpy() + img_np = np.transpose(img_np[[2, 1, 0], :, :], (1, 2, 0)) # HWC, BGR + elif n_dim == 3: + img_np = tensor.numpy() + img_np = np.transpose(img_np[[2, 1, 0], :, :], (1, 2, 0)) # HWC, BGR + elif n_dim == 2: + img_np = tensor.numpy() + else: + raise TypeError( + 'Only support 4D, 3D and 2D tensor. But received with dimension: {:d}'.format(n_dim)) + if out_type == np.uint8: + img_np = (img_np * 255.0).round() + # Important. Unlike matlab, numpy.unit8() WILL NOT round by default. + return img_np.astype(out_type) + + +''' +# -------------------------------------------- +# Augmentation, flipe and/or rotate +# -------------------------------------------- +# The following two are enough. +# (1) augmet_img: numpy image of WxHxC or WxH +# (2) augment_img_tensor4: tensor image 1xCxWxH +# -------------------------------------------- +''' + + +def augment_img(img, mode=0): + '''Kai Zhang (github: https://github.com/cszn) + ''' + if mode == 0: + return img + elif mode == 1: + return np.flipud(np.rot90(img)) + elif mode == 2: + return np.flipud(img) + elif mode == 3: + return np.rot90(img, k=3) + elif mode == 4: + return np.flipud(np.rot90(img, k=2)) + elif mode == 5: + return np.rot90(img) + elif mode == 6: + return np.rot90(img, k=2) + elif mode == 7: + return np.flipud(np.rot90(img, k=3)) + + +def augment_img_tensor4(img, mode=0): + '''Kai Zhang (github: https://github.com/cszn) + ''' + if mode == 0: + return img + elif mode == 1: + return img.rot90(1, [2, 3]).flip([2]) + elif mode == 2: + return img.flip([2]) + elif mode == 3: + return img.rot90(3, [2, 3]) + elif mode == 4: + return img.rot90(2, [2, 3]).flip([2]) + elif mode == 5: + return img.rot90(1, [2, 3]) + elif mode == 6: + return img.rot90(2, [2, 3]) + elif mode == 7: + return img.rot90(3, [2, 3]).flip([2]) + + +def augment_img_tensor(img, mode=0): + '''Kai Zhang (github: https://github.com/cszn) + ''' + img_size = img.size() + img_np = img.data.cpu().numpy() + if len(img_size) == 3: + img_np = np.transpose(img_np, (1, 2, 0)) + elif len(img_size) == 4: + img_np = np.transpose(img_np, (2, 3, 1, 0)) + img_np = augment_img(img_np, mode=mode) + img_tensor = torch.from_numpy(np.ascontiguousarray(img_np)) + if len(img_size) == 3: + img_tensor = img_tensor.permute(2, 0, 1) + elif len(img_size) == 4: + img_tensor = img_tensor.permute(3, 2, 0, 1) + + return img_tensor.type_as(img) + + +def augment_img_np3(img, mode=0): + if mode == 0: + return img + elif mode == 1: + return img.transpose(1, 0, 2) + elif mode == 2: + return img[::-1, :, :] + elif mode == 3: + img = img[::-1, :, :] + img = img.transpose(1, 0, 2) + return img + elif mode == 4: + return img[:, ::-1, :] + elif mode == 5: + img = img[:, ::-1, :] + img = img.transpose(1, 0, 2) + return img + elif mode == 6: + img = img[:, ::-1, :] + img = img[::-1, :, :] + return img + elif mode == 7: + img = img[:, ::-1, :] + img = img[::-1, :, :] + img = img.transpose(1, 0, 2) + return img + + +def augment_imgs(img_list, hflip=True, rot=True): + # horizontal flip OR rotate + hflip = hflip and random.random() < 0.5 + vflip = rot and random.random() < 0.5 + rot90 = rot and random.random() < 0.5 + + def _augment(img): + if hflip: + img = img[:, ::-1, :] + if vflip: + img = img[::-1, :, :] + if rot90: + img = img.transpose(1, 0, 2) + return img + + return [_augment(img) for img in img_list] + + +''' +# -------------------------------------------- +# modcrop and shave +# -------------------------------------------- +''' + + +def modcrop(img_in, scale): + # img_in: Numpy, HWC or HW + img = np.copy(img_in) + if img.ndim == 2: + H, W = img.shape + H_r, W_r = H % scale, W % scale + img = img[:H - H_r, :W - W_r] + elif img.ndim == 3: + H, W, C = img.shape + H_r, W_r = H % scale, W % scale + img = img[:H - H_r, :W - W_r, :] + else: + raise ValueError('Wrong img ndim: [{:d}].'.format(img.ndim)) + return img + + +def shave(img_in, border=0): + # img_in: Numpy, HWC or HW + img = np.copy(img_in) + h, w = img.shape[:2] + img = img[border:h-border, border:w-border] + return img + + +''' +# -------------------------------------------- +# image processing process on numpy image +# channel_convert(in_c, tar_type, img_list): +# rgb2ycbcr(img, only_y=True): +# bgr2ycbcr(img, only_y=True): +# ycbcr2rgb(img): +# -------------------------------------------- +''' + + +def rgb2ycbcr(img, only_y=True): + '''same as matlab rgb2ycbcr + only_y: only return Y channel + Input: + uint8, [0, 255] + float, [0, 1] + ''' + in_img_type = img.dtype + img.astype(np.float32) + if in_img_type != np.uint8: + img *= 255. + # convert + if only_y: + rlt = np.dot(img, [65.481, 128.553, 24.966]) / 255.0 + 16.0 + else: + rlt = np.matmul(img, [[65.481, -37.797, 112.0], [128.553, -74.203, -93.786], + [24.966, 112.0, -18.214]]) / 255.0 + [16, 128, 128] + if in_img_type == np.uint8: + rlt = rlt.round() + else: + rlt /= 255. + return rlt.astype(in_img_type) + + +def ycbcr2rgb(img): + '''same as matlab ycbcr2rgb + Input: + uint8, [0, 255] + float, [0, 1] + ''' + in_img_type = img.dtype + img.astype(np.float32) + if in_img_type != np.uint8: + img *= 255. + # convert + rlt = np.matmul(img, [[0.00456621, 0.00456621, 0.00456621], [0, -0.00153632, 0.00791071], + [0.00625893, -0.00318811, 0]]) * 255.0 + [-222.921, 135.576, -276.836] + if in_img_type == np.uint8: + rlt = rlt.round() + else: + rlt /= 255. + return rlt.astype(in_img_type) + + +def bgr2ycbcr(img, only_y=True): + '''bgr version of rgb2ycbcr + only_y: only return Y channel + Input: + uint8, [0, 255] + float, [0, 1] + ''' + in_img_type = img.dtype + img.astype(np.float32) + if in_img_type != np.uint8: + img *= 255. + # convert + if only_y: + rlt = np.dot(img, [24.966, 128.553, 65.481]) / 255.0 + 16.0 + else: + rlt = np.matmul(img, [[24.966, 112.0, -18.214], [128.553, -74.203, -93.786], + [65.481, -37.797, 112.0]]) / 255.0 + [16, 128, 128] + if in_img_type == np.uint8: + rlt = rlt.round() + else: + rlt /= 255. + return rlt.astype(in_img_type) + + +def channel_convert(in_c, tar_type, img_list): + # conversion among BGR, gray and y + if in_c == 3 and tar_type == 'gray': # BGR to gray + gray_list = [cv2.cvtColor(img, cv2.COLOR_BGR2GRAY) for img in img_list] + return [np.expand_dims(img, axis=2) for img in gray_list] + elif in_c == 3 and tar_type == 'y': # BGR to y + y_list = [bgr2ycbcr(img, only_y=True) for img in img_list] + return [np.expand_dims(img, axis=2) for img in y_list] + elif in_c == 1 and tar_type == 'RGB': # gray/y to BGR + return [cv2.cvtColor(img, cv2.COLOR_GRAY2BGR) for img in img_list] + else: + return img_list + + +''' +# -------------------------------------------- +# metric, PSNR and SSIM +# -------------------------------------------- +''' + + +# -------------------------------------------- +# PSNR +# -------------------------------------------- +def calculate_psnr(img1, img2, border=0): + # img1 and img2 have range [0, 255] + #img1 = img1.squeeze() + #img2 = img2.squeeze() + if not img1.shape == img2.shape: + raise ValueError('Input images must have the same dimensions.') + h, w = img1.shape[:2] + img1 = img1[border:h-border, border:w-border] + img2 = img2[border:h-border, border:w-border] + + img1 = img1.astype(np.float64) + img2 = img2.astype(np.float64) + mse = np.mean((img1 - img2)**2) + if mse == 0: + return float('inf') + return 20 * math.log10(255.0 / math.sqrt(mse)) + + +# -------------------------------------------- +# SSIM +# -------------------------------------------- +def calculate_ssim(img1, img2, border=0): + '''calculate SSIM + the same outputs as MATLAB's + img1, img2: [0, 255] + ''' + #img1 = img1.squeeze() + #img2 = img2.squeeze() + if not img1.shape == img2.shape: + raise ValueError('Input images must have the same dimensions.') + h, w = img1.shape[:2] + img1 = img1[border:h-border, border:w-border] + img2 = img2[border:h-border, border:w-border] + + if img1.ndim == 2: + return ssim(img1, img2) + elif img1.ndim == 3: + if img1.shape[2] == 3: + ssims = [] + for i in range(3): + ssims.append(ssim(img1[:,:,i], img2[:,:,i])) + return np.array(ssims).mean() + elif img1.shape[2] == 1: + return ssim(np.squeeze(img1), np.squeeze(img2)) + else: + raise ValueError('Wrong input image dimensions.') + + +def ssim(img1, img2): + C1 = (0.01 * 255)**2 + C2 = (0.03 * 255)**2 + + img1 = img1.astype(np.float64) + img2 = img2.astype(np.float64) + kernel = cv2.getGaussianKernel(11, 1.5) + window = np.outer(kernel, kernel.transpose()) + + mu1 = cv2.filter2D(img1, -1, window)[5:-5, 5:-5] # valid + mu2 = cv2.filter2D(img2, -1, window)[5:-5, 5:-5] + mu1_sq = mu1**2 + mu2_sq = mu2**2 + mu1_mu2 = mu1 * mu2 + sigma1_sq = cv2.filter2D(img1**2, -1, window)[5:-5, 5:-5] - mu1_sq + sigma2_sq = cv2.filter2D(img2**2, -1, window)[5:-5, 5:-5] - mu2_sq + sigma12 = cv2.filter2D(img1 * img2, -1, window)[5:-5, 5:-5] - mu1_mu2 + + ssim_map = ((2 * mu1_mu2 + C1) * (2 * sigma12 + C2)) / ((mu1_sq + mu2_sq + C1) * + (sigma1_sq + sigma2_sq + C2)) + return ssim_map.mean() + + +''' +# -------------------------------------------- +# matlab's bicubic imresize (numpy and torch) [0, 1] +# -------------------------------------------- +''' + + +# matlab 'imresize' function, now only support 'bicubic' +def cubic(x): + absx = torch.abs(x) + absx2 = absx**2 + absx3 = absx**3 + return (1.5*absx3 - 2.5*absx2 + 1) * ((absx <= 1).type_as(absx)) + \ + (-0.5*absx3 + 2.5*absx2 - 4*absx + 2) * (((absx > 1)*(absx <= 2)).type_as(absx)) + + +def calculate_weights_indices(in_length, out_length, scale, kernel, kernel_width, antialiasing): + if (scale < 1) and (antialiasing): + # Use a modified kernel to simultaneously interpolate and antialias- larger kernel width + kernel_width = kernel_width / scale + + # Output-space coordinates + x = torch.linspace(1, out_length, out_length) + + # Input-space coordinates. Calculate the inverse mapping such that 0.5 + # in output space maps to 0.5 in input space, and 0.5+scale in output + # space maps to 1.5 in input space. + u = x / scale + 0.5 * (1 - 1 / scale) + + # What is the left-most pixel that can be involved in the computation? + left = torch.floor(u - kernel_width / 2) + + # What is the maximum number of pixels that can be involved in the + # computation? Note: it's OK to use an extra pixel here; if the + # corresponding weights are all zero, it will be eliminated at the end + # of this function. + P = math.ceil(kernel_width) + 2 + + # The indices of the input pixels involved in computing the k-th output + # pixel are in row k of the indices matrix. + indices = left.view(out_length, 1).expand(out_length, P) + torch.linspace(0, P - 1, P).view( + 1, P).expand(out_length, P) + + # The weights used to compute the k-th output pixel are in row k of the + # weights matrix. + distance_to_center = u.view(out_length, 1).expand(out_length, P) - indices + # apply cubic kernel + if (scale < 1) and (antialiasing): + weights = scale * cubic(distance_to_center * scale) + else: + weights = cubic(distance_to_center) + # Normalize the weights matrix so that each row sums to 1. + weights_sum = torch.sum(weights, 1).view(out_length, 1) + weights = weights / weights_sum.expand(out_length, P) + + # If a column in weights is all zero, get rid of it. only consider the first and last column. + weights_zero_tmp = torch.sum((weights == 0), 0) + if not math.isclose(weights_zero_tmp[0], 0, rel_tol=1e-6): + indices = indices.narrow(1, 1, P - 2) + weights = weights.narrow(1, 1, P - 2) + if not math.isclose(weights_zero_tmp[-1], 0, rel_tol=1e-6): + indices = indices.narrow(1, 0, P - 2) + weights = weights.narrow(1, 0, P - 2) + weights = weights.contiguous() + indices = indices.contiguous() + sym_len_s = -indices.min() + 1 + sym_len_e = indices.max() - in_length + indices = indices + sym_len_s - 1 + return weights, indices, int(sym_len_s), int(sym_len_e) + + +# -------------------------------------------- +# imresize for tensor image [0, 1] +# -------------------------------------------- +def imresize(img, scale, antialiasing=True): + # Now the scale should be the same for H and W + # input: img: pytorch tensor, CHW or HW [0,1] + # output: CHW or HW [0,1] w/o round + need_squeeze = True if img.dim() == 2 else False + if need_squeeze: + img.unsqueeze_(0) + in_C, in_H, in_W = img.size() + out_C, out_H, out_W = in_C, math.ceil(in_H * scale), math.ceil(in_W * scale) + kernel_width = 4 + kernel = 'cubic' + + # Return the desired dimension order for performing the resize. The + # strategy is to perform the resize first along the dimension with the + # smallest scale factor. + # Now we do not support this. + + # get weights and indices + weights_H, indices_H, sym_len_Hs, sym_len_He = calculate_weights_indices( + in_H, out_H, scale, kernel, kernel_width, antialiasing) + weights_W, indices_W, sym_len_Ws, sym_len_We = calculate_weights_indices( + in_W, out_W, scale, kernel, kernel_width, antialiasing) + # process H dimension + # symmetric copying + img_aug = torch.FloatTensor(in_C, in_H + sym_len_Hs + sym_len_He, in_W) + img_aug.narrow(1, sym_len_Hs, in_H).copy_(img) + + sym_patch = img[:, :sym_len_Hs, :] + inv_idx = torch.arange(sym_patch.size(1) - 1, -1, -1).long() + sym_patch_inv = sym_patch.index_select(1, inv_idx) + img_aug.narrow(1, 0, sym_len_Hs).copy_(sym_patch_inv) + + sym_patch = img[:, -sym_len_He:, :] + inv_idx = torch.arange(sym_patch.size(1) - 1, -1, -1).long() + sym_patch_inv = sym_patch.index_select(1, inv_idx) + img_aug.narrow(1, sym_len_Hs + in_H, sym_len_He).copy_(sym_patch_inv) + + out_1 = torch.FloatTensor(in_C, out_H, in_W) + kernel_width = weights_H.size(1) + for i in range(out_H): + idx = int(indices_H[i][0]) + for j in range(out_C): + out_1[j, i, :] = img_aug[j, idx:idx + kernel_width, :].transpose(0, 1).mv(weights_H[i]) + + # process W dimension + # symmetric copying + out_1_aug = torch.FloatTensor(in_C, out_H, in_W + sym_len_Ws + sym_len_We) + out_1_aug.narrow(2, sym_len_Ws, in_W).copy_(out_1) + + sym_patch = out_1[:, :, :sym_len_Ws] + inv_idx = torch.arange(sym_patch.size(2) - 1, -1, -1).long() + sym_patch_inv = sym_patch.index_select(2, inv_idx) + out_1_aug.narrow(2, 0, sym_len_Ws).copy_(sym_patch_inv) + + sym_patch = out_1[:, :, -sym_len_We:] + inv_idx = torch.arange(sym_patch.size(2) - 1, -1, -1).long() + sym_patch_inv = sym_patch.index_select(2, inv_idx) + out_1_aug.narrow(2, sym_len_Ws + in_W, sym_len_We).copy_(sym_patch_inv) + + out_2 = torch.FloatTensor(in_C, out_H, out_W) + kernel_width = weights_W.size(1) + for i in range(out_W): + idx = int(indices_W[i][0]) + for j in range(out_C): + out_2[j, :, i] = out_1_aug[j, :, idx:idx + kernel_width].mv(weights_W[i]) + if need_squeeze: + out_2.squeeze_() + return out_2 + + +# -------------------------------------------- +# imresize for numpy image [0, 1] +# -------------------------------------------- +def imresize_np(img, scale, antialiasing=True): + # Now the scale should be the same for H and W + # input: img: Numpy, HWC or HW [0,1] + # output: HWC or HW [0,1] w/o round + img = torch.from_numpy(img) + need_squeeze = True if img.dim() == 2 else False + if need_squeeze: + img.unsqueeze_(2) + + in_H, in_W, in_C = img.size() + out_C, out_H, out_W = in_C, math.ceil(in_H * scale), math.ceil(in_W * scale) + kernel_width = 4 + kernel = 'cubic' + + # Return the desired dimension order for performing the resize. The + # strategy is to perform the resize first along the dimension with the + # smallest scale factor. + # Now we do not support this. + + # get weights and indices + weights_H, indices_H, sym_len_Hs, sym_len_He = calculate_weights_indices( + in_H, out_H, scale, kernel, kernel_width, antialiasing) + weights_W, indices_W, sym_len_Ws, sym_len_We = calculate_weights_indices( + in_W, out_W, scale, kernel, kernel_width, antialiasing) + # process H dimension + # symmetric copying + img_aug = torch.FloatTensor(in_H + sym_len_Hs + sym_len_He, in_W, in_C) + img_aug.narrow(0, sym_len_Hs, in_H).copy_(img) + + sym_patch = img[:sym_len_Hs, :, :] + inv_idx = torch.arange(sym_patch.size(0) - 1, -1, -1).long() + sym_patch_inv = sym_patch.index_select(0, inv_idx) + img_aug.narrow(0, 0, sym_len_Hs).copy_(sym_patch_inv) + + sym_patch = img[-sym_len_He:, :, :] + inv_idx = torch.arange(sym_patch.size(0) - 1, -1, -1).long() + sym_patch_inv = sym_patch.index_select(0, inv_idx) + img_aug.narrow(0, sym_len_Hs + in_H, sym_len_He).copy_(sym_patch_inv) + + out_1 = torch.FloatTensor(out_H, in_W, in_C) + kernel_width = weights_H.size(1) + for i in range(out_H): + idx = int(indices_H[i][0]) + for j in range(out_C): + out_1[i, :, j] = img_aug[idx:idx + kernel_width, :, j].transpose(0, 1).mv(weights_H[i]) + + # process W dimension + # symmetric copying + out_1_aug = torch.FloatTensor(out_H, in_W + sym_len_Ws + sym_len_We, in_C) + out_1_aug.narrow(1, sym_len_Ws, in_W).copy_(out_1) + + sym_patch = out_1[:, :sym_len_Ws, :] + inv_idx = torch.arange(sym_patch.size(1) - 1, -1, -1).long() + sym_patch_inv = sym_patch.index_select(1, inv_idx) + out_1_aug.narrow(1, 0, sym_len_Ws).copy_(sym_patch_inv) + + sym_patch = out_1[:, -sym_len_We:, :] + inv_idx = torch.arange(sym_patch.size(1) - 1, -1, -1).long() + sym_patch_inv = sym_patch.index_select(1, inv_idx) + out_1_aug.narrow(1, sym_len_Ws + in_W, sym_len_We).copy_(sym_patch_inv) + + out_2 = torch.FloatTensor(out_H, out_W, in_C) + kernel_width = weights_W.size(1) + for i in range(out_W): + idx = int(indices_W[i][0]) + for j in range(out_C): + out_2[:, i, j] = out_1_aug[:, idx:idx + kernel_width, j].mv(weights_W[i]) + if need_squeeze: + out_2.squeeze_() + + return out_2.numpy() + + +if __name__ == '__main__': + print('---') +# img = imread_uint('test.bmp', 3) +# img = uint2single(img) +# img_bicubic = imresize_np(img, 1/4) \ No newline at end of file diff --git a/ldm/modules/karlo/__init__.py b/ldm/modules/karlo/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/ldm/modules/karlo/diffusers_pipeline.py b/ldm/modules/karlo/diffusers_pipeline.py new file mode 100644 index 0000000000000000000000000000000000000000..07f72b35a6e05390430880f190ed425985a4a5ef --- /dev/null +++ b/ldm/modules/karlo/diffusers_pipeline.py @@ -0,0 +1,512 @@ +# Copyright 2022 Kakao Brain and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from typing import List, Optional, Tuple, Union + +import torch +from torch.nn import functional as F + +from transformers import CLIPTextModelWithProjection, CLIPTokenizer +from transformers.models.clip.modeling_clip import CLIPTextModelOutput + +from ...models import PriorTransformer, UNet2DConditionModel, UNet2DModel +from ...pipelines import DiffusionPipeline, ImagePipelineOutput +from ...schedulers import UnCLIPScheduler +from ...utils import is_accelerate_available, logging, randn_tensor +from .text_proj import UnCLIPTextProjModel + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +class UnCLIPPipeline(DiffusionPipeline): + """ + Pipeline for text-to-image generation using unCLIP + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.) + Args: + text_encoder ([`CLIPTextModelWithProjection`]): + Frozen text-encoder. + tokenizer (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + prior ([`PriorTransformer`]): + The canonincal unCLIP prior to approximate the image embedding from the text embedding. + text_proj ([`UnCLIPTextProjModel`]): + Utility class to prepare and combine the embeddings before they are passed to the decoder. + decoder ([`UNet2DConditionModel`]): + The decoder to invert the image embedding into an image. + super_res_first ([`UNet2DModel`]): + Super resolution unet. Used in all but the last step of the super resolution diffusion process. + super_res_last ([`UNet2DModel`]): + Super resolution unet. Used in the last step of the super resolution diffusion process. + prior_scheduler ([`UnCLIPScheduler`]): + Scheduler used in the prior denoising process. Just a modified DDPMScheduler. + decoder_scheduler ([`UnCLIPScheduler`]): + Scheduler used in the decoder denoising process. Just a modified DDPMScheduler. + super_res_scheduler ([`UnCLIPScheduler`]): + Scheduler used in the super resolution denoising process. Just a modified DDPMScheduler. + """ + + prior: PriorTransformer + decoder: UNet2DConditionModel + text_proj: UnCLIPTextProjModel + text_encoder: CLIPTextModelWithProjection + tokenizer: CLIPTokenizer + super_res_first: UNet2DModel + super_res_last: UNet2DModel + + prior_scheduler: UnCLIPScheduler + decoder_scheduler: UnCLIPScheduler + super_res_scheduler: UnCLIPScheduler + + def __init__( + self, + prior: PriorTransformer, + decoder: UNet2DConditionModel, + text_encoder: CLIPTextModelWithProjection, + tokenizer: CLIPTokenizer, + text_proj: UnCLIPTextProjModel, + super_res_first: UNet2DModel, + super_res_last: UNet2DModel, + prior_scheduler: UnCLIPScheduler, + decoder_scheduler: UnCLIPScheduler, + super_res_scheduler: UnCLIPScheduler, + ): + super().__init__() + + self.register_modules( + prior=prior, + decoder=decoder, + text_encoder=text_encoder, + tokenizer=tokenizer, + text_proj=text_proj, + super_res_first=super_res_first, + super_res_last=super_res_last, + prior_scheduler=prior_scheduler, + decoder_scheduler=decoder_scheduler, + super_res_scheduler=super_res_scheduler, + ) + + def prepare_latents(self, shape, dtype, device, generator, latents, scheduler): + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + if latents.shape != shape: + raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}") + latents = latents.to(device) + + latents = latents * scheduler.init_noise_sigma + return latents + + def _encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + text_model_output: Optional[Union[CLIPTextModelOutput, Tuple]] = None, + text_attention_mask: Optional[torch.Tensor] = None, + ): + if text_model_output is None: + batch_size = len(prompt) if isinstance(prompt, list) else 1 + # get prompt text embeddings + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + text_mask = text_inputs.attention_mask.bool().to(device) + + if text_input_ids.shape[-1] > self.tokenizer.model_max_length: + removed_text = self.tokenizer.batch_decode(text_input_ids[:, self.tokenizer.model_max_length :]) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + text_input_ids = text_input_ids[:, : self.tokenizer.model_max_length] + + text_encoder_output = self.text_encoder(text_input_ids.to(device)) + + text_embeddings = text_encoder_output.text_embeds + text_encoder_hidden_states = text_encoder_output.last_hidden_state + + else: + batch_size = text_model_output[0].shape[0] + text_embeddings, text_encoder_hidden_states = text_model_output[0], text_model_output[1] + text_mask = text_attention_mask + + text_embeddings = text_embeddings.repeat_interleave(num_images_per_prompt, dim=0) + text_encoder_hidden_states = text_encoder_hidden_states.repeat_interleave(num_images_per_prompt, dim=0) + text_mask = text_mask.repeat_interleave(num_images_per_prompt, dim=0) + + if do_classifier_free_guidance: + uncond_tokens = [""] * batch_size + + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + uncond_text_mask = uncond_input.attention_mask.bool().to(device) + uncond_embeddings_text_encoder_output = self.text_encoder(uncond_input.input_ids.to(device)) + + uncond_embeddings = uncond_embeddings_text_encoder_output.text_embeds + uncond_text_encoder_hidden_states = uncond_embeddings_text_encoder_output.last_hidden_state + + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + + seq_len = uncond_embeddings.shape[1] + uncond_embeddings = uncond_embeddings.repeat(1, num_images_per_prompt) + uncond_embeddings = uncond_embeddings.view(batch_size * num_images_per_prompt, seq_len) + + seq_len = uncond_text_encoder_hidden_states.shape[1] + uncond_text_encoder_hidden_states = uncond_text_encoder_hidden_states.repeat(1, num_images_per_prompt, 1) + uncond_text_encoder_hidden_states = uncond_text_encoder_hidden_states.view( + batch_size * num_images_per_prompt, seq_len, -1 + ) + uncond_text_mask = uncond_text_mask.repeat_interleave(num_images_per_prompt, dim=0) + + # done duplicates + + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + text_embeddings = torch.cat([uncond_embeddings, text_embeddings]) + text_encoder_hidden_states = torch.cat([uncond_text_encoder_hidden_states, text_encoder_hidden_states]) + + text_mask = torch.cat([uncond_text_mask, text_mask]) + + return text_embeddings, text_encoder_hidden_states, text_mask + + def enable_sequential_cpu_offload(self, gpu_id=0): + r""" + Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, the pipeline's + models have their state dicts saved to CPU and then are moved to a `torch.device('meta') and loaded to GPU only + when their specific submodule has its `forward` method called. + """ + if is_accelerate_available(): + from accelerate import cpu_offload + else: + raise ImportError("Please install accelerate via `pip install accelerate`") + + device = torch.device(f"cuda:{gpu_id}") + + # TODO: self.prior.post_process_latents is not covered by the offload hooks, so it fails if added to the list + models = [ + self.decoder, + self.text_proj, + self.text_encoder, + self.super_res_first, + self.super_res_last, + ] + for cpu_offloaded_model in models: + if cpu_offloaded_model is not None: + cpu_offload(cpu_offloaded_model, device) + + @property + def _execution_device(self): + r""" + Returns the device on which the pipeline's models will be executed. After calling + `pipeline.enable_sequential_cpu_offload()` the execution device can only be inferred from Accelerate's module + hooks. + """ + if self.device != torch.device("meta") or not hasattr(self.decoder, "_hf_hook"): + return self.device + for module in self.decoder.modules(): + if ( + hasattr(module, "_hf_hook") + and hasattr(module._hf_hook, "execution_device") + and module._hf_hook.execution_device is not None + ): + return torch.device(module._hf_hook.execution_device) + return self.device + + @torch.no_grad() + def __call__( + self, + prompt: Optional[Union[str, List[str]]] = None, + num_images_per_prompt: int = 1, + prior_num_inference_steps: int = 25, + decoder_num_inference_steps: int = 25, + super_res_num_inference_steps: int = 7, + generator: Optional[torch.Generator] = None, + prior_latents: Optional[torch.FloatTensor] = None, + decoder_latents: Optional[torch.FloatTensor] = None, + super_res_latents: Optional[torch.FloatTensor] = None, + text_model_output: Optional[Union[CLIPTextModelOutput, Tuple]] = None, + text_attention_mask: Optional[torch.Tensor] = None, + prior_guidance_scale: float = 4.0, + decoder_guidance_scale: float = 8.0, + output_type: Optional[str] = "pil", + return_dict: bool = True, + ): + """ + Function invoked when calling the pipeline for generation. + Args: + prompt (`str` or `List[str]`): + The prompt or prompts to guide the image generation. This can only be left undefined if + `text_model_output` and `text_attention_mask` is passed. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + prior_num_inference_steps (`int`, *optional*, defaults to 25): + The number of denoising steps for the prior. More denoising steps usually lead to a higher quality + image at the expense of slower inference. + decoder_num_inference_steps (`int`, *optional*, defaults to 25): + The number of denoising steps for the decoder. More denoising steps usually lead to a higher quality + image at the expense of slower inference. + super_res_num_inference_steps (`int`, *optional*, defaults to 7): + The number of denoising steps for super resolution. More denoising steps usually lead to a higher + quality image at the expense of slower inference. + generator (`torch.Generator`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + prior_latents (`torch.FloatTensor` of shape (batch size, embeddings dimension), *optional*): + Pre-generated noisy latents to be used as inputs for the prior. + decoder_latents (`torch.FloatTensor` of shape (batch size, channels, height, width), *optional*): + Pre-generated noisy latents to be used as inputs for the decoder. + super_res_latents (`torch.FloatTensor` of shape (batch size, channels, super res height, super res width), *optional*): + Pre-generated noisy latents to be used as inputs for the decoder. + prior_guidance_scale (`float`, *optional*, defaults to 4.0): + Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598). + `guidance_scale` is defined as `w` of equation 2. of [Imagen + Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale > + 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`, + usually at the expense of lower image quality. + decoder_guidance_scale (`float`, *optional*, defaults to 4.0): + Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598). + `guidance_scale` is defined as `w` of equation 2. of [Imagen + Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale > + 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`, + usually at the expense of lower image quality. + text_model_output (`CLIPTextModelOutput`, *optional*): + Pre-defined CLIPTextModel outputs that can be derived from the text encoder. Pre-defined text outputs + can be passed for tasks like text embedding interpolations. Make sure to also pass + `text_attention_mask` in this case. `prompt` can the be left to `None`. + text_attention_mask (`torch.Tensor`, *optional*): + Pre-defined CLIP text attention mask that can be derived from the tokenizer. Pre-defined text attention + masks are necessary when passing `text_model_output`. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.ImagePipelineOutput`] instead of a plain tuple. + """ + if prompt is not None: + if isinstance(prompt, str): + batch_size = 1 + elif isinstance(prompt, list): + batch_size = len(prompt) + else: + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + else: + batch_size = text_model_output[0].shape[0] + + device = self._execution_device + + batch_size = batch_size * num_images_per_prompt + + do_classifier_free_guidance = prior_guidance_scale > 1.0 or decoder_guidance_scale > 1.0 + + text_embeddings, text_encoder_hidden_states, text_mask = self._encode_prompt( + prompt, device, num_images_per_prompt, do_classifier_free_guidance, text_model_output, text_attention_mask + ) + + # prior + + self.prior_scheduler.set_timesteps(prior_num_inference_steps, device=device) + prior_timesteps_tensor = self.prior_scheduler.timesteps + + embedding_dim = self.prior.config.embedding_dim + + prior_latents = self.prepare_latents( + (batch_size, embedding_dim), + text_embeddings.dtype, + device, + generator, + prior_latents, + self.prior_scheduler, + ) + + for i, t in enumerate(self.progress_bar(prior_timesteps_tensor)): + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([prior_latents] * 2) if do_classifier_free_guidance else prior_latents + + predicted_image_embedding = self.prior( + latent_model_input, + timestep=t, + proj_embedding=text_embeddings, + encoder_hidden_states=text_encoder_hidden_states, + attention_mask=text_mask, + ).predicted_image_embedding + + if do_classifier_free_guidance: + predicted_image_embedding_uncond, predicted_image_embedding_text = predicted_image_embedding.chunk(2) + predicted_image_embedding = predicted_image_embedding_uncond + prior_guidance_scale * ( + predicted_image_embedding_text - predicted_image_embedding_uncond + ) + + if i + 1 == prior_timesteps_tensor.shape[0]: + prev_timestep = None + else: + prev_timestep = prior_timesteps_tensor[i + 1] + + prior_latents = self.prior_scheduler.step( + predicted_image_embedding, + timestep=t, + sample=prior_latents, + generator=generator, + prev_timestep=prev_timestep, + ).prev_sample + + prior_latents = self.prior.post_process_latents(prior_latents) + + image_embeddings = prior_latents + + # done prior + + # decoder + + text_encoder_hidden_states, additive_clip_time_embeddings = self.text_proj( + image_embeddings=image_embeddings, + text_embeddings=text_embeddings, + text_encoder_hidden_states=text_encoder_hidden_states, + do_classifier_free_guidance=do_classifier_free_guidance, + ) + + decoder_text_mask = F.pad(text_mask, (self.text_proj.clip_extra_context_tokens, 0), value=1) + + self.decoder_scheduler.set_timesteps(decoder_num_inference_steps, device=device) + decoder_timesteps_tensor = self.decoder_scheduler.timesteps + + num_channels_latents = self.decoder.in_channels + height = self.decoder.sample_size + width = self.decoder.sample_size + + decoder_latents = self.prepare_latents( + (batch_size, num_channels_latents, height, width), + text_encoder_hidden_states.dtype, + device, + generator, + decoder_latents, + self.decoder_scheduler, + ) + + for i, t in enumerate(self.progress_bar(decoder_timesteps_tensor)): + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([decoder_latents] * 2) if do_classifier_free_guidance else decoder_latents + + noise_pred = self.decoder( + sample=latent_model_input, + timestep=t, + encoder_hidden_states=text_encoder_hidden_states, + class_labels=additive_clip_time_embeddings, + attention_mask=decoder_text_mask, + ).sample + + if do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred_uncond, _ = noise_pred_uncond.split(latent_model_input.shape[1], dim=1) + noise_pred_text, predicted_variance = noise_pred_text.split(latent_model_input.shape[1], dim=1) + noise_pred = noise_pred_uncond + decoder_guidance_scale * (noise_pred_text - noise_pred_uncond) + noise_pred = torch.cat([noise_pred, predicted_variance], dim=1) + + if i + 1 == decoder_timesteps_tensor.shape[0]: + prev_timestep = None + else: + prev_timestep = decoder_timesteps_tensor[i + 1] + + # compute the previous noisy sample x_t -> x_t-1 + decoder_latents = self.decoder_scheduler.step( + noise_pred, t, decoder_latents, prev_timestep=prev_timestep, generator=generator + ).prev_sample + + decoder_latents = decoder_latents.clamp(-1, 1) + + image_small = decoder_latents + + # done decoder + + # super res + + self.super_res_scheduler.set_timesteps(super_res_num_inference_steps, device=device) + super_res_timesteps_tensor = self.super_res_scheduler.timesteps + + channels = self.super_res_first.in_channels // 2 + height = self.super_res_first.sample_size + width = self.super_res_first.sample_size + + super_res_latents = self.prepare_latents( + (batch_size, channels, height, width), + image_small.dtype, + device, + generator, + super_res_latents, + self.super_res_scheduler, + ) + + interpolate_antialias = {} + if "antialias" in inspect.signature(F.interpolate).parameters: + interpolate_antialias["antialias"] = True + + image_upscaled = F.interpolate( + image_small, size=[height, width], mode="bicubic", align_corners=False, **interpolate_antialias + ) + + for i, t in enumerate(self.progress_bar(super_res_timesteps_tensor)): + # no classifier free guidance + + if i == super_res_timesteps_tensor.shape[0] - 1: + unet = self.super_res_last + else: + unet = self.super_res_first + + latent_model_input = torch.cat([super_res_latents, image_upscaled], dim=1) + + noise_pred = unet( + sample=latent_model_input, + timestep=t, + ).sample + + if i + 1 == super_res_timesteps_tensor.shape[0]: + prev_timestep = None + else: + prev_timestep = super_res_timesteps_tensor[i + 1] + + # compute the previous noisy sample x_t -> x_t-1 + super_res_latents = self.super_res_scheduler.step( + noise_pred, t, super_res_latents, prev_timestep=prev_timestep, generator=generator + ).prev_sample + + image = super_res_latents + # done super res + + # post processing + + image = image * 0.5 + 0.5 + image = image.clamp(0, 1) + image = image.cpu().permute(0, 2, 3, 1).float().numpy() + + if output_type == "pil": + image = self.numpy_to_pil(image) + + if not return_dict: + return (image,) + + return ImagePipelineOutput(images=image) \ No newline at end of file diff --git a/ldm/modules/karlo/kakao/__init__.py b/ldm/modules/karlo/kakao/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/ldm/modules/karlo/kakao/models/__init__.py b/ldm/modules/karlo/kakao/models/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/ldm/modules/karlo/kakao/models/clip.py b/ldm/modules/karlo/kakao/models/clip.py new file mode 100644 index 0000000000000000000000000000000000000000..961d81502a069ae29b6752478f048c979b27fb9b --- /dev/null +++ b/ldm/modules/karlo/kakao/models/clip.py @@ -0,0 +1,182 @@ +# ------------------------------------------------------------------------------------ +# Karlo-v1.0.alpha +# Copyright (c) 2022 KakaoBrain. All Rights Reserved. +# ------------------------------------------------------------------------------------ +# ------------------------------------------------------------------------------------ +# Adapted from OpenAI's CLIP (https://github.com/openai/CLIP/) +# ------------------------------------------------------------------------------------ + + +import torch +import torch.nn as nn +import torch.nn.functional as F +import clip + +from clip.model import CLIP, convert_weights +from clip.simple_tokenizer import SimpleTokenizer, default_bpe + + +"""===== Monkey-Patching original CLIP for JIT compile =====""" + + +class LayerNorm(nn.LayerNorm): + """Subclass torch's LayerNorm to handle fp16.""" + + def forward(self, x: torch.Tensor): + orig_type = x.dtype + ret = F.layer_norm( + x.type(torch.float32), + self.normalized_shape, + self.weight, + self.bias, + self.eps, + ) + return ret.type(orig_type) + + +clip.model.LayerNorm = LayerNorm +delattr(clip.model.CLIP, "forward") + +"""===== End of Monkey-Patching =====""" + + +class CustomizedCLIP(CLIP): + def __init__(self, *args, **kwargs): + super().__init__(*args, **kwargs) + + @torch.jit.export + def encode_image(self, image): + return self.visual(image) + + @torch.jit.export + def encode_text(self, text): + # re-define this function to return unpooled text features + + x = self.token_embedding(text).type(self.dtype) # [batch_size, n_ctx, d_model] + + x = x + self.positional_embedding.type(self.dtype) + x = x.permute(1, 0, 2) # NLD -> LND + x = self.transformer(x) + x = x.permute(1, 0, 2) # LND -> NLD + x = self.ln_final(x).type(self.dtype) + + x_seq = x + # x.shape = [batch_size, n_ctx, transformer.width] + # take features from the eot embedding (eot_token is the highest number in each sequence) + x_out = x[torch.arange(x.shape[0]), text.argmax(dim=-1)] @ self.text_projection + + return x_out, x_seq + + @torch.jit.ignore + def forward(self, image, text): + super().forward(image, text) + + @classmethod + def load_from_checkpoint(cls, ckpt_path: str): + state_dict = torch.load(ckpt_path, map_location="cpu").state_dict() + + vit = "visual.proj" in state_dict + if vit: + vision_width = state_dict["visual.conv1.weight"].shape[0] + vision_layers = len( + [ + k + for k in state_dict.keys() + if k.startswith("visual.") and k.endswith(".attn.in_proj_weight") + ] + ) + vision_patch_size = state_dict["visual.conv1.weight"].shape[-1] + grid_size = round( + (state_dict["visual.positional_embedding"].shape[0] - 1) ** 0.5 + ) + image_resolution = vision_patch_size * grid_size + else: + counts: list = [ + len( + set( + k.split(".")[2] + for k in state_dict + if k.startswith(f"visual.layer{b}") + ) + ) + for b in [1, 2, 3, 4] + ] + vision_layers = tuple(counts) + vision_width = state_dict["visual.layer1.0.conv1.weight"].shape[0] + output_width = round( + (state_dict["visual.attnpool.positional_embedding"].shape[0] - 1) ** 0.5 + ) + vision_patch_size = None + assert ( + output_width**2 + 1 + == state_dict["visual.attnpool.positional_embedding"].shape[0] + ) + image_resolution = output_width * 32 + + embed_dim = state_dict["text_projection"].shape[1] + context_length = state_dict["positional_embedding"].shape[0] + vocab_size = state_dict["token_embedding.weight"].shape[0] + transformer_width = state_dict["ln_final.weight"].shape[0] + transformer_heads = transformer_width // 64 + transformer_layers = len( + set( + k.split(".")[2] + for k in state_dict + if k.startswith("transformer.resblocks") + ) + ) + + model = cls( + embed_dim, + image_resolution, + vision_layers, + vision_width, + vision_patch_size, + context_length, + vocab_size, + transformer_width, + transformer_heads, + transformer_layers, + ) + + for key in ["input_resolution", "context_length", "vocab_size"]: + if key in state_dict: + del state_dict[key] + + convert_weights(model) + model.load_state_dict(state_dict) + model.eval() + model.float() + return model + + +class CustomizedTokenizer(SimpleTokenizer): + def __init__(self): + super().__init__(bpe_path=default_bpe()) + + self.sot_token = self.encoder["<|startoftext|>"] + self.eot_token = self.encoder["<|endoftext|>"] + + def padded_tokens_and_mask(self, texts, text_ctx): + assert isinstance(texts, list) and all( + isinstance(elem, str) for elem in texts + ), "texts should be a list of strings" + + all_tokens = [ + [self.sot_token] + self.encode(text) + [self.eot_token] for text in texts + ] + + mask = [ + [True] * min(text_ctx, len(tokens)) + + [False] * max(text_ctx - len(tokens), 0) + for tokens in all_tokens + ] + mask = torch.tensor(mask, dtype=torch.bool) + result = torch.zeros(len(all_tokens), text_ctx, dtype=torch.int) + for i, tokens in enumerate(all_tokens): + if len(tokens) > text_ctx: + tokens = tokens[:text_ctx] + tokens[-1] = self.eot_token + result[i, : len(tokens)] = torch.tensor(tokens) + + return result, mask diff --git a/ldm/modules/karlo/kakao/models/decoder_model.py b/ldm/modules/karlo/kakao/models/decoder_model.py new file mode 100644 index 0000000000000000000000000000000000000000..84e96c9b2f3f54561a5d950097d5d069ce4ce3a0 --- /dev/null +++ b/ldm/modules/karlo/kakao/models/decoder_model.py @@ -0,0 +1,193 @@ +# ------------------------------------------------------------------------------------ +# Karlo-v1.0.alpha +# Copyright (c) 2022 KakaoBrain. All Rights Reserved. +# ------------------------------------------------------------------------------------ + +import copy +import torch + +from ldm.modules.karlo.kakao.modules import create_gaussian_diffusion +from ldm.modules.karlo.kakao.modules.unet import PLMImUNet + + +class Text2ImProgressiveModel(torch.nn.Module): + """ + A decoder that generates 64x64px images based on the text prompt. + + :param config: yaml config to define the decoder. + :param tokenizer: tokenizer used in clip. + """ + + def __init__( + self, + config, + tokenizer, + ): + super().__init__() + + self._conf = config + self._model_conf = config.model.hparams + self._diffusion_kwargs = dict( + steps=config.diffusion.steps, + learn_sigma=config.diffusion.learn_sigma, + sigma_small=config.diffusion.sigma_small, + noise_schedule=config.diffusion.noise_schedule, + use_kl=config.diffusion.use_kl, + predict_xstart=config.diffusion.predict_xstart, + rescale_learned_sigmas=config.diffusion.rescale_learned_sigmas, + timestep_respacing=config.diffusion.timestep_respacing, + ) + self._tokenizer = tokenizer + + self.model = self.create_plm_dec_model() + + cf_token, cf_mask = self.set_cf_text_tensor() + self.register_buffer("cf_token", cf_token, persistent=False) + self.register_buffer("cf_mask", cf_mask, persistent=False) + + @classmethod + def load_from_checkpoint(cls, config, tokenizer, ckpt_path, strict: bool = True): + ckpt = torch.load(ckpt_path, map_location="cpu")["state_dict"] + + model = cls(config, tokenizer) + model.load_state_dict(ckpt, strict=strict) + return model + + def create_plm_dec_model(self): + image_size = self._model_conf.image_size + if self._model_conf.channel_mult == "": + if image_size == 256: + channel_mult = (1, 1, 2, 2, 4, 4) + elif image_size == 128: + channel_mult = (1, 1, 2, 3, 4) + elif image_size == 64: + channel_mult = (1, 2, 3, 4) + else: + raise ValueError(f"unsupported image size: {image_size}") + else: + channel_mult = tuple( + int(ch_mult) for ch_mult in self._model_conf.channel_mult.split(",") + ) + assert 2 ** (len(channel_mult) + 2) == image_size + + attention_ds = [] + for res in self._model_conf.attention_resolutions.split(","): + attention_ds.append(image_size // int(res)) + + return PLMImUNet( + text_ctx=self._model_conf.text_ctx, + xf_width=self._model_conf.xf_width, + in_channels=3, + model_channels=self._model_conf.num_channels, + out_channels=6 if self._model_conf.learn_sigma else 3, + num_res_blocks=self._model_conf.num_res_blocks, + attention_resolutions=tuple(attention_ds), + dropout=self._model_conf.dropout, + channel_mult=channel_mult, + num_heads=self._model_conf.num_heads, + num_head_channels=self._model_conf.num_head_channels, + num_heads_upsample=self._model_conf.num_heads_upsample, + use_scale_shift_norm=self._model_conf.use_scale_shift_norm, + resblock_updown=self._model_conf.resblock_updown, + clip_dim=self._model_conf.clip_dim, + clip_emb_mult=self._model_conf.clip_emb_mult, + clip_emb_type=self._model_conf.clip_emb_type, + clip_emb_drop=self._model_conf.clip_emb_drop, + ) + + def set_cf_text_tensor(self): + return self._tokenizer.padded_tokens_and_mask([""], self.model.text_ctx) + + def get_sample_fn(self, timestep_respacing): + use_ddim = timestep_respacing.startswith(("ddim", "fast")) + + diffusion_kwargs = copy.deepcopy(self._diffusion_kwargs) + diffusion_kwargs.update(timestep_respacing=timestep_respacing) + diffusion = create_gaussian_diffusion(**diffusion_kwargs) + sample_fn = ( + diffusion.ddim_sample_loop_progressive + if use_ddim + else diffusion.p_sample_loop_progressive + ) + + return sample_fn + + def forward( + self, + txt_feat, + txt_feat_seq, + tok, + mask, + img_feat=None, + cf_guidance_scales=None, + timestep_respacing=None, + ): + # cfg should be enabled in inference + assert cf_guidance_scales is not None and all(cf_guidance_scales > 0.0) + assert img_feat is not None + + bsz = txt_feat.shape[0] + img_sz = self._model_conf.image_size + + def guided_model_fn(x_t, ts, **kwargs): + half = x_t[: len(x_t) // 2] + combined = torch.cat([half, half], dim=0) + model_out = self.model(combined, ts, **kwargs) + eps, rest = model_out[:, :3], model_out[:, 3:] + cond_eps, uncond_eps = torch.split(eps, len(eps) // 2, dim=0) + half_eps = uncond_eps + cf_guidance_scales.view(-1, 1, 1, 1) * ( + cond_eps - uncond_eps + ) + eps = torch.cat([half_eps, half_eps], dim=0) + return torch.cat([eps, rest], dim=1) + + cf_feat = self.model.cf_param.unsqueeze(0) + cf_feat = cf_feat.expand(bsz // 2, -1) + feat = torch.cat([img_feat, cf_feat.to(txt_feat.device)], dim=0) + + cond = { + "y": feat, + "txt_feat": txt_feat, + "txt_feat_seq": txt_feat_seq, + "mask": mask, + } + sample_fn = self.get_sample_fn(timestep_respacing) + sample_outputs = sample_fn( + guided_model_fn, + (bsz, 3, img_sz, img_sz), + noise=None, + device=txt_feat.device, + clip_denoised=True, + model_kwargs=cond, + ) + + for out in sample_outputs: + sample = out["sample"] + yield sample if cf_guidance_scales is None else sample[ + : sample.shape[0] // 2 + ] + + +class Text2ImModel(Text2ImProgressiveModel): + def forward( + self, + txt_feat, + txt_feat_seq, + tok, + mask, + img_feat=None, + cf_guidance_scales=None, + timestep_respacing=None, + ): + last_out = None + for out in super().forward( + txt_feat, + txt_feat_seq, + tok, + mask, + img_feat, + cf_guidance_scales, + timestep_respacing, + ): + last_out = out + return last_out diff --git a/ldm/modules/karlo/kakao/models/prior_model.py b/ldm/modules/karlo/kakao/models/prior_model.py new file mode 100644 index 0000000000000000000000000000000000000000..03ef230d2ac5e7f785c138f710337980e6754022 --- /dev/null +++ b/ldm/modules/karlo/kakao/models/prior_model.py @@ -0,0 +1,138 @@ +# ------------------------------------------------------------------------------------ +# Karlo-v1.0.alpha +# Copyright (c) 2022 KakaoBrain. All Rights Reserved. +# ------------------------------------------------------------------------------------ + +import copy +import torch + +from ldm.modules.karlo.kakao.modules import create_gaussian_diffusion +from ldm.modules.karlo.kakao.modules.xf import PriorTransformer + + +class PriorDiffusionModel(torch.nn.Module): + """ + A prior that generates clip image feature based on the text prompt. + + :param config: yaml config to define the decoder. + :param tokenizer: tokenizer used in clip. + :param clip_mean: mean to normalize the clip image feature (zero-mean, unit variance). + :param clip_std: std to noramlize the clip image feature (zero-mean, unit variance). + """ + + def __init__(self, config, tokenizer, clip_mean, clip_std): + super().__init__() + + self._conf = config + self._model_conf = config.model.hparams + self._diffusion_kwargs = dict( + steps=config.diffusion.steps, + learn_sigma=config.diffusion.learn_sigma, + sigma_small=config.diffusion.sigma_small, + noise_schedule=config.diffusion.noise_schedule, + use_kl=config.diffusion.use_kl, + predict_xstart=config.diffusion.predict_xstart, + rescale_learned_sigmas=config.diffusion.rescale_learned_sigmas, + timestep_respacing=config.diffusion.timestep_respacing, + ) + self._tokenizer = tokenizer + + self.register_buffer("clip_mean", clip_mean[None, :], persistent=False) + self.register_buffer("clip_std", clip_std[None, :], persistent=False) + + causal_mask = self.get_causal_mask() + self.register_buffer("causal_mask", causal_mask, persistent=False) + + self.model = PriorTransformer( + text_ctx=self._model_conf.text_ctx, + xf_width=self._model_conf.xf_width, + xf_layers=self._model_conf.xf_layers, + xf_heads=self._model_conf.xf_heads, + xf_final_ln=self._model_conf.xf_final_ln, + clip_dim=self._model_conf.clip_dim, + ) + + cf_token, cf_mask = self.set_cf_text_tensor() + self.register_buffer("cf_token", cf_token, persistent=False) + self.register_buffer("cf_mask", cf_mask, persistent=False) + + @classmethod + def load_from_checkpoint( + cls, config, tokenizer, clip_mean, clip_std, ckpt_path, strict: bool = True + ): + ckpt = torch.load(ckpt_path, map_location="cpu")["state_dict"] + + model = cls(config, tokenizer, clip_mean, clip_std) + model.load_state_dict(ckpt, strict=strict) + return model + + def set_cf_text_tensor(self): + return self._tokenizer.padded_tokens_and_mask([""], self.model.text_ctx) + + def get_sample_fn(self, timestep_respacing): + use_ddim = timestep_respacing.startswith(("ddim", "fast")) + + diffusion_kwargs = copy.deepcopy(self._diffusion_kwargs) + diffusion_kwargs.update(timestep_respacing=timestep_respacing) + diffusion = create_gaussian_diffusion(**diffusion_kwargs) + sample_fn = diffusion.ddim_sample_loop if use_ddim else diffusion.p_sample_loop + + return sample_fn + + def get_causal_mask(self): + seq_len = self._model_conf.text_ctx + 4 + mask = torch.empty(seq_len, seq_len) + mask.fill_(float("-inf")) + mask.triu_(1) + mask = mask[None, ...] + return mask + + def forward( + self, + txt_feat, + txt_feat_seq, + mask, + cf_guidance_scales=None, + timestep_respacing=None, + denoised_fn=True, + ): + # cfg should be enabled in inference + assert cf_guidance_scales is not None and all(cf_guidance_scales > 0.0) + + bsz_ = txt_feat.shape[0] + bsz = bsz_ // 2 + + def guided_model_fn(x_t, ts, **kwargs): + half = x_t[: len(x_t) // 2] + combined = torch.cat([half, half], dim=0) + model_out = self.model(combined, ts, **kwargs) + eps, rest = ( + model_out[:, : int(x_t.shape[1])], + model_out[:, int(x_t.shape[1]) :], + ) + cond_eps, uncond_eps = torch.split(eps, len(eps) // 2, dim=0) + half_eps = uncond_eps + cf_guidance_scales.view(-1, 1) * ( + cond_eps - uncond_eps + ) + eps = torch.cat([half_eps, half_eps], dim=0) + return torch.cat([eps, rest], dim=1) + + cond = { + "text_emb": txt_feat, + "text_enc": txt_feat_seq, + "mask": mask, + "causal_mask": self.causal_mask, + } + sample_fn = self.get_sample_fn(timestep_respacing) + sample = sample_fn( + guided_model_fn, + (bsz_, self.model.clip_dim), + noise=None, + device=txt_feat.device, + clip_denoised=False, + denoised_fn=lambda x: torch.clamp(x, -10, 10), + model_kwargs=cond, + ) + sample = (sample * self.clip_std) + self.clip_mean + + return sample[:bsz] diff --git a/ldm/modules/karlo/kakao/models/sr_256_1k.py b/ldm/modules/karlo/kakao/models/sr_256_1k.py new file mode 100644 index 0000000000000000000000000000000000000000..1e874f6f1b8286d85456ae335316880cb5c15336 --- /dev/null +++ b/ldm/modules/karlo/kakao/models/sr_256_1k.py @@ -0,0 +1,10 @@ +# ------------------------------------------------------------------------------------ +# Karlo-v1.0.alpha +# Copyright (c) 2022 KakaoBrain. All Rights Reserved. +# ------------------------------------------------------------------------------------ + +from ldm.modules.karlo.kakao.models.sr_64_256 import SupRes64to256Progressive + + +class SupRes256to1kProgressive(SupRes64to256Progressive): + pass # no difference currently diff --git a/ldm/modules/karlo/kakao/models/sr_64_256.py b/ldm/modules/karlo/kakao/models/sr_64_256.py new file mode 100644 index 0000000000000000000000000000000000000000..32687afe38134c4eed06083fc9c345b1dbe958a9 --- /dev/null +++ b/ldm/modules/karlo/kakao/models/sr_64_256.py @@ -0,0 +1,88 @@ +# ------------------------------------------------------------------------------------ +# Karlo-v1.0.alpha +# Copyright (c) 2022 KakaoBrain. All Rights Reserved. +# ------------------------------------------------------------------------------------ + +import copy +import torch + +from ldm.modules.karlo.kakao.modules.unet import SuperResUNetModel +from ldm.modules.karlo.kakao.modules import create_gaussian_diffusion + + +class ImprovedSupRes64to256ProgressiveModel(torch.nn.Module): + """ + ImprovedSR model fine-tunes the pretrained DDPM-based SR model by using adversarial and perceptual losses. + In specific, the low-resolution sample is iteratively recovered by 6 steps with the frozen pretrained SR model. + In the following additional one step, a seperate fine-tuned model recovers high-frequency details. + This approach greatly improves the fidelity of images of 256x256px, even with small number of reverse steps. + """ + + def __init__(self, config): + super().__init__() + + self._config = config + self._diffusion_kwargs = dict( + steps=config.diffusion.steps, + learn_sigma=config.diffusion.learn_sigma, + sigma_small=config.diffusion.sigma_small, + noise_schedule=config.diffusion.noise_schedule, + use_kl=config.diffusion.use_kl, + predict_xstart=config.diffusion.predict_xstart, + rescale_learned_sigmas=config.diffusion.rescale_learned_sigmas, + ) + + self.model_first_steps = SuperResUNetModel( + in_channels=3, # auto-changed to 6 inside the model + model_channels=config.model.hparams.channels, + out_channels=3, + num_res_blocks=config.model.hparams.depth, + attention_resolutions=(), # no attention + dropout=config.model.hparams.dropout, + channel_mult=config.model.hparams.channels_multiple, + resblock_updown=True, + use_middle_attention=False, + ) + self.model_last_step = SuperResUNetModel( + in_channels=3, # auto-changed to 6 inside the model + model_channels=config.model.hparams.channels, + out_channels=3, + num_res_blocks=config.model.hparams.depth, + attention_resolutions=(), # no attention + dropout=config.model.hparams.dropout, + channel_mult=config.model.hparams.channels_multiple, + resblock_updown=True, + use_middle_attention=False, + ) + + @classmethod + def load_from_checkpoint(cls, config, ckpt_path, strict: bool = True): + ckpt = torch.load(ckpt_path, map_location="cpu")["state_dict"] + + model = cls(config) + model.load_state_dict(ckpt, strict=strict) + return model + + def get_sample_fn(self, timestep_respacing): + diffusion_kwargs = copy.deepcopy(self._diffusion_kwargs) + diffusion_kwargs.update(timestep_respacing=timestep_respacing) + diffusion = create_gaussian_diffusion(**diffusion_kwargs) + return diffusion.p_sample_loop_progressive_for_improved_sr + + def forward(self, low_res, timestep_respacing="7", **kwargs): + assert ( + timestep_respacing == "7" + ), "different respacing method may work, but no guaranteed" + + sample_fn = self.get_sample_fn(timestep_respacing) + sample_outputs = sample_fn( + self.model_first_steps, + self.model_last_step, + shape=low_res.shape, + clip_denoised=True, + model_kwargs=dict(low_res=low_res), + **kwargs, + ) + for x in sample_outputs: + sample = x["sample"] + yield sample diff --git a/ldm/modules/karlo/kakao/modules/__init__.py b/ldm/modules/karlo/kakao/modules/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..11d4358a6475243b8b789efb7595566993ccd9e1 --- /dev/null +++ b/ldm/modules/karlo/kakao/modules/__init__.py @@ -0,0 +1,49 @@ +# ------------------------------------------------------------------------------------ +# Adapted from Guided-Diffusion repo (https://github.com/openai/guided-diffusion) +# ------------------------------------------------------------------------------------ + + +from .diffusion import gaussian_diffusion as gd +from .diffusion.respace import ( + SpacedDiffusion, + space_timesteps, +) + + +def create_gaussian_diffusion( + steps, + learn_sigma, + sigma_small, + noise_schedule, + use_kl, + predict_xstart, + rescale_learned_sigmas, + timestep_respacing, +): + betas = gd.get_named_beta_schedule(noise_schedule, steps) + if use_kl: + loss_type = gd.LossType.RESCALED_KL + elif rescale_learned_sigmas: + loss_type = gd.LossType.RESCALED_MSE + else: + loss_type = gd.LossType.MSE + if not timestep_respacing: + timestep_respacing = [steps] + + return SpacedDiffusion( + use_timesteps=space_timesteps(steps, timestep_respacing), + betas=betas, + model_mean_type=( + gd.ModelMeanType.EPSILON if not predict_xstart else gd.ModelMeanType.START_X + ), + model_var_type=( + ( + gd.ModelVarType.FIXED_LARGE + if not sigma_small + else gd.ModelVarType.FIXED_SMALL + ) + if not learn_sigma + else gd.ModelVarType.LEARNED_RANGE + ), + loss_type=loss_type, + ) diff --git a/ldm/modules/karlo/kakao/modules/diffusion/gaussian_diffusion.py b/ldm/modules/karlo/kakao/modules/diffusion/gaussian_diffusion.py new file mode 100644 index 0000000000000000000000000000000000000000..6a111aa09ea84157735cc2a4a66a68c4faaf42f4 --- /dev/null +++ b/ldm/modules/karlo/kakao/modules/diffusion/gaussian_diffusion.py @@ -0,0 +1,828 @@ +# ------------------------------------------------------------------------------------ +# Adapted from Guided-Diffusion repo (https://github.com/openai/guided-diffusion) +# ------------------------------------------------------------------------------------ + +import enum +import math + +import numpy as np +import torch as th + + +def _warmup_beta(beta_start, beta_end, num_diffusion_timesteps, warmup_frac): + betas = beta_end * np.ones(num_diffusion_timesteps, dtype=np.float64) + warmup_time = int(num_diffusion_timesteps * warmup_frac) + betas[:warmup_time] = np.linspace( + beta_start, beta_end, warmup_time, dtype=np.float64 + ) + return betas + + +def get_beta_schedule(beta_schedule, *, beta_start, beta_end, num_diffusion_timesteps): + """ + This is the deprecated API for creating beta schedules. + See get_named_beta_schedule() for the new library of schedules. + """ + if beta_schedule == "quad": + betas = ( + np.linspace( + beta_start**0.5, + beta_end**0.5, + num_diffusion_timesteps, + dtype=np.float64, + ) + ** 2 + ) + elif beta_schedule == "linear": + betas = np.linspace( + beta_start, beta_end, num_diffusion_timesteps, dtype=np.float64 + ) + elif beta_schedule == "warmup10": + betas = _warmup_beta(beta_start, beta_end, num_diffusion_timesteps, 0.1) + elif beta_schedule == "warmup50": + betas = _warmup_beta(beta_start, beta_end, num_diffusion_timesteps, 0.5) + elif beta_schedule == "const": + betas = beta_end * np.ones(num_diffusion_timesteps, dtype=np.float64) + elif beta_schedule == "jsd": # 1/T, 1/(T-1), 1/(T-2), ..., 1 + betas = 1.0 / np.linspace( + num_diffusion_timesteps, 1, num_diffusion_timesteps, dtype=np.float64 + ) + else: + raise NotImplementedError(beta_schedule) + assert betas.shape == (num_diffusion_timesteps,) + return betas + + +def get_named_beta_schedule(schedule_name, num_diffusion_timesteps): + """ + Get a pre-defined beta schedule for the given name. + The beta schedule library consists of beta schedules which remain similar + in the limit of num_diffusion_timesteps. + Beta schedules may be added, but should not be removed or changed once + they are committed to maintain backwards compatibility. + """ + if schedule_name == "linear": + # Linear schedule from Ho et al, extended to work for any number of + # diffusion steps. + scale = 1000 / num_diffusion_timesteps + return get_beta_schedule( + "linear", + beta_start=scale * 0.0001, + beta_end=scale * 0.02, + num_diffusion_timesteps=num_diffusion_timesteps, + ) + elif schedule_name == "squaredcos_cap_v2": + return betas_for_alpha_bar( + num_diffusion_timesteps, + lambda t: math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2, + ) + else: + raise NotImplementedError(f"unknown beta schedule: {schedule_name}") + + +def betas_for_alpha_bar(num_diffusion_timesteps, alpha_bar, max_beta=0.999): + """ + Create a beta schedule that discretizes the given alpha_t_bar function, + which defines the cumulative product of (1-beta) over time from t = [0,1]. + :param num_diffusion_timesteps: the number of betas to produce. + :param alpha_bar: a lambda that takes an argument t from 0 to 1 and + produces the cumulative product of (1-beta) up to that + part of the diffusion process. + :param max_beta: the maximum beta to use; use values lower than 1 to + prevent singularities. + """ + betas = [] + for i in range(num_diffusion_timesteps): + t1 = i / num_diffusion_timesteps + t2 = (i + 1) / num_diffusion_timesteps + betas.append(min(1 - alpha_bar(t2) / alpha_bar(t1), max_beta)) + return np.array(betas) + + +class ModelMeanType(enum.Enum): + """ + Which type of output the model predicts. + """ + + PREVIOUS_X = enum.auto() # the model predicts x_{t-1} + START_X = enum.auto() # the model predicts x_0 + EPSILON = enum.auto() # the model predicts epsilon + + +class ModelVarType(enum.Enum): + """ + What is used as the model's output variance. + The LEARNED_RANGE option has been added to allow the model to predict + values between FIXED_SMALL and FIXED_LARGE, making its job easier. + """ + + LEARNED = enum.auto() + FIXED_SMALL = enum.auto() + FIXED_LARGE = enum.auto() + LEARNED_RANGE = enum.auto() + + +class LossType(enum.Enum): + MSE = enum.auto() # use raw MSE loss (and KL when learning variances) + RESCALED_MSE = ( + enum.auto() + ) # use raw MSE loss (with RESCALED_KL when learning variances) + KL = enum.auto() # use the variational lower-bound + RESCALED_KL = enum.auto() # like KL, but rescale to estimate the full VLB + + def is_vb(self): + return self == LossType.KL or self == LossType.RESCALED_KL + + +class GaussianDiffusion(th.nn.Module): + """ + Utilities for training and sampling diffusion models. + Original ported from this codebase: + https://github.com/hojonathanho/diffusion/blob/1e0dceb3b3495bbe19116a5e1b3596cd0706c543/diffusion_tf/diffusion_utils_2.py#L42 + :param betas: a 1-D numpy array of betas for each diffusion timestep, + starting at T and going to 1. + """ + + def __init__( + self, + *, + betas, + model_mean_type, + model_var_type, + loss_type, + ): + super(GaussianDiffusion, self).__init__() + self.model_mean_type = model_mean_type + self.model_var_type = model_var_type + self.loss_type = loss_type + + # Use float64 for accuracy. + betas = np.array(betas, dtype=np.float64) + assert len(betas.shape) == 1, "betas must be 1-D" + assert (betas > 0).all() and (betas <= 1).all() + + self.num_timesteps = int(betas.shape[0]) + + alphas = 1.0 - betas + alphas_cumprod = np.cumprod(alphas, axis=0) + alphas_cumprod_prev = np.append(1.0, alphas_cumprod[:-1]) + alphas_cumprod_next = np.append(alphas_cumprod[1:], 0.0) + assert alphas_cumprod_prev.shape == (self.num_timesteps,) + + # calculations for diffusion q(x_t | x_{t-1}) and others + sqrt_alphas_cumprod = np.sqrt(alphas_cumprod) + sqrt_one_minus_alphas_cumprod = np.sqrt(1.0 - alphas_cumprod) + log_one_minus_alphas_cumprod = np.log(1.0 - alphas_cumprod) + sqrt_recip_alphas_cumprod = np.sqrt(1.0 / alphas_cumprod) + sqrt_recipm1_alphas_cumprod = np.sqrt(1.0 / alphas_cumprod - 1) + + # calculations for posterior q(x_{t-1} | x_t, x_0) + posterior_variance = ( + betas * (1.0 - alphas_cumprod_prev) / (1.0 - alphas_cumprod) + ) + # below: log calculation clipped because the posterior variance is 0 at the beginning of the diffusion chain + posterior_log_variance_clipped = np.log( + np.append(posterior_variance[1], posterior_variance[1:]) + ) + posterior_mean_coef1 = ( + betas * np.sqrt(alphas_cumprod_prev) / (1.0 - alphas_cumprod) + ) + posterior_mean_coef2 = ( + (1.0 - alphas_cumprod_prev) * np.sqrt(alphas) / (1.0 - alphas_cumprod) + ) + + self.register_buffer("betas", th.from_numpy(betas), persistent=False) + self.register_buffer( + "alphas_cumprod", th.from_numpy(alphas_cumprod), persistent=False + ) + self.register_buffer( + "alphas_cumprod_prev", th.from_numpy(alphas_cumprod_prev), persistent=False + ) + self.register_buffer( + "alphas_cumprod_next", th.from_numpy(alphas_cumprod_next), persistent=False + ) + + self.register_buffer( + "sqrt_alphas_cumprod", th.from_numpy(sqrt_alphas_cumprod), persistent=False + ) + self.register_buffer( + "sqrt_one_minus_alphas_cumprod", + th.from_numpy(sqrt_one_minus_alphas_cumprod), + persistent=False, + ) + self.register_buffer( + "log_one_minus_alphas_cumprod", + th.from_numpy(log_one_minus_alphas_cumprod), + persistent=False, + ) + self.register_buffer( + "sqrt_recip_alphas_cumprod", + th.from_numpy(sqrt_recip_alphas_cumprod), + persistent=False, + ) + self.register_buffer( + "sqrt_recipm1_alphas_cumprod", + th.from_numpy(sqrt_recipm1_alphas_cumprod), + persistent=False, + ) + + self.register_buffer( + "posterior_variance", th.from_numpy(posterior_variance), persistent=False + ) + self.register_buffer( + "posterior_log_variance_clipped", + th.from_numpy(posterior_log_variance_clipped), + persistent=False, + ) + self.register_buffer( + "posterior_mean_coef1", + th.from_numpy(posterior_mean_coef1), + persistent=False, + ) + self.register_buffer( + "posterior_mean_coef2", + th.from_numpy(posterior_mean_coef2), + persistent=False, + ) + + def q_mean_variance(self, x_start, t): + """ + Get the distribution q(x_t | x_0). + :param x_start: the [N x C x ...] tensor of noiseless inputs. + :param t: the number of diffusion steps (minus 1). Here, 0 means one step. + :return: A tuple (mean, variance, log_variance), all of x_start's shape. + """ + mean = ( + _extract_into_tensor(self.sqrt_alphas_cumprod, t, x_start.shape) * x_start + ) + variance = _extract_into_tensor(1.0 - self.alphas_cumprod, t, x_start.shape) + log_variance = _extract_into_tensor( + self.log_one_minus_alphas_cumprod, t, x_start.shape + ) + return mean, variance, log_variance + + def q_sample(self, x_start, t, noise=None): + """ + Diffuse the data for a given number of diffusion steps. + In other words, sample from q(x_t | x_0). + :param x_start: the initial data batch. + :param t: the number of diffusion steps (minus 1). Here, 0 means one step. + :param noise: if specified, the split-out normal noise. + :return: A noisy version of x_start. + """ + if noise is None: + noise = th.randn_like(x_start) + assert noise.shape == x_start.shape + return ( + _extract_into_tensor(self.sqrt_alphas_cumprod, t, x_start.shape) * x_start + + _extract_into_tensor(self.sqrt_one_minus_alphas_cumprod, t, x_start.shape) + * noise + ) + + def q_posterior_mean_variance(self, x_start, x_t, t): + """ + Compute the mean and variance of the diffusion posterior: + q(x_{t-1} | x_t, x_0) + """ + assert x_start.shape == x_t.shape + posterior_mean = ( + _extract_into_tensor(self.posterior_mean_coef1, t, x_t.shape) * x_start + + _extract_into_tensor(self.posterior_mean_coef2, t, x_t.shape) * x_t + ) + posterior_variance = _extract_into_tensor(self.posterior_variance, t, x_t.shape) + posterior_log_variance_clipped = _extract_into_tensor( + self.posterior_log_variance_clipped, t, x_t.shape + ) + assert ( + posterior_mean.shape[0] + == posterior_variance.shape[0] + == posterior_log_variance_clipped.shape[0] + == x_start.shape[0] + ) + return posterior_mean, posterior_variance, posterior_log_variance_clipped + + def p_mean_variance( + self, + model, + x, + t, + clip_denoised=True, + denoised_fn=None, + model_kwargs=None, + **ignore_kwargs, + ): + """ + Apply the model to get p(x_{t-1} | x_t), as well as a prediction of + the initial x, x_0. + :param model: the model, which takes a signal and a batch of timesteps + as input. + :param x: the [N x C x ...] tensor at time t. + :param t: a 1-D Tensor of timesteps. + :param clip_denoised: if True, clip the denoised signal into [-1, 1]. + :param denoised_fn: if not None, a function which applies to the + x_start prediction before it is used to sample. Applies before + clip_denoised. + :param model_kwargs: if not None, a dict of extra keyword arguments to + pass to the model. This can be used for conditioning. + :return: a dict with the following keys: + - 'mean': the model mean output. + - 'variance': the model variance output. + - 'log_variance': the log of 'variance'. + - 'pred_xstart': the prediction for x_0. + """ + if model_kwargs is None: + model_kwargs = {} + + B, C = x.shape[:2] + assert t.shape == (B,) + model_output = model(x, t, **model_kwargs) + if isinstance(model_output, tuple): + model_output, extra = model_output + else: + extra = None + + if self.model_var_type in [ModelVarType.LEARNED, ModelVarType.LEARNED_RANGE]: + assert model_output.shape == (B, C * 2, *x.shape[2:]) + model_output, model_var_values = th.split(model_output, C, dim=1) + if self.model_var_type == ModelVarType.LEARNED: + model_log_variance = model_var_values + model_variance = th.exp(model_log_variance) + else: + min_log = _extract_into_tensor( + self.posterior_log_variance_clipped, t, x.shape + ) + max_log = _extract_into_tensor(th.log(self.betas), t, x.shape) + # The model_var_values is [-1, 1] for [min_var, max_var]. + frac = (model_var_values + 1) / 2 + model_log_variance = frac * max_log + (1 - frac) * min_log + model_variance = th.exp(model_log_variance) + else: + model_variance, model_log_variance = { + # for fixedlarge, we set the initial (log-)variance like so + # to get a better decoder log likelihood. + ModelVarType.FIXED_LARGE: ( + th.cat([self.posterior_variance[1][None], self.betas[1:]]), + th.log(th.cat([self.posterior_variance[1][None], self.betas[1:]])), + ), + ModelVarType.FIXED_SMALL: ( + self.posterior_variance, + self.posterior_log_variance_clipped, + ), + }[self.model_var_type] + model_variance = _extract_into_tensor(model_variance, t, x.shape) + model_log_variance = _extract_into_tensor(model_log_variance, t, x.shape) + + def process_xstart(x): + if denoised_fn is not None: + x = denoised_fn(x) + if clip_denoised: + return x.clamp(-1, 1) + return x + + if self.model_mean_type == ModelMeanType.PREVIOUS_X: + pred_xstart = process_xstart( + self._predict_xstart_from_xprev(x_t=x, t=t, xprev=model_output) + ) + model_mean = model_output + elif self.model_mean_type in [ModelMeanType.START_X, ModelMeanType.EPSILON]: + if self.model_mean_type == ModelMeanType.START_X: + pred_xstart = process_xstart(model_output) + else: + pred_xstart = process_xstart( + self._predict_xstart_from_eps(x_t=x, t=t, eps=model_output) + ) + model_mean, _, _ = self.q_posterior_mean_variance( + x_start=pred_xstart, x_t=x, t=t + ) + else: + raise NotImplementedError(self.model_mean_type) + + assert ( + model_mean.shape == model_log_variance.shape == pred_xstart.shape == x.shape + ) + return { + "mean": model_mean, + "variance": model_variance, + "log_variance": model_log_variance, + "pred_xstart": pred_xstart, + } + + def _predict_xstart_from_eps(self, x_t, t, eps): + assert x_t.shape == eps.shape + return ( + _extract_into_tensor(self.sqrt_recip_alphas_cumprod, t, x_t.shape) * x_t + - _extract_into_tensor(self.sqrt_recipm1_alphas_cumprod, t, x_t.shape) * eps + ) + + def _predict_eps_from_xstart(self, x_t, t, pred_xstart): + return ( + _extract_into_tensor(self.sqrt_recip_alphas_cumprod, t, x_t.shape) * x_t + - pred_xstart + ) / _extract_into_tensor(self.sqrt_recipm1_alphas_cumprod, t, x_t.shape) + + def condition_mean(self, cond_fn, p_mean_var, x, t, model_kwargs=None): + """ + Compute the mean for the previous step, given a function cond_fn that + computes the gradient of a conditional log probability with respect to + x. In particular, cond_fn computes grad(log(p(y|x))), and we want to + condition on y. + This uses the conditioning strategy from Sohl-Dickstein et al. (2015). + """ + gradient = cond_fn(x, t, **model_kwargs) + new_mean = ( + p_mean_var["mean"].float() + p_mean_var["variance"] * gradient.float() + ) + return new_mean + + def condition_score(self, cond_fn, p_mean_var, x, t, model_kwargs=None): + """ + Compute what the p_mean_variance output would have been, should the + model's score function be conditioned by cond_fn. + See condition_mean() for details on cond_fn. + Unlike condition_mean(), this instead uses the conditioning strategy + from Song et al (2020). + """ + alpha_bar = _extract_into_tensor(self.alphas_cumprod, t, x.shape) + + eps = self._predict_eps_from_xstart(x, t, p_mean_var["pred_xstart"]) + eps = eps - (1 - alpha_bar).sqrt() * cond_fn(x, t, **model_kwargs) + + out = p_mean_var.copy() + out["pred_xstart"] = self._predict_xstart_from_eps(x, t, eps) + out["mean"], _, _ = self.q_posterior_mean_variance( + x_start=out["pred_xstart"], x_t=x, t=t + ) + return out + + def p_sample( + self, + model, + x, + t, + clip_denoised=True, + denoised_fn=None, + cond_fn=None, + model_kwargs=None, + ): + """ + Sample x_{t-1} from the model at the given timestep. + :param model: the model to sample from. + :param x: the current tensor at x_{t-1}. + :param t: the value of t, starting at 0 for the first diffusion step. + :param clip_denoised: if True, clip the x_start prediction to [-1, 1]. + :param denoised_fn: if not None, a function which applies to the + x_start prediction before it is used to sample. + :param cond_fn: if not None, this is a gradient function that acts + similarly to the model. + :param model_kwargs: if not None, a dict of extra keyword arguments to + pass to the model. This can be used for conditioning. + :return: a dict containing the following keys: + - 'sample': a random sample from the model. + - 'pred_xstart': a prediction of x_0. + """ + out = self.p_mean_variance( + model, + x, + t, + clip_denoised=clip_denoised, + denoised_fn=denoised_fn, + model_kwargs=model_kwargs, + ) + noise = th.randn_like(x) + nonzero_mask = ( + (t != 0).float().view(-1, *([1] * (len(x.shape) - 1))) + ) # no noise when t == 0 + if cond_fn is not None: + out["mean"] = self.condition_mean( + cond_fn, out, x, t, model_kwargs=model_kwargs + ) + sample = out["mean"] + nonzero_mask * th.exp(0.5 * out["log_variance"]) * noise + return {"sample": sample, "pred_xstart": out["pred_xstart"]} + + def p_sample_loop( + self, + model, + shape, + noise=None, + clip_denoised=True, + denoised_fn=None, + cond_fn=None, + model_kwargs=None, + device=None, + progress=False, + ): + """ + Generate samples from the model. + :param model: the model module. + :param shape: the shape of the samples, (N, C, H, W). + :param noise: if specified, the noise from the encoder to sample. + Should be of the same shape as `shape`. + :param clip_denoised: if True, clip x_start predictions to [-1, 1]. + :param denoised_fn: if not None, a function which applies to the + x_start prediction before it is used to sample. + :param cond_fn: if not None, this is a gradient function that acts + similarly to the model. + :param model_kwargs: if not None, a dict of extra keyword arguments to + pass to the model. This can be used for conditioning. + :param device: if specified, the device to create the samples on. + If not specified, use a model parameter's device. + :param progress: if True, show a tqdm progress bar. + :return: a non-differentiable batch of samples. + """ + final = None + for sample in self.p_sample_loop_progressive( + model, + shape, + noise=noise, + clip_denoised=clip_denoised, + denoised_fn=denoised_fn, + cond_fn=cond_fn, + model_kwargs=model_kwargs, + device=device, + progress=progress, + ): + final = sample + return final["sample"] + + def p_sample_loop_progressive( + self, + model, + shape, + noise=None, + clip_denoised=True, + denoised_fn=None, + cond_fn=None, + model_kwargs=None, + device=None, + progress=False, + ): + """ + Generate samples from the model and yield intermediate samples from + each timestep of diffusion. + Arguments are the same as p_sample_loop(). + Returns a generator over dicts, where each dict is the return value of + p_sample(). + """ + if device is None: + device = next(model.parameters()).device + assert isinstance(shape, (tuple, list)) + if noise is not None: + img = noise + else: + img = th.randn(*shape, device=device) + indices = list(range(self.num_timesteps))[::-1] + + if progress: + # Lazy import so that we don't depend on tqdm. + from tqdm.auto import tqdm + + indices = tqdm(indices) + + for idx, i in enumerate(indices): + t = th.tensor([i] * shape[0], device=device) + with th.no_grad(): + out = self.p_sample( + model, + img, + t, + clip_denoised=clip_denoised, + denoised_fn=denoised_fn, + cond_fn=cond_fn, + model_kwargs=model_kwargs, + ) + yield out + img = out["sample"] + + def p_sample_loop_progressive_for_improved_sr( + self, + model, + model_aux, + shape, + noise=None, + clip_denoised=True, + denoised_fn=None, + cond_fn=None, + model_kwargs=None, + device=None, + progress=False, + ): + """ + Modified version of p_sample_loop_progressive for sampling from the improved sr model + """ + + if device is None: + device = next(model.parameters()).device + assert isinstance(shape, (tuple, list)) + if noise is not None: + img = noise + else: + img = th.randn(*shape, device=device) + indices = list(range(self.num_timesteps))[::-1] + + if progress: + # Lazy import so that we don't depend on tqdm. + from tqdm.auto import tqdm + + indices = tqdm(indices) + + for idx, i in enumerate(indices): + t = th.tensor([i] * shape[0], device=device) + with th.no_grad(): + out = self.p_sample( + model_aux if len(indices) - 1 == idx else model, + img, + t, + clip_denoised=clip_denoised, + denoised_fn=denoised_fn, + cond_fn=cond_fn, + model_kwargs=model_kwargs, + ) + yield out + img = out["sample"] + + def ddim_sample( + self, + model, + x, + t, + clip_denoised=True, + denoised_fn=None, + cond_fn=None, + model_kwargs=None, + eta=0.0, + ): + """ + Sample x_{t-1} from the model using DDIM. + Same usage as p_sample(). + """ + out = self.p_mean_variance( + model, + x, + t, + clip_denoised=clip_denoised, + denoised_fn=denoised_fn, + model_kwargs=model_kwargs, + ) + if cond_fn is not None: + out = self.condition_score(cond_fn, out, x, t, model_kwargs=model_kwargs) + + # Usually our model outputs epsilon, but we re-derive it + # in case we used x_start or x_prev prediction. + eps = self._predict_eps_from_xstart(x, t, out["pred_xstart"]) + + alpha_bar = _extract_into_tensor(self.alphas_cumprod, t, x.shape) + alpha_bar_prev = _extract_into_tensor(self.alphas_cumprod_prev, t, x.shape) + sigma = ( + eta + * th.sqrt((1 - alpha_bar_prev) / (1 - alpha_bar)) + * th.sqrt(1 - alpha_bar / alpha_bar_prev) + ) + # Equation 12. + noise = th.randn_like(x) + mean_pred = ( + out["pred_xstart"] * th.sqrt(alpha_bar_prev) + + th.sqrt(1 - alpha_bar_prev - sigma**2) * eps + ) + nonzero_mask = ( + (t != 0).float().view(-1, *([1] * (len(x.shape) - 1))) + ) # no noise when t == 0 + sample = mean_pred + nonzero_mask * sigma * noise + return {"sample": sample, "pred_xstart": out["pred_xstart"]} + + def ddim_reverse_sample( + self, + model, + x, + t, + clip_denoised=True, + denoised_fn=None, + cond_fn=None, + model_kwargs=None, + eta=0.0, + ): + """ + Sample x_{t+1} from the model using DDIM reverse ODE. + """ + assert eta == 0.0, "Reverse ODE only for deterministic path" + out = self.p_mean_variance( + model, + x, + t, + clip_denoised=clip_denoised, + denoised_fn=denoised_fn, + model_kwargs=model_kwargs, + ) + if cond_fn is not None: + out = self.condition_score(cond_fn, out, x, t, model_kwargs=model_kwargs) + # Usually our model outputs epsilon, but we re-derive it + # in case we used x_start or x_prev prediction. + eps = ( + _extract_into_tensor(self.sqrt_recip_alphas_cumprod, t, x.shape) * x + - out["pred_xstart"] + ) / _extract_into_tensor(self.sqrt_recipm1_alphas_cumprod, t, x.shape) + alpha_bar_next = _extract_into_tensor(self.alphas_cumprod_next, t, x.shape) + + # Equation 12. reversed + mean_pred = ( + out["pred_xstart"] * th.sqrt(alpha_bar_next) + + th.sqrt(1 - alpha_bar_next) * eps + ) + + return {"sample": mean_pred, "pred_xstart": out["pred_xstart"]} + + def ddim_sample_loop( + self, + model, + shape, + noise=None, + clip_denoised=True, + denoised_fn=None, + cond_fn=None, + model_kwargs=None, + device=None, + progress=False, + eta=0.0, + ): + """ + Generate samples from the model using DDIM. + Same usage as p_sample_loop(). + """ + final = None + for sample in self.ddim_sample_loop_progressive( + model, + shape, + noise=noise, + clip_denoised=clip_denoised, + denoised_fn=denoised_fn, + cond_fn=cond_fn, + model_kwargs=model_kwargs, + device=device, + progress=progress, + eta=eta, + ): + final = sample + return final["sample"] + + def ddim_sample_loop_progressive( + self, + model, + shape, + noise=None, + clip_denoised=True, + denoised_fn=None, + cond_fn=None, + model_kwargs=None, + device=None, + progress=False, + eta=0.0, + ): + """ + Use DDIM to sample from the model and yield intermediate samples from + each timestep of DDIM. + Same usage as p_sample_loop_progressive(). + """ + if device is None: + device = next(model.parameters()).device + assert isinstance(shape, (tuple, list)) + if noise is not None: + img = noise + else: + img = th.randn(*shape, device=device) + indices = list(range(self.num_timesteps))[::-1] + + if progress: + # Lazy import so that we don't depend on tqdm. + from tqdm.auto import tqdm + + indices = tqdm(indices) + + for i in indices: + t = th.tensor([i] * shape[0], device=device) + with th.no_grad(): + out = self.ddim_sample( + model, + img, + t, + clip_denoised=clip_denoised, + denoised_fn=denoised_fn, + cond_fn=cond_fn, + model_kwargs=model_kwargs, + eta=eta, + ) + yield out + img = out["sample"] + + +def _extract_into_tensor(arr, timesteps, broadcast_shape): + """ + Extract values from a 1-D numpy array for a batch of indices. + :param arr: the 1-D numpy array. + :param timesteps: a tensor of indices into the array to extract. + :param broadcast_shape: a larger shape of K dimensions with the batch + dimension equal to the length of timesteps. + :return: a tensor of shape [batch_size, 1, ...] where the shape has K dims. + """ + res = arr.to(device=timesteps.device)[timesteps].float() + while len(res.shape) < len(broadcast_shape): + res = res[..., None] + return res + th.zeros(broadcast_shape, device=timesteps.device) diff --git a/ldm/modules/karlo/kakao/modules/diffusion/respace.py b/ldm/modules/karlo/kakao/modules/diffusion/respace.py new file mode 100644 index 0000000000000000000000000000000000000000..70c808f8b3ce947a8466026ca0d4ceccabda3745 --- /dev/null +++ b/ldm/modules/karlo/kakao/modules/diffusion/respace.py @@ -0,0 +1,112 @@ +# ------------------------------------------------------------------------------------ +# Adapted from Guided-Diffusion repo (https://github.com/openai/guided-diffusion) +# ------------------------------------------------------------------------------------ + + +import torch as th + +from .gaussian_diffusion import GaussianDiffusion + + +def space_timesteps(num_timesteps, section_counts): + """ + Create a list of timesteps to use from an original diffusion process, + given the number of timesteps we want to take from equally-sized portions + of the original process. + + For example, if there's 300 timesteps and the section counts are [10,15,20] + then the first 100 timesteps are strided to be 10 timesteps, the second 100 + are strided to be 15 timesteps, and the final 100 are strided to be 20. + + :param num_timesteps: the number of diffusion steps in the original + process to divide up. + :param section_counts: either a list of numbers, or a string containing + comma-separated numbers, indicating the step count + per section. As a special case, use "ddimN" where N + is a number of steps to use the striding from the + DDIM paper. + :return: a set of diffusion steps from the original process to use. + """ + if isinstance(section_counts, str): + if section_counts.startswith("ddim"): + desired_count = int(section_counts[len("ddim") :]) + for i in range(1, num_timesteps): + if len(range(0, num_timesteps, i)) == desired_count: + return set(range(0, num_timesteps, i)) + raise ValueError( + f"cannot create exactly {num_timesteps} steps with an integer stride" + ) + elif section_counts == "fast27": + steps = space_timesteps(num_timesteps, "10,10,3,2,2") + # Help reduce DDIM artifacts from noisiest timesteps. + steps.remove(num_timesteps - 1) + steps.add(num_timesteps - 3) + return steps + section_counts = [int(x) for x in section_counts.split(",")] + size_per = num_timesteps // len(section_counts) + extra = num_timesteps % len(section_counts) + start_idx = 0 + all_steps = [] + for i, section_count in enumerate(section_counts): + size = size_per + (1 if i < extra else 0) + if size < section_count: + raise ValueError( + f"cannot divide section of {size} steps into {section_count}" + ) + if section_count <= 1: + frac_stride = 1 + else: + frac_stride = (size - 1) / (section_count - 1) + cur_idx = 0.0 + taken_steps = [] + for _ in range(section_count): + taken_steps.append(start_idx + round(cur_idx)) + cur_idx += frac_stride + all_steps += taken_steps + start_idx += size + return set(all_steps) + + +class SpacedDiffusion(GaussianDiffusion): + """ + A diffusion process which can skip steps in a base diffusion process. + + :param use_timesteps: a collection (sequence or set) of timesteps from the + original diffusion process to retain. + :param kwargs: the kwargs to create the base diffusion process. + """ + + def __init__(self, use_timesteps, **kwargs): + self.use_timesteps = set(use_timesteps) + self.original_num_steps = len(kwargs["betas"]) + + base_diffusion = GaussianDiffusion(**kwargs) # pylint: disable=missing-kwoa + last_alpha_cumprod = 1.0 + new_betas = [] + timestep_map = [] + for i, alpha_cumprod in enumerate(base_diffusion.alphas_cumprod): + if i in self.use_timesteps: + new_betas.append(1 - alpha_cumprod / last_alpha_cumprod) + last_alpha_cumprod = alpha_cumprod + timestep_map.append(i) + kwargs["betas"] = th.tensor(new_betas).numpy() + super().__init__(**kwargs) + self.register_buffer("timestep_map", th.tensor(timestep_map), persistent=False) + + def p_mean_variance(self, model, *args, **kwargs): + return super().p_mean_variance(self._wrap_model(model), *args, **kwargs) + + def condition_mean(self, cond_fn, *args, **kwargs): + return super().condition_mean(self._wrap_model(cond_fn), *args, **kwargs) + + def condition_score(self, cond_fn, *args, **kwargs): + return super().condition_score(self._wrap_model(cond_fn), *args, **kwargs) + + def _wrap_model(self, model): + def wrapped(x, ts, **kwargs): + ts_cpu = ts.detach().to("cpu") + return model( + x, self.timestep_map[ts_cpu].to(device=ts.device, dtype=ts.dtype), **kwargs + ) + + return wrapped diff --git a/ldm/modules/karlo/kakao/modules/nn.py b/ldm/modules/karlo/kakao/modules/nn.py new file mode 100644 index 0000000000000000000000000000000000000000..2eef3f5a06778b269f99168d2a2962df84422ad0 --- /dev/null +++ b/ldm/modules/karlo/kakao/modules/nn.py @@ -0,0 +1,114 @@ +# ------------------------------------------------------------------------------------ +# Adapted from Guided-Diffusion repo (https://github.com/openai/guided-diffusion) +# ------------------------------------------------------------------------------------ + +import math + +import torch as th +import torch.nn as nn +import torch.nn.functional as F + + +class GroupNorm32(nn.GroupNorm): + def __init__(self, num_groups, num_channels, swish, eps=1e-5): + super().__init__(num_groups=num_groups, num_channels=num_channels, eps=eps) + self.swish = swish + + def forward(self, x): + y = super().forward(x.float()).to(x.dtype) + if self.swish == 1.0: + y = F.silu(y) + elif self.swish: + y = y * F.sigmoid(y * float(self.swish)) + return y + + +def conv_nd(dims, *args, **kwargs): + """ + Create a 1D, 2D, or 3D convolution module. + """ + if dims == 1: + return nn.Conv1d(*args, **kwargs) + elif dims == 2: + return nn.Conv2d(*args, **kwargs) + elif dims == 3: + return nn.Conv3d(*args, **kwargs) + raise ValueError(f"unsupported dimensions: {dims}") + + +def linear(*args, **kwargs): + """ + Create a linear module. + """ + return nn.Linear(*args, **kwargs) + + +def avg_pool_nd(dims, *args, **kwargs): + """ + Create a 1D, 2D, or 3D average pooling module. + """ + if dims == 1: + return nn.AvgPool1d(*args, **kwargs) + elif dims == 2: + return nn.AvgPool2d(*args, **kwargs) + elif dims == 3: + return nn.AvgPool3d(*args, **kwargs) + raise ValueError(f"unsupported dimensions: {dims}") + + +def zero_module(module): + """ + Zero out the parameters of a module and return it. + """ + for p in module.parameters(): + p.detach().zero_() + return module + + +def scale_module(module, scale): + """ + Scale the parameters of a module and return it. + """ + for p in module.parameters(): + p.detach().mul_(scale) + return module + + +def normalization(channels, swish=0.0): + """ + Make a standard normalization layer, with an optional swish activation. + + :param channels: number of input channels. + :return: an nn.Module for normalization. + """ + return GroupNorm32(num_channels=channels, num_groups=32, swish=swish) + + +def timestep_embedding(timesteps, dim, max_period=10000): + """ + Create sinusoidal timestep embeddings. + + :param timesteps: a 1-D Tensor of N indices, one per batch element. + These may be fractional. + :param dim: the dimension of the output. + :param max_period: controls the minimum frequency of the embeddings. + :return: an [N x dim] Tensor of positional embeddings. + """ + half = dim // 2 + freqs = th.exp( + -math.log(max_period) + * th.arange(start=0, end=half, dtype=th.float32, device=timesteps.device) + / half + ) + args = timesteps[:, None].float() * freqs[None] + embedding = th.cat([th.cos(args), th.sin(args)], dim=-1) + if dim % 2: + embedding = th.cat([embedding, th.zeros_like(embedding[:, :1])], dim=-1) + return embedding + + +def mean_flat(tensor): + """ + Take the mean over all non-batch dimensions. + """ + return tensor.mean(dim=list(range(1, len(tensor.shape)))) diff --git a/ldm/modules/karlo/kakao/modules/resample.py b/ldm/modules/karlo/kakao/modules/resample.py new file mode 100644 index 0000000000000000000000000000000000000000..485421aa4066a34ec01bdba8a26bd113598461ac --- /dev/null +++ b/ldm/modules/karlo/kakao/modules/resample.py @@ -0,0 +1,68 @@ +# ------------------------------------------------------------------------------------ +# Modified from Guided-Diffusion (https://github.com/openai/guided-diffusion) +# ------------------------------------------------------------------------------------ + +from abc import abstractmethod + +import torch as th + + +def create_named_schedule_sampler(name, diffusion): + """ + Create a ScheduleSampler from a library of pre-defined samplers. + + :param name: the name of the sampler. + :param diffusion: the diffusion object to sample for. + """ + if name == "uniform": + return UniformSampler(diffusion) + else: + raise NotImplementedError(f"unknown schedule sampler: {name}") + + +class ScheduleSampler(th.nn.Module): + """ + A distribution over timesteps in the diffusion process, intended to reduce + variance of the objective. + + By default, samplers perform unbiased importance sampling, in which the + objective's mean is unchanged. + However, subclasses may override sample() to change how the resampled + terms are reweighted, allowing for actual changes in the objective. + """ + + @abstractmethod + def weights(self): + """ + Get a numpy array of weights, one per diffusion step. + + The weights needn't be normalized, but must be positive. + """ + + def sample(self, batch_size, device): + """ + Importance-sample timesteps for a batch. + + :param batch_size: the number of timesteps. + :param device: the torch device to save to. + :return: a tuple (timesteps, weights): + - timesteps: a tensor of timestep indices. + - weights: a tensor of weights to scale the resulting losses. + """ + w = self.weights() + p = w / th.sum(w) + indices = p.multinomial(batch_size, replacement=True) + weights = 1 / (len(p) * p[indices]) + return indices, weights + + +class UniformSampler(ScheduleSampler): + def __init__(self, diffusion): + super(UniformSampler, self).__init__() + self.diffusion = diffusion + self.register_buffer( + "_weights", th.ones([diffusion.num_timesteps]), persistent=False + ) + + def weights(self): + return self._weights diff --git a/ldm/modules/karlo/kakao/modules/unet.py b/ldm/modules/karlo/kakao/modules/unet.py new file mode 100644 index 0000000000000000000000000000000000000000..c99d0b791876add8b85354d05d61a6ae2c852e68 --- /dev/null +++ b/ldm/modules/karlo/kakao/modules/unet.py @@ -0,0 +1,792 @@ +# ------------------------------------------------------------------------------------ +# Modified from Guided-Diffusion (https://github.com/openai/guided-diffusion) +# ------------------------------------------------------------------------------------ + +import math +from abc import abstractmethod + +import torch as th +import torch.nn as nn +import torch.nn.functional as F + +from .nn import ( + avg_pool_nd, + conv_nd, + linear, + normalization, + timestep_embedding, + zero_module, +) +from .xf import LayerNorm + + +class TimestepBlock(nn.Module): + """ + Any module where forward() takes timestep embeddings as a second argument. + """ + + @abstractmethod + def forward(self, x, emb): + """ + Apply the module to `x` given `emb` timestep embeddings. + """ + + +class TimestepEmbedSequential(nn.Sequential, TimestepBlock): + """ + A sequential module that passes timestep embeddings to the children that + support it as an extra input. + """ + + def forward(self, x, emb, encoder_out=None, mask=None): + for layer in self: + if isinstance(layer, TimestepBlock): + x = layer(x, emb) + elif isinstance(layer, AttentionBlock): + x = layer(x, encoder_out, mask=mask) + else: + x = layer(x) + return x + + +class Upsample(nn.Module): + """ + An upsampling layer with an optional convolution. + + :param channels: channels in the inputs and outputs. + :param use_conv: a bool determining if a convolution is applied. + :param dims: determines if the signal is 1D, 2D, or 3D. If 3D, then + upsampling occurs in the inner-two dimensions. + """ + + def __init__(self, channels, use_conv, dims=2, out_channels=None): + super().__init__() + self.channels = channels + self.out_channels = out_channels or channels + self.use_conv = use_conv + self.dims = dims + if use_conv: + self.conv = conv_nd(dims, self.channels, self.out_channels, 3, padding=1) + + def forward(self, x): + assert x.shape[1] == self.channels + if self.dims == 3: + x = F.interpolate( + x, (x.shape[2], x.shape[3] * 2, x.shape[4] * 2), mode="nearest" + ) + else: + x = F.interpolate(x, scale_factor=2, mode="nearest") + if self.use_conv: + x = self.conv(x) + return x + + +class Downsample(nn.Module): + """ + A downsampling layer with an optional convolution. + + :param channels: channels in the inputs and outputs. + :param use_conv: a bool determining if a convolution is applied. + :param dims: determines if the signal is 1D, 2D, or 3D. If 3D, then + downsampling occurs in the inner-two dimensions. + """ + + def __init__(self, channels, use_conv, dims=2, out_channels=None): + super().__init__() + self.channels = channels + self.out_channels = out_channels or channels + self.use_conv = use_conv + self.dims = dims + stride = 2 if dims != 3 else (1, 2, 2) + if use_conv: + self.op = conv_nd( + dims, self.channels, self.out_channels, 3, stride=stride, padding=1 + ) + else: + assert self.channels == self.out_channels + self.op = avg_pool_nd(dims, kernel_size=stride, stride=stride) + + def forward(self, x): + assert x.shape[1] == self.channels + return self.op(x) + + +class ResBlock(TimestepBlock): + """ + A residual block that can optionally change the number of channels. + + :param channels: the number of input channels. + :param emb_channels: the number of timestep embedding channels. + :param dropout: the rate of dropout. + :param out_channels: if specified, the number of out channels. + :param use_conv: if True and out_channels is specified, use a spatial + convolution instead of a smaller 1x1 convolution to change the + channels in the skip connection. + :param dims: determines if the signal is 1D, 2D, or 3D. + :param use_checkpoint: if True, use gradient checkpointing on this module. + :param up: if True, use this block for upsampling. + :param down: if True, use this block for downsampling. + """ + + def __init__( + self, + channels, + emb_channels, + dropout, + out_channels=None, + use_conv=False, + use_scale_shift_norm=False, + dims=2, + use_checkpoint=False, + up=False, + down=False, + ): + super().__init__() + self.channels = channels + self.emb_channels = emb_channels + self.dropout = dropout + self.out_channels = out_channels or channels + self.use_conv = use_conv + self.use_checkpoint = use_checkpoint + self.use_scale_shift_norm = use_scale_shift_norm + + self.in_layers = nn.Sequential( + normalization(channels, swish=1.0), + nn.Identity(), + conv_nd(dims, channels, self.out_channels, 3, padding=1), + ) + + self.updown = up or down + + if up: + self.h_upd = Upsample(channels, False, dims) + self.x_upd = Upsample(channels, False, dims) + elif down: + self.h_upd = Downsample(channels, False, dims) + self.x_upd = Downsample(channels, False, dims) + else: + self.h_upd = self.x_upd = nn.Identity() + + self.emb_layers = nn.Sequential( + nn.SiLU(), + linear( + emb_channels, + 2 * self.out_channels if use_scale_shift_norm else self.out_channels, + ), + ) + self.out_layers = nn.Sequential( + normalization( + self.out_channels, swish=0.0 if use_scale_shift_norm else 1.0 + ), + nn.SiLU() if use_scale_shift_norm else nn.Identity(), + nn.Dropout(p=dropout), + zero_module( + conv_nd(dims, self.out_channels, self.out_channels, 3, padding=1) + ), + ) + + if self.out_channels == channels: + self.skip_connection = nn.Identity() + elif use_conv: + self.skip_connection = conv_nd( + dims, channels, self.out_channels, 3, padding=1 + ) + else: + self.skip_connection = conv_nd(dims, channels, self.out_channels, 1) + + def forward(self, x, emb): + """ + Apply the block to a Tensor, conditioned on a timestep embedding. + + :param x: an [N x C x ...] Tensor of features. + :param emb: an [N x emb_channels] Tensor of timestep embeddings. + :return: an [N x C x ...] Tensor of outputs. + """ + if self.updown: + in_rest, in_conv = self.in_layers[:-1], self.in_layers[-1] + h = in_rest(x) + h = self.h_upd(h) + x = self.x_upd(x) + h = in_conv(h) + else: + h = self.in_layers(x) + emb_out = self.emb_layers(emb) + while len(emb_out.shape) < len(h.shape): + emb_out = emb_out[..., None] + if self.use_scale_shift_norm: + out_norm, out_rest = self.out_layers[0], self.out_layers[1:] + scale, shift = th.chunk(emb_out, 2, dim=1) + h = out_norm(h) * (1 + scale) + shift + h = out_rest(h) + else: + h = h + emb_out + h = self.out_layers(h) + return self.skip_connection(x) + h + + +class ResBlockNoTimeEmbedding(nn.Module): + """ + A residual block without time embedding + + :param channels: the number of input channels. + :param emb_channels: the number of timestep embedding channels. + :param dropout: the rate of dropout. + :param out_channels: if specified, the number of out channels. + :param use_conv: if True and out_channels is specified, use a spatial + convolution instead of a smaller 1x1 convolution to change the + channels in the skip connection. + :param dims: determines if the signal is 1D, 2D, or 3D. + :param use_checkpoint: if True, use gradient checkpointing on this module. + :param up: if True, use this block for upsampling. + :param down: if True, use this block for downsampling. + """ + + def __init__( + self, + channels, + emb_channels, + dropout, + out_channels=None, + use_conv=False, + dims=2, + use_checkpoint=False, + up=False, + down=False, + **kwargs, + ): + super().__init__() + self.channels = channels + self.emb_channels = emb_channels + self.dropout = dropout + self.out_channels = out_channels or channels + self.use_conv = use_conv + self.use_checkpoint = use_checkpoint + + self.in_layers = nn.Sequential( + normalization(channels, swish=1.0), + nn.Identity(), + conv_nd(dims, channels, self.out_channels, 3, padding=1), + ) + + self.updown = up or down + + if up: + self.h_upd = Upsample(channels, False, dims) + self.x_upd = Upsample(channels, False, dims) + elif down: + self.h_upd = Downsample(channels, False, dims) + self.x_upd = Downsample(channels, False, dims) + else: + self.h_upd = self.x_upd = nn.Identity() + + self.out_layers = nn.Sequential( + normalization(self.out_channels, swish=1.0), + nn.Dropout(p=dropout), + zero_module( + conv_nd(dims, self.out_channels, self.out_channels, 3, padding=1) + ), + ) + + if self.out_channels == channels: + self.skip_connection = nn.Identity() + elif use_conv: + self.skip_connection = conv_nd( + dims, channels, self.out_channels, 3, padding=1 + ) + else: + self.skip_connection = conv_nd(dims, channels, self.out_channels, 1) + + def forward(self, x, emb=None): + """ + Apply the block to a Tensor, NOT conditioned on a timestep embedding. + + :param x: an [N x C x ...] Tensor of features. + :param emb: an [N x emb_channels] Tensor of timestep embeddings. + :return: an [N x C x ...] Tensor of outputs. + """ + assert emb is None + + if self.updown: + in_rest, in_conv = self.in_layers[:-1], self.in_layers[-1] + h = in_rest(x) + h = self.h_upd(h) + x = self.x_upd(x) + h = in_conv(h) + else: + h = self.in_layers(x) + h = self.out_layers(h) + return self.skip_connection(x) + h + + +class AttentionBlock(nn.Module): + """ + An attention block that allows spatial positions to attend to each other. + + Originally ported from here, but adapted to the N-d case. + https://github.com/hojonathanho/diffusion/blob/1e0dceb3b3495bbe19116a5e1b3596cd0706c543/diffusion_tf/models/unet.py#L66. + """ + + def __init__( + self, + channels, + num_heads=1, + num_head_channels=-1, + use_checkpoint=False, + encoder_channels=None, + ): + super().__init__() + self.channels = channels + if num_head_channels == -1: + self.num_heads = num_heads + else: + assert ( + channels % num_head_channels == 0 + ), f"q,k,v channels {channels} is not divisible by num_head_channels {num_head_channels}" + self.num_heads = channels // num_head_channels + self.use_checkpoint = use_checkpoint + self.norm = normalization(channels, swish=0.0) + self.qkv = conv_nd(1, channels, channels * 3, 1) + self.attention = QKVAttention(self.num_heads) + + if encoder_channels is not None: + self.encoder_kv = conv_nd(1, encoder_channels, channels * 2, 1) + self.proj_out = zero_module(conv_nd(1, channels, channels, 1)) + + def forward(self, x, encoder_out=None, mask=None): + b, c, *spatial = x.shape + qkv = self.qkv(self.norm(x).view(b, c, -1)) + if encoder_out is not None: + encoder_out = self.encoder_kv(encoder_out) + h = self.attention(qkv, encoder_out, mask=mask) + else: + h = self.attention(qkv) + h = self.proj_out(h) + return x + h.reshape(b, c, *spatial) + + +class QKVAttention(nn.Module): + """ + A module which performs QKV attention. Matches legacy QKVAttention + input/ouput heads shaping + """ + + def __init__(self, n_heads): + super().__init__() + self.n_heads = n_heads + + def forward(self, qkv, encoder_kv=None, mask=None): + """ + Apply QKV attention. + + :param qkv: an [N x (H * 3 * C) x T] tensor of Qs, Ks, and Vs. + :return: an [N x (H * C) x T] tensor after attention. + """ + bs, width, length = qkv.shape + assert width % (3 * self.n_heads) == 0 + ch = width // (3 * self.n_heads) + q, k, v = qkv.reshape(bs * self.n_heads, ch * 3, length).split(ch, dim=1) + if encoder_kv is not None: + assert encoder_kv.shape[1] == self.n_heads * ch * 2 + ek, ev = encoder_kv.reshape(bs * self.n_heads, ch * 2, -1).split(ch, dim=1) + k = th.cat([ek, k], dim=-1) + v = th.cat([ev, v], dim=-1) + scale = 1 / math.sqrt(math.sqrt(ch)) + weight = th.einsum("bct,bcs->bts", q * scale, k * scale) + if mask is not None: + mask = F.pad(mask, (0, length), value=0.0) + mask = ( + mask.unsqueeze(1) + .expand(-1, self.n_heads, -1) + .reshape(bs * self.n_heads, 1, -1) + ) + weight = weight + mask + weight = th.softmax(weight, dim=-1) + a = th.einsum("bts,bcs->bct", weight, v) + return a.reshape(bs, -1, length) + + +class UNetModel(nn.Module): + """ + The full UNet model with attention and timestep embedding. + + :param in_channels: channels in the input Tensor. + :param model_channels: base channel count for the model. + :param out_channels: channels in the output Tensor. + :param num_res_blocks: number of residual blocks per downsample. + :param attention_resolutions: a collection of downsample rates at which + attention will take place. May be a set, list, or tuple. + For example, if this contains 4, then at 4x downsampling, attention + will be used. + :param dropout: the dropout probability. + :param channel_mult: channel multiplier for each level of the UNet. + :param conv_resample: if True, use learned convolutions for upsampling and + downsampling. + :param dims: determines if the signal is 1D, 2D, or 3D. + :param clip_dim: dimension of clip feature. + :param num_classes: if specified (as an int), then this model will be + class-conditional with `num_classes` classes. + :param use_checkpoint: use gradient checkpointing to reduce memory usage. + :param num_heads: the number of attention heads in each attention layer. + :param num_heads_channels: if specified, ignore num_heads and instead use + a fixed channel width per attention head. + :param num_heads_upsample: works with num_heads to set a different number + of heads for upsampling. Deprecated. + :param use_scale_shift_norm: use a FiLM-like conditioning mechanism. + :param resblock_updown: use residual blocks for up/downsampling. + :param encoder_channels: use to make the dimension of query and kv same in AttentionBlock. + :param use_time_embedding: use time embedding for condition. + """ + + def __init__( + self, + in_channels, + model_channels, + out_channels, + num_res_blocks, + attention_resolutions, + dropout=0, + channel_mult=(1, 2, 4, 8), + conv_resample=True, + dims=2, + clip_dim=None, + use_checkpoint=False, + num_heads=1, + num_head_channels=-1, + num_heads_upsample=-1, + use_scale_shift_norm=False, + use_middle_attention=True, + resblock_updown=False, + encoder_channels=None, + use_time_embedding=True, + ): + super().__init__() + + if num_heads_upsample == -1: + num_heads_upsample = num_heads + + self.in_channels = in_channels + self.model_channels = model_channels + self.out_channels = out_channels + self.num_res_blocks = num_res_blocks + self.attention_resolutions = attention_resolutions + self.dropout = dropout + self.channel_mult = channel_mult + self.conv_resample = conv_resample + self.clip_dim = clip_dim + self.use_checkpoint = use_checkpoint + self.num_heads = num_heads + self.num_head_channels = num_head_channels + self.num_heads_upsample = num_heads_upsample + self.use_middle_attention = use_middle_attention + self.use_time_embedding = use_time_embedding + + if self.use_time_embedding: + time_embed_dim = model_channels * 4 + self.time_embed = nn.Sequential( + linear(model_channels, time_embed_dim), + nn.SiLU(), + linear(time_embed_dim, time_embed_dim), + ) + + if self.clip_dim is not None: + self.clip_emb = nn.Linear(clip_dim, time_embed_dim) + else: + time_embed_dim = None + + CustomResidualBlock = ( + ResBlock if self.use_time_embedding else ResBlockNoTimeEmbedding + ) + ch = input_ch = int(channel_mult[0] * model_channels) + self.input_blocks = nn.ModuleList( + [TimestepEmbedSequential(conv_nd(dims, in_channels, ch, 3, padding=1))] + ) + self._feature_size = ch + input_block_chans = [ch] + ds = 1 + for level, mult in enumerate(channel_mult): + for _ in range(num_res_blocks): + layers = [ + CustomResidualBlock( + ch, + time_embed_dim, + dropout, + out_channels=int(mult * model_channels), + dims=dims, + use_checkpoint=use_checkpoint, + use_scale_shift_norm=use_scale_shift_norm, + ) + ] + ch = int(mult * model_channels) + if ds in attention_resolutions: + layers.append( + AttentionBlock( + ch, + use_checkpoint=use_checkpoint, + num_heads=num_heads, + num_head_channels=num_head_channels, + encoder_channels=encoder_channels, + ) + ) + self.input_blocks.append(TimestepEmbedSequential(*layers)) + self._feature_size += ch + input_block_chans.append(ch) + if level != len(channel_mult) - 1: + out_ch = ch + self.input_blocks.append( + TimestepEmbedSequential( + CustomResidualBlock( + ch, + time_embed_dim, + dropout, + out_channels=out_ch, + dims=dims, + use_checkpoint=use_checkpoint, + use_scale_shift_norm=use_scale_shift_norm, + down=True, + ) + if resblock_updown + else Downsample( + ch, conv_resample, dims=dims, out_channels=out_ch + ) + ) + ) + ch = out_ch + input_block_chans.append(ch) + ds *= 2 + self._feature_size += ch + + self.middle_block = TimestepEmbedSequential( + CustomResidualBlock( + ch, + time_embed_dim, + dropout, + dims=dims, + use_checkpoint=use_checkpoint, + use_scale_shift_norm=use_scale_shift_norm, + ), + *( + AttentionBlock( + ch, + use_checkpoint=use_checkpoint, + num_heads=num_heads, + num_head_channels=num_head_channels, + encoder_channels=encoder_channels, + ), + ) + if self.use_middle_attention + else tuple(), # add AttentionBlock or not + CustomResidualBlock( + ch, + time_embed_dim, + dropout, + dims=dims, + use_checkpoint=use_checkpoint, + use_scale_shift_norm=use_scale_shift_norm, + ), + ) + self._feature_size += ch + + self.output_blocks = nn.ModuleList([]) + for level, mult in list(enumerate(channel_mult))[::-1]: + for i in range(num_res_blocks + 1): + ich = input_block_chans.pop() + layers = [ + CustomResidualBlock( + ch + ich, + time_embed_dim, + dropout, + out_channels=int(model_channels * mult), + dims=dims, + use_checkpoint=use_checkpoint, + use_scale_shift_norm=use_scale_shift_norm, + ) + ] + ch = int(model_channels * mult) + if ds in attention_resolutions: + layers.append( + AttentionBlock( + ch, + use_checkpoint=use_checkpoint, + num_heads=num_heads_upsample, + num_head_channels=num_head_channels, + encoder_channels=encoder_channels, + ) + ) + if level and i == num_res_blocks: + out_ch = ch + layers.append( + CustomResidualBlock( + ch, + time_embed_dim, + dropout, + out_channels=out_ch, + dims=dims, + use_checkpoint=use_checkpoint, + use_scale_shift_norm=use_scale_shift_norm, + up=True, + ) + if resblock_updown + else Upsample(ch, conv_resample, dims=dims, out_channels=out_ch) + ) + ds //= 2 + self.output_blocks.append(TimestepEmbedSequential(*layers)) + self._feature_size += ch + + self.out = nn.Sequential( + normalization(ch, swish=1.0), + nn.Identity(), + zero_module(conv_nd(dims, input_ch, out_channels, 3, padding=1)), + ) + + def forward(self, x, timesteps, y=None): + """ + Apply the model to an input batch. + + :param x: an [N x C x ...] Tensor of inputs. + :param timesteps: a 1-D batch of timesteps. + :param y: an [N] Tensor of labels, if class-conditional. + :return: an [N x C x ...] Tensor of outputs. + """ + assert (y is not None) == ( + self.clip_dim is not None + ), "must specify y if and only if the model is clip-rep-conditional" + + hs = [] + if self.use_time_embedding: + emb = self.time_embed(timestep_embedding(timesteps, self.model_channels)) + if self.clip_dim is not None: + emb = emb + self.clip_emb(y) + else: + emb = None + + h = x + for module in self.input_blocks: + h = module(h, emb) + hs.append(h) + h = self.middle_block(h, emb) + for module in self.output_blocks: + h = th.cat([h, hs.pop()], dim=1) + h = module(h, emb) + + return self.out(h) + + +class SuperResUNetModel(UNetModel): + """ + A UNetModel that performs super-resolution. + + Expects an extra kwarg `low_res` to condition on a low-resolution image. + Assumes that the shape of low-resolution and the input should be the same. + """ + + def __init__(self, *args, **kwargs): + if "in_channels" in kwargs: + kwargs = dict(kwargs) + kwargs["in_channels"] = kwargs["in_channels"] * 2 + else: + # Curse you, Python. Or really, just curse positional arguments :|. + args = list(args) + args[1] = args[1] * 2 + super().__init__(*args, **kwargs) + + def forward(self, x, timesteps, low_res=None, **kwargs): + _, _, new_height, new_width = x.shape + assert new_height == low_res.shape[2] and new_width == low_res.shape[3] + + x = th.cat([x, low_res], dim=1) + return super().forward(x, timesteps, **kwargs) + + +class PLMImUNet(UNetModel): + """ + A UNetModel that conditions on text with a pretrained text encoder in CLIP. + + :param text_ctx: number of text tokens to expect. + :param xf_width: width of the transformer. + :param clip_emb_mult: #extra tokens by projecting clip text feature. + :param clip_emb_type: type of condition (here, we fix clip image feature). + :param clip_emb_drop: dropout rato of clip image feature for cfg. + """ + + def __init__( + self, + text_ctx, + xf_width, + *args, + clip_emb_mult=None, + clip_emb_type="image", + clip_emb_drop=0.0, + **kwargs, + ): + self.text_ctx = text_ctx + self.xf_width = xf_width + self.clip_emb_mult = clip_emb_mult + self.clip_emb_type = clip_emb_type + self.clip_emb_drop = clip_emb_drop + + if not xf_width: + super().__init__(*args, **kwargs, encoder_channels=None) + else: + super().__init__(*args, **kwargs, encoder_channels=xf_width) + + # Project text encoded feat seq from pre-trained text encoder in CLIP + self.text_seq_proj = nn.Sequential( + nn.Linear(self.clip_dim, xf_width), + LayerNorm(xf_width), + ) + # Project CLIP text feat + self.text_feat_proj = nn.Linear(self.clip_dim, self.model_channels * 4) + + assert clip_emb_mult is not None + assert clip_emb_type == "image" + assert self.clip_dim is not None, "CLIP representation dim should be specified" + + self.clip_tok_proj = nn.Linear( + self.clip_dim, self.xf_width * self.clip_emb_mult + ) + if self.clip_emb_drop > 0: + self.cf_param = nn.Parameter(th.empty(self.clip_dim, dtype=th.float32)) + + def proc_clip_emb_drop(self, feat): + if self.clip_emb_drop > 0: + bsz, feat_dim = feat.shape + assert ( + feat_dim == self.clip_dim + ), f"CLIP input dim: {feat_dim}, model CLIP dim: {self.clip_dim}" + drop_idx = th.rand((bsz,), device=feat.device) < self.clip_emb_drop + feat = th.where( + drop_idx[..., None], self.cf_param[None].type_as(feat), feat + ) + return feat + + def forward( + self, x, timesteps, txt_feat=None, txt_feat_seq=None, mask=None, y=None + ): + bsz = x.shape[0] + hs = [] + emb = self.time_embed(timestep_embedding(timesteps, self.model_channels)) + emb = emb + self.clip_emb(y) + + xf_out = self.text_seq_proj(txt_feat_seq) + xf_out = xf_out.permute(0, 2, 1) + emb = emb + self.text_feat_proj(txt_feat) + xf_out = th.cat( + [ + self.clip_tok_proj(y).reshape(bsz, -1, self.clip_emb_mult), + xf_out, + ], + dim=2, + ) + mask = F.pad(mask, (self.clip_emb_mult, 0), value=True) + mask = th.where(mask, 0.0, float("-inf")) + + h = x + for module in self.input_blocks: + h = module(h, emb, xf_out, mask=mask) + hs.append(h) + h = self.middle_block(h, emb, xf_out, mask=mask) + for module in self.output_blocks: + h = th.cat([h, hs.pop()], dim=1) + h = module(h, emb, xf_out, mask=mask) + h = self.out(h) + + return h diff --git a/ldm/modules/karlo/kakao/modules/xf.py b/ldm/modules/karlo/kakao/modules/xf.py new file mode 100644 index 0000000000000000000000000000000000000000..66d7d4a2f3fd9fec420ac8272ae00b96c634a3a9 --- /dev/null +++ b/ldm/modules/karlo/kakao/modules/xf.py @@ -0,0 +1,231 @@ +# ------------------------------------------------------------------------------------ +# Adapted from the repos below: +# (a) Guided-Diffusion (https://github.com/openai/guided-diffusion) +# (b) CLIP ViT (https://github.com/openai/CLIP/) +# ------------------------------------------------------------------------------------ + +import math + +import torch as th +import torch.nn as nn +import torch.nn.functional as F + +from .nn import timestep_embedding + + +def convert_module_to_f16(param): + """ + Convert primitive modules to float16. + """ + if isinstance(param, (nn.Linear, nn.Conv2d, nn.ConvTranspose2d)): + param.weight.data = param.weight.data.half() + if param.bias is not None: + param.bias.data = param.bias.data.half() + + +class LayerNorm(nn.LayerNorm): + """ + Implementation that supports fp16 inputs but fp32 gains/biases. + """ + + def forward(self, x: th.Tensor): + return super().forward(x.float()).to(x.dtype) + + +class MultiheadAttention(nn.Module): + def __init__(self, n_ctx, width, heads): + super().__init__() + self.n_ctx = n_ctx + self.width = width + self.heads = heads + self.c_qkv = nn.Linear(width, width * 3) + self.c_proj = nn.Linear(width, width) + self.attention = QKVMultiheadAttention(heads, n_ctx) + + def forward(self, x, mask=None): + x = self.c_qkv(x) + x = self.attention(x, mask=mask) + x = self.c_proj(x) + return x + + +class MLP(nn.Module): + def __init__(self, width): + super().__init__() + self.width = width + self.c_fc = nn.Linear(width, width * 4) + self.c_proj = nn.Linear(width * 4, width) + self.gelu = nn.GELU() + + def forward(self, x): + return self.c_proj(self.gelu(self.c_fc(x))) + + +class QKVMultiheadAttention(nn.Module): + def __init__(self, n_heads: int, n_ctx: int): + super().__init__() + self.n_heads = n_heads + self.n_ctx = n_ctx + + def forward(self, qkv, mask=None): + bs, n_ctx, width = qkv.shape + attn_ch = width // self.n_heads // 3 + scale = 1 / math.sqrt(math.sqrt(attn_ch)) + qkv = qkv.view(bs, n_ctx, self.n_heads, -1) + q, k, v = th.split(qkv, attn_ch, dim=-1) + weight = th.einsum("bthc,bshc->bhts", q * scale, k * scale) + wdtype = weight.dtype + if mask is not None: + weight = weight + mask[:, None, ...] + weight = th.softmax(weight, dim=-1).type(wdtype) + return th.einsum("bhts,bshc->bthc", weight, v).reshape(bs, n_ctx, -1) + + +class ResidualAttentionBlock(nn.Module): + def __init__( + self, + n_ctx: int, + width: int, + heads: int, + ): + super().__init__() + + self.attn = MultiheadAttention( + n_ctx, + width, + heads, + ) + self.ln_1 = LayerNorm(width) + self.mlp = MLP(width) + self.ln_2 = LayerNorm(width) + + def forward(self, x, mask=None): + x = x + self.attn(self.ln_1(x), mask=mask) + x = x + self.mlp(self.ln_2(x)) + return x + + +class Transformer(nn.Module): + def __init__( + self, + n_ctx: int, + width: int, + layers: int, + heads: int, + ): + super().__init__() + self.n_ctx = n_ctx + self.width = width + self.layers = layers + self.resblocks = nn.ModuleList( + [ + ResidualAttentionBlock( + n_ctx, + width, + heads, + ) + for _ in range(layers) + ] + ) + + def forward(self, x, mask=None): + for block in self.resblocks: + x = block(x, mask=mask) + return x + + +class PriorTransformer(nn.Module): + """ + A Causal Transformer that conditions on CLIP text embedding, text. + + :param text_ctx: number of text tokens to expect. + :param xf_width: width of the transformer. + :param xf_layers: depth of the transformer. + :param xf_heads: heads in the transformer. + :param xf_final_ln: use a LayerNorm after the output layer. + :param clip_dim: dimension of clip feature. + """ + + def __init__( + self, + text_ctx, + xf_width, + xf_layers, + xf_heads, + xf_final_ln, + clip_dim, + ): + super().__init__() + + self.text_ctx = text_ctx + self.xf_width = xf_width + self.xf_layers = xf_layers + self.xf_heads = xf_heads + self.clip_dim = clip_dim + self.ext_len = 4 + + self.time_embed = nn.Sequential( + nn.Linear(xf_width, xf_width), + nn.SiLU(), + nn.Linear(xf_width, xf_width), + ) + self.text_enc_proj = nn.Linear(clip_dim, xf_width) + self.text_emb_proj = nn.Linear(clip_dim, xf_width) + self.clip_img_proj = nn.Linear(clip_dim, xf_width) + self.out_proj = nn.Linear(xf_width, clip_dim) + self.transformer = Transformer( + text_ctx + self.ext_len, + xf_width, + xf_layers, + xf_heads, + ) + if xf_final_ln: + self.final_ln = LayerNorm(xf_width) + else: + self.final_ln = None + + self.positional_embedding = nn.Parameter( + th.empty(1, text_ctx + self.ext_len, xf_width) + ) + self.prd_emb = nn.Parameter(th.randn((1, 1, xf_width))) + + nn.init.normal_(self.prd_emb, std=0.01) + nn.init.normal_(self.positional_embedding, std=0.01) + + def forward( + self, + x, + timesteps, + text_emb=None, + text_enc=None, + mask=None, + causal_mask=None, + ): + bsz = x.shape[0] + mask = F.pad(mask, (0, self.ext_len), value=True) + + t_emb = self.time_embed(timestep_embedding(timesteps, self.xf_width)) + text_enc = self.text_enc_proj(text_enc) + text_emb = self.text_emb_proj(text_emb) + x = self.clip_img_proj(x) + + input_seq = [ + text_enc, + text_emb[:, None, :], + t_emb[:, None, :], + x[:, None, :], + self.prd_emb.to(x.dtype).expand(bsz, -1, -1), + ] + input = th.cat(input_seq, dim=1) + input = input + self.positional_embedding.to(input.dtype) + + mask = th.where(mask, 0.0, float("-inf")) + mask = (mask[:, None, :] + causal_mask).to(input.dtype) + + out = self.transformer(input, mask=mask) + if self.final_ln is not None: + out = self.final_ln(out) + + out = self.out_proj(out[:, -1]) + + return out diff --git a/ldm/modules/karlo/kakao/sampler.py b/ldm/modules/karlo/kakao/sampler.py new file mode 100644 index 0000000000000000000000000000000000000000..b56bf2f20c280d735a2be10af03d7cd3e7bf85ac --- /dev/null +++ b/ldm/modules/karlo/kakao/sampler.py @@ -0,0 +1,272 @@ +# ------------------------------------------------------------------------------------ +# Karlo-v1.0.alpha +# Copyright (c) 2022 KakaoBrain. All Rights Reserved. + +# source: https://github.com/kakaobrain/karlo/blob/3c68a50a16d76b48a15c181d1c5a5e0879a90f85/karlo/sampler/t2i.py#L15 +# ------------------------------------------------------------------------------------ + +from typing import Iterator + +import torch +import torchvision.transforms.functional as TVF +from torchvision.transforms import InterpolationMode + +from .template import BaseSampler, CKPT_PATH + + +class T2ISampler(BaseSampler): + """ + A sampler for text-to-image generation. + :param root_dir: directory for model checkpoints. + :param sampling_type: ["default", "fast"] + """ + + def __init__( + self, + root_dir: str, + sampling_type: str = "default", + ): + super().__init__(root_dir, sampling_type) + + @classmethod + def from_pretrained( + cls, + root_dir: str, + clip_model_path: str, + clip_stat_path: str, + sampling_type: str = "default", + ): + + model = cls( + root_dir=root_dir, + sampling_type=sampling_type, + ) + model.load_clip(clip_model_path) + model.load_prior( + f"{CKPT_PATH['prior']}", + clip_stat_path=clip_stat_path, + prior_config="configs/karlo/prior_1B_vit_l.yaml" + ) + model.load_decoder(f"{CKPT_PATH['decoder']}", decoder_config="configs/karlo/decoder_900M_vit_l.yaml") + model.load_sr_64_256(CKPT_PATH["sr_256"], sr_config="configs/karlo/improved_sr_64_256_1.4B.yaml") + return model + + def preprocess( + self, + prompt: str, + bsz: int, + ): + """Setup prompts & cfg scales""" + prompts_batch = [prompt for _ in range(bsz)] + + prior_cf_scales_batch = [self._prior_cf_scale] * len(prompts_batch) + prior_cf_scales_batch = torch.tensor(prior_cf_scales_batch, device="cuda") + + decoder_cf_scales_batch = [self._decoder_cf_scale] * len(prompts_batch) + decoder_cf_scales_batch = torch.tensor(decoder_cf_scales_batch, device="cuda") + + """ Get CLIP text feature """ + clip_model = self._clip + tokenizer = self._tokenizer + max_txt_length = self._prior.model.text_ctx + + tok, mask = tokenizer.padded_tokens_and_mask(prompts_batch, max_txt_length) + cf_token, cf_mask = tokenizer.padded_tokens_and_mask([""], max_txt_length) + if not (cf_token.shape == tok.shape): + cf_token = cf_token.expand(tok.shape[0], -1) + cf_mask = cf_mask.expand(tok.shape[0], -1) + + tok = torch.cat([tok, cf_token], dim=0) + mask = torch.cat([mask, cf_mask], dim=0) + + tok, mask = tok.to(device="cuda"), mask.to(device="cuda") + txt_feat, txt_feat_seq = clip_model.encode_text(tok) + + return ( + prompts_batch, + prior_cf_scales_batch, + decoder_cf_scales_batch, + txt_feat, + txt_feat_seq, + tok, + mask, + ) + + def __call__( + self, + prompt: str, + bsz: int, + progressive_mode=None, + ) -> Iterator[torch.Tensor]: + assert progressive_mode in ("loop", "stage", "final") + with torch.no_grad(), torch.cuda.amp.autocast(): + ( + prompts_batch, + prior_cf_scales_batch, + decoder_cf_scales_batch, + txt_feat, + txt_feat_seq, + tok, + mask, + ) = self.preprocess( + prompt, + bsz, + ) + + """ Transform CLIP text feature into image feature """ + img_feat = self._prior( + txt_feat, + txt_feat_seq, + mask, + prior_cf_scales_batch, + timestep_respacing=self._prior_sm, + ) + + """ Generate 64x64px images """ + images_64_outputs = self._decoder( + txt_feat, + txt_feat_seq, + tok, + mask, + img_feat, + cf_guidance_scales=decoder_cf_scales_batch, + timestep_respacing=self._decoder_sm, + ) + + images_64 = None + for k, out in enumerate(images_64_outputs): + images_64 = out + if progressive_mode == "loop": + yield torch.clamp(out * 0.5 + 0.5, 0.0, 1.0) + if progressive_mode == "stage": + yield torch.clamp(out * 0.5 + 0.5, 0.0, 1.0) + + images_64 = torch.clamp(images_64, -1, 1) + + """ Upsample 64x64 to 256x256 """ + images_256 = TVF.resize( + images_64, + [256, 256], + interpolation=InterpolationMode.BICUBIC, + antialias=True, + ) + images_256_outputs = self._sr_64_256( + images_256, timestep_respacing=self._sr_sm + ) + + for k, out in enumerate(images_256_outputs): + images_256 = out + if progressive_mode == "loop": + yield torch.clamp(out * 0.5 + 0.5, 0.0, 1.0) + if progressive_mode == "stage": + yield torch.clamp(out * 0.5 + 0.5, 0.0, 1.0) + + yield torch.clamp(images_256 * 0.5 + 0.5, 0.0, 1.0) + + +class PriorSampler(BaseSampler): + """ + A sampler for text-to-image generation, but only the prior. + :param root_dir: directory for model checkpoints. + :param sampling_type: ["default", "fast"] + """ + + def __init__( + self, + root_dir: str, + sampling_type: str = "default", + ): + super().__init__(root_dir, sampling_type) + + @classmethod + def from_pretrained( + cls, + root_dir: str, + clip_model_path: str, + clip_stat_path: str, + sampling_type: str = "default", + ): + model = cls( + root_dir=root_dir, + sampling_type=sampling_type, + ) + model.load_clip(clip_model_path) + model.load_prior( + f"{CKPT_PATH['prior']}", + clip_stat_path=clip_stat_path, + prior_config="configs/karlo/prior_1B_vit_l.yaml" + ) + return model + + def preprocess( + self, + prompt: str, + bsz: int, + ): + """Setup prompts & cfg scales""" + prompts_batch = [prompt for _ in range(bsz)] + + prior_cf_scales_batch = [self._prior_cf_scale] * len(prompts_batch) + prior_cf_scales_batch = torch.tensor(prior_cf_scales_batch, device="cuda") + + decoder_cf_scales_batch = [self._decoder_cf_scale] * len(prompts_batch) + decoder_cf_scales_batch = torch.tensor(decoder_cf_scales_batch, device="cuda") + + """ Get CLIP text feature """ + clip_model = self._clip + tokenizer = self._tokenizer + max_txt_length = self._prior.model.text_ctx + + tok, mask = tokenizer.padded_tokens_and_mask(prompts_batch, max_txt_length) + cf_token, cf_mask = tokenizer.padded_tokens_and_mask([""], max_txt_length) + if not (cf_token.shape == tok.shape): + cf_token = cf_token.expand(tok.shape[0], -1) + cf_mask = cf_mask.expand(tok.shape[0], -1) + + tok = torch.cat([tok, cf_token], dim=0) + mask = torch.cat([mask, cf_mask], dim=0) + + tok, mask = tok.to(device="cuda"), mask.to(device="cuda") + txt_feat, txt_feat_seq = clip_model.encode_text(tok) + + return ( + prompts_batch, + prior_cf_scales_batch, + decoder_cf_scales_batch, + txt_feat, + txt_feat_seq, + tok, + mask, + ) + + def __call__( + self, + prompt: str, + bsz: int, + progressive_mode=None, + ) -> Iterator[torch.Tensor]: + assert progressive_mode in ("loop", "stage", "final") + with torch.no_grad(), torch.cuda.amp.autocast(): + ( + prompts_batch, + prior_cf_scales_batch, + decoder_cf_scales_batch, + txt_feat, + txt_feat_seq, + tok, + mask, + ) = self.preprocess( + prompt, + bsz, + ) + + """ Transform CLIP text feature into image feature """ + img_feat = self._prior( + txt_feat, + txt_feat_seq, + mask, + prior_cf_scales_batch, + timestep_respacing=self._prior_sm, + ) + + yield img_feat diff --git a/ldm/modules/karlo/kakao/template.py b/ldm/modules/karlo/kakao/template.py new file mode 100644 index 0000000000000000000000000000000000000000..949e80e67b6a8a7c7bbd2c7baad5ec8c8c7e2fdf --- /dev/null +++ b/ldm/modules/karlo/kakao/template.py @@ -0,0 +1,141 @@ +# ------------------------------------------------------------------------------------ +# Karlo-v1.0.alpha +# Copyright (c) 2022 KakaoBrain. All Rights Reserved. +# ------------------------------------------------------------------------------------ + +import os +import logging +import torch + +from omegaconf import OmegaConf + +from ldm.modules.karlo.kakao.models.clip import CustomizedCLIP, CustomizedTokenizer +from ldm.modules.karlo.kakao.models.prior_model import PriorDiffusionModel +from ldm.modules.karlo.kakao.models.decoder_model import Text2ImProgressiveModel +from ldm.modules.karlo.kakao.models.sr_64_256 import ImprovedSupRes64to256ProgressiveModel + + +SAMPLING_CONF = { + "default": { + "prior_sm": "25", + "prior_n_samples": 1, + "prior_cf_scale": 4.0, + "decoder_sm": "50", + "decoder_cf_scale": 8.0, + "sr_sm": "7", + }, + "fast": { + "prior_sm": "25", + "prior_n_samples": 1, + "prior_cf_scale": 4.0, + "decoder_sm": "25", + "decoder_cf_scale": 8.0, + "sr_sm": "7", + }, +} + +CKPT_PATH = { + "prior": "prior-ckpt-step=01000000-of-01000000.ckpt", + "decoder": "decoder-ckpt-step=01000000-of-01000000.ckpt", + "sr_256": "improved-sr-ckpt-step=1.2M.ckpt", +} + + +class BaseSampler: + _PRIOR_CLASS = PriorDiffusionModel + _DECODER_CLASS = Text2ImProgressiveModel + _SR256_CLASS = ImprovedSupRes64to256ProgressiveModel + + def __init__( + self, + root_dir: str, + sampling_type: str = "fast", + ): + self._root_dir = root_dir + + sampling_type = SAMPLING_CONF[sampling_type] + self._prior_sm = sampling_type["prior_sm"] + self._prior_n_samples = sampling_type["prior_n_samples"] + self._prior_cf_scale = sampling_type["prior_cf_scale"] + + assert self._prior_n_samples == 1 + + self._decoder_sm = sampling_type["decoder_sm"] + self._decoder_cf_scale = sampling_type["decoder_cf_scale"] + + self._sr_sm = sampling_type["sr_sm"] + + def __repr__(self): + line = "" + line += f"Prior, sampling method: {self._prior_sm}, cf_scale: {self._prior_cf_scale}\n" + line += f"Decoder, sampling method: {self._decoder_sm}, cf_scale: {self._decoder_cf_scale}\n" + line += f"SR(64->256), sampling method: {self._sr_sm}" + + return line + + def load_clip(self, clip_path: str): + clip = CustomizedCLIP.load_from_checkpoint( + os.path.join(self._root_dir, clip_path) + ) + clip = torch.jit.script(clip) + clip.cuda() + clip.eval() + + self._clip = clip + self._tokenizer = CustomizedTokenizer() + + def load_prior( + self, + ckpt_path: str, + clip_stat_path: str, + prior_config: str = "configs/prior_1B_vit_l.yaml" + ): + logging.info(f"Loading prior: {ckpt_path}") + + config = OmegaConf.load(prior_config) + clip_mean, clip_std = torch.load( + os.path.join(self._root_dir, clip_stat_path), map_location="cpu" + ) + + prior = self._PRIOR_CLASS.load_from_checkpoint( + config, + self._tokenizer, + clip_mean, + clip_std, + os.path.join(self._root_dir, ckpt_path), + strict=True, + ) + prior.cuda() + prior.eval() + logging.info("done.") + + self._prior = prior + + def load_decoder(self, ckpt_path: str, decoder_config: str = "configs/decoder_900M_vit_l.yaml"): + logging.info(f"Loading decoder: {ckpt_path}") + + config = OmegaConf.load(decoder_config) + decoder = self._DECODER_CLASS.load_from_checkpoint( + config, + self._tokenizer, + os.path.join(self._root_dir, ckpt_path), + strict=True, + ) + decoder.cuda() + decoder.eval() + logging.info("done.") + + self._decoder = decoder + + def load_sr_64_256(self, ckpt_path: str, sr_config: str = "configs/improved_sr_64_256_1.4B.yaml"): + logging.info(f"Loading SR(64->256): {ckpt_path}") + + config = OmegaConf.load(sr_config) + sr = self._SR256_CLASS.load_from_checkpoint( + config, os.path.join(self._root_dir, ckpt_path), strict=True + ) + sr.cuda() + sr.eval() + logging.info("done.") + + self._sr_64_256 = sr \ No newline at end of file diff --git a/ldm/modules/midas/__init__.py b/ldm/modules/midas/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/ldm/modules/midas/api.py b/ldm/modules/midas/api.py new file mode 100644 index 0000000000000000000000000000000000000000..b58ebbffd942a2fc22264f0ab47e400c26b9f41c --- /dev/null +++ b/ldm/modules/midas/api.py @@ -0,0 +1,170 @@ +# based on https://github.com/isl-org/MiDaS + +import cv2 +import torch +import torch.nn as nn +from torchvision.transforms import Compose + +from ldm.modules.midas.midas.dpt_depth import DPTDepthModel +from ldm.modules.midas.midas.midas_net import MidasNet +from ldm.modules.midas.midas.midas_net_custom import MidasNet_small +from ldm.modules.midas.midas.transforms import Resize, NormalizeImage, PrepareForNet + + +ISL_PATHS = { + "dpt_large": "midas_models/dpt_large-midas-2f21e586.pt", + "dpt_hybrid": "midas_models/dpt_hybrid-midas-501f0c75.pt", + "midas_v21": "", + "midas_v21_small": "", +} + + +def disabled_train(self, mode=True): + """Overwrite model.train with this function to make sure train/eval mode + does not change anymore.""" + return self + + +def load_midas_transform(model_type): + # https://github.com/isl-org/MiDaS/blob/master/run.py + # load transform only + if model_type == "dpt_large": # DPT-Large + net_w, net_h = 384, 384 + resize_mode = "minimal" + normalization = NormalizeImage(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]) + + elif model_type == "dpt_hybrid": # DPT-Hybrid + net_w, net_h = 384, 384 + resize_mode = "minimal" + normalization = NormalizeImage(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]) + + elif model_type == "midas_v21": + net_w, net_h = 384, 384 + resize_mode = "upper_bound" + normalization = NormalizeImage(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) + + elif model_type == "midas_v21_small": + net_w, net_h = 256, 256 + resize_mode = "upper_bound" + normalization = NormalizeImage(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) + + else: + assert False, f"model_type '{model_type}' not implemented, use: --model_type large" + + transform = Compose( + [ + Resize( + net_w, + net_h, + resize_target=None, + keep_aspect_ratio=True, + ensure_multiple_of=32, + resize_method=resize_mode, + image_interpolation_method=cv2.INTER_CUBIC, + ), + normalization, + PrepareForNet(), + ] + ) + + return transform + + +def load_model(model_type): + # https://github.com/isl-org/MiDaS/blob/master/run.py + # load network + model_path = ISL_PATHS[model_type] + if model_type == "dpt_large": # DPT-Large + model = DPTDepthModel( + path=model_path, + backbone="vitl16_384", + non_negative=True, + ) + net_w, net_h = 384, 384 + resize_mode = "minimal" + normalization = NormalizeImage(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]) + + elif model_type == "dpt_hybrid": # DPT-Hybrid + model = DPTDepthModel( + path=model_path, + backbone="vitb_rn50_384", + non_negative=True, + ) + net_w, net_h = 384, 384 + resize_mode = "minimal" + normalization = NormalizeImage(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]) + + elif model_type == "midas_v21": + model = MidasNet(model_path, non_negative=True) + net_w, net_h = 384, 384 + resize_mode = "upper_bound" + normalization = NormalizeImage( + mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225] + ) + + elif model_type == "midas_v21_small": + model = MidasNet_small(model_path, features=64, backbone="efficientnet_lite3", exportable=True, + non_negative=True, blocks={'expand': True}) + net_w, net_h = 256, 256 + resize_mode = "upper_bound" + normalization = NormalizeImage( + mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225] + ) + + else: + print(f"model_type '{model_type}' not implemented, use: --model_type large") + assert False + + transform = Compose( + [ + Resize( + net_w, + net_h, + resize_target=None, + keep_aspect_ratio=True, + ensure_multiple_of=32, + resize_method=resize_mode, + image_interpolation_method=cv2.INTER_CUBIC, + ), + normalization, + PrepareForNet(), + ] + ) + + return model.eval(), transform + + +class MiDaSInference(nn.Module): + MODEL_TYPES_TORCH_HUB = [ + "DPT_Large", + "DPT_Hybrid", + "MiDaS_small" + ] + MODEL_TYPES_ISL = [ + "dpt_large", + "dpt_hybrid", + "midas_v21", + "midas_v21_small", + ] + + def __init__(self, model_type): + super().__init__() + assert (model_type in self.MODEL_TYPES_ISL) + model, _ = load_model(model_type) + self.model = model + self.model.train = disabled_train + + def forward(self, x): + # x in 0..1 as produced by calling self.transform on a 0..1 float64 numpy array + # NOTE: we expect that the correct transform has been called during dataloading. + with torch.no_grad(): + prediction = self.model(x) + prediction = torch.nn.functional.interpolate( + prediction.unsqueeze(1), + size=x.shape[2:], + mode="bicubic", + align_corners=False, + ) + assert prediction.shape == (x.shape[0], 1, x.shape[2], x.shape[3]) + return prediction + diff --git a/ldm/modules/midas/midas/__init__.py b/ldm/modules/midas/midas/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/ldm/modules/midas/midas/base_model.py b/ldm/modules/midas/midas/base_model.py new file mode 100644 index 0000000000000000000000000000000000000000..5cf430239b47ec5ec07531263f26f5c24a2311cd --- /dev/null +++ b/ldm/modules/midas/midas/base_model.py @@ -0,0 +1,16 @@ +import torch + + +class BaseModel(torch.nn.Module): + def load(self, path): + """Load model from file. + + Args: + path (str): file path + """ + parameters = torch.load(path, map_location=torch.device('cpu')) + + if "optimizer" in parameters: + parameters = parameters["model"] + + self.load_state_dict(parameters) diff --git a/ldm/modules/midas/midas/blocks.py b/ldm/modules/midas/midas/blocks.py new file mode 100644 index 0000000000000000000000000000000000000000..2145d18fa98060a618536d9a64fe6589e9be4f78 --- /dev/null +++ b/ldm/modules/midas/midas/blocks.py @@ -0,0 +1,342 @@ +import torch +import torch.nn as nn + +from .vit import ( + _make_pretrained_vitb_rn50_384, + _make_pretrained_vitl16_384, + _make_pretrained_vitb16_384, + forward_vit, +) + +def _make_encoder(backbone, features, use_pretrained, groups=1, expand=False, exportable=True, hooks=None, use_vit_only=False, use_readout="ignore",): + if backbone == "vitl16_384": + pretrained = _make_pretrained_vitl16_384( + use_pretrained, hooks=hooks, use_readout=use_readout + ) + scratch = _make_scratch( + [256, 512, 1024, 1024], features, groups=groups, expand=expand + ) # ViT-L/16 - 85.0% Top1 (backbone) + elif backbone == "vitb_rn50_384": + pretrained = _make_pretrained_vitb_rn50_384( + use_pretrained, + hooks=hooks, + use_vit_only=use_vit_only, + use_readout=use_readout, + ) + scratch = _make_scratch( + [256, 512, 768, 768], features, groups=groups, expand=expand + ) # ViT-H/16 - 85.0% Top1 (backbone) + elif backbone == "vitb16_384": + pretrained = _make_pretrained_vitb16_384( + use_pretrained, hooks=hooks, use_readout=use_readout + ) + scratch = _make_scratch( + [96, 192, 384, 768], features, groups=groups, expand=expand + ) # ViT-B/16 - 84.6% Top1 (backbone) + elif backbone == "resnext101_wsl": + pretrained = _make_pretrained_resnext101_wsl(use_pretrained) + scratch = _make_scratch([256, 512, 1024, 2048], features, groups=groups, expand=expand) # efficientnet_lite3 + elif backbone == "efficientnet_lite3": + pretrained = _make_pretrained_efficientnet_lite3(use_pretrained, exportable=exportable) + scratch = _make_scratch([32, 48, 136, 384], features, groups=groups, expand=expand) # efficientnet_lite3 + else: + print(f"Backbone '{backbone}' not implemented") + assert False + + return pretrained, scratch + + +def _make_scratch(in_shape, out_shape, groups=1, expand=False): + scratch = nn.Module() + + out_shape1 = out_shape + out_shape2 = out_shape + out_shape3 = out_shape + out_shape4 = out_shape + if expand==True: + out_shape1 = out_shape + out_shape2 = out_shape*2 + out_shape3 = out_shape*4 + out_shape4 = out_shape*8 + + scratch.layer1_rn = nn.Conv2d( + in_shape[0], out_shape1, kernel_size=3, stride=1, padding=1, bias=False, groups=groups + ) + scratch.layer2_rn = nn.Conv2d( + in_shape[1], out_shape2, kernel_size=3, stride=1, padding=1, bias=False, groups=groups + ) + scratch.layer3_rn = nn.Conv2d( + in_shape[2], out_shape3, kernel_size=3, stride=1, padding=1, bias=False, groups=groups + ) + scratch.layer4_rn = nn.Conv2d( + in_shape[3], out_shape4, kernel_size=3, stride=1, padding=1, bias=False, groups=groups + ) + + return scratch + + +def _make_pretrained_efficientnet_lite3(use_pretrained, exportable=False): + efficientnet = torch.hub.load( + "rwightman/gen-efficientnet-pytorch", + "tf_efficientnet_lite3", + pretrained=use_pretrained, + exportable=exportable + ) + return _make_efficientnet_backbone(efficientnet) + + +def _make_efficientnet_backbone(effnet): + pretrained = nn.Module() + + pretrained.layer1 = nn.Sequential( + effnet.conv_stem, effnet.bn1, effnet.act1, *effnet.blocks[0:2] + ) + pretrained.layer2 = nn.Sequential(*effnet.blocks[2:3]) + pretrained.layer3 = nn.Sequential(*effnet.blocks[3:5]) + pretrained.layer4 = nn.Sequential(*effnet.blocks[5:9]) + + return pretrained + + +def _make_resnet_backbone(resnet): + pretrained = nn.Module() + pretrained.layer1 = nn.Sequential( + resnet.conv1, resnet.bn1, resnet.relu, resnet.maxpool, resnet.layer1 + ) + + pretrained.layer2 = resnet.layer2 + pretrained.layer3 = resnet.layer3 + pretrained.layer4 = resnet.layer4 + + return pretrained + + +def _make_pretrained_resnext101_wsl(use_pretrained): + resnet = torch.hub.load("facebookresearch/WSL-Images", "resnext101_32x8d_wsl") + return _make_resnet_backbone(resnet) + + + +class Interpolate(nn.Module): + """Interpolation module. + """ + + def __init__(self, scale_factor, mode, align_corners=False): + """Init. + + Args: + scale_factor (float): scaling + mode (str): interpolation mode + """ + super(Interpolate, self).__init__() + + self.interp = nn.functional.interpolate + self.scale_factor = scale_factor + self.mode = mode + self.align_corners = align_corners + + def forward(self, x): + """Forward pass. + + Args: + x (tensor): input + + Returns: + tensor: interpolated data + """ + + x = self.interp( + x, scale_factor=self.scale_factor, mode=self.mode, align_corners=self.align_corners + ) + + return x + + +class ResidualConvUnit(nn.Module): + """Residual convolution module. + """ + + def __init__(self, features): + """Init. + + Args: + features (int): number of features + """ + super().__init__() + + self.conv1 = nn.Conv2d( + features, features, kernel_size=3, stride=1, padding=1, bias=True + ) + + self.conv2 = nn.Conv2d( + features, features, kernel_size=3, stride=1, padding=1, bias=True + ) + + self.relu = nn.ReLU(inplace=True) + + def forward(self, x): + """Forward pass. + + Args: + x (tensor): input + + Returns: + tensor: output + """ + out = self.relu(x) + out = self.conv1(out) + out = self.relu(out) + out = self.conv2(out) + + return out + x + + +class FeatureFusionBlock(nn.Module): + """Feature fusion block. + """ + + def __init__(self, features): + """Init. + + Args: + features (int): number of features + """ + super(FeatureFusionBlock, self).__init__() + + self.resConfUnit1 = ResidualConvUnit(features) + self.resConfUnit2 = ResidualConvUnit(features) + + def forward(self, *xs): + """Forward pass. + + Returns: + tensor: output + """ + output = xs[0] + + if len(xs) == 2: + output += self.resConfUnit1(xs[1]) + + output = self.resConfUnit2(output) + + output = nn.functional.interpolate( + output, scale_factor=2, mode="bilinear", align_corners=True + ) + + return output + + + + +class ResidualConvUnit_custom(nn.Module): + """Residual convolution module. + """ + + def __init__(self, features, activation, bn): + """Init. + + Args: + features (int): number of features + """ + super().__init__() + + self.bn = bn + + self.groups=1 + + self.conv1 = nn.Conv2d( + features, features, kernel_size=3, stride=1, padding=1, bias=True, groups=self.groups + ) + + self.conv2 = nn.Conv2d( + features, features, kernel_size=3, stride=1, padding=1, bias=True, groups=self.groups + ) + + if self.bn==True: + self.bn1 = nn.BatchNorm2d(features) + self.bn2 = nn.BatchNorm2d(features) + + self.activation = activation + + self.skip_add = nn.quantized.FloatFunctional() + + def forward(self, x): + """Forward pass. + + Args: + x (tensor): input + + Returns: + tensor: output + """ + + out = self.activation(x) + out = self.conv1(out) + if self.bn==True: + out = self.bn1(out) + + out = self.activation(out) + out = self.conv2(out) + if self.bn==True: + out = self.bn2(out) + + if self.groups > 1: + out = self.conv_merge(out) + + return self.skip_add.add(out, x) + + # return out + x + + +class FeatureFusionBlock_custom(nn.Module): + """Feature fusion block. + """ + + def __init__(self, features, activation, deconv=False, bn=False, expand=False, align_corners=True): + """Init. + + Args: + features (int): number of features + """ + super(FeatureFusionBlock_custom, self).__init__() + + self.deconv = deconv + self.align_corners = align_corners + + self.groups=1 + + self.expand = expand + out_features = features + if self.expand==True: + out_features = features//2 + + self.out_conv = nn.Conv2d(features, out_features, kernel_size=1, stride=1, padding=0, bias=True, groups=1) + + self.resConfUnit1 = ResidualConvUnit_custom(features, activation, bn) + self.resConfUnit2 = ResidualConvUnit_custom(features, activation, bn) + + self.skip_add = nn.quantized.FloatFunctional() + + def forward(self, *xs): + """Forward pass. + + Returns: + tensor: output + """ + output = xs[0] + + if len(xs) == 2: + res = self.resConfUnit1(xs[1]) + output = self.skip_add.add(output, res) + # output += res + + output = self.resConfUnit2(output) + + output = nn.functional.interpolate( + output, scale_factor=2, mode="bilinear", align_corners=self.align_corners + ) + + output = self.out_conv(output) + + return output + diff --git a/ldm/modules/midas/midas/dpt_depth.py b/ldm/modules/midas/midas/dpt_depth.py new file mode 100644 index 0000000000000000000000000000000000000000..4e9aab5d2767dffea39da5b3f30e2798688216f1 --- /dev/null +++ b/ldm/modules/midas/midas/dpt_depth.py @@ -0,0 +1,109 @@ +import torch +import torch.nn as nn +import torch.nn.functional as F + +from .base_model import BaseModel +from .blocks import ( + FeatureFusionBlock, + FeatureFusionBlock_custom, + Interpolate, + _make_encoder, + forward_vit, +) + + +def _make_fusion_block(features, use_bn): + return FeatureFusionBlock_custom( + features, + nn.ReLU(False), + deconv=False, + bn=use_bn, + expand=False, + align_corners=True, + ) + + +class DPT(BaseModel): + def __init__( + self, + head, + features=256, + backbone="vitb_rn50_384", + readout="project", + channels_last=False, + use_bn=False, + ): + + super(DPT, self).__init__() + + self.channels_last = channels_last + + hooks = { + "vitb_rn50_384": [0, 1, 8, 11], + "vitb16_384": [2, 5, 8, 11], + "vitl16_384": [5, 11, 17, 23], + } + + # Instantiate backbone and reassemble blocks + self.pretrained, self.scratch = _make_encoder( + backbone, + features, + False, # Set to true of you want to train from scratch, uses ImageNet weights + groups=1, + expand=False, + exportable=False, + hooks=hooks[backbone], + use_readout=readout, + ) + + self.scratch.refinenet1 = _make_fusion_block(features, use_bn) + self.scratch.refinenet2 = _make_fusion_block(features, use_bn) + self.scratch.refinenet3 = _make_fusion_block(features, use_bn) + self.scratch.refinenet4 = _make_fusion_block(features, use_bn) + + self.scratch.output_conv = head + + + def forward(self, x): + if self.channels_last == True: + x.contiguous(memory_format=torch.channels_last) + + layer_1, layer_2, layer_3, layer_4 = forward_vit(self.pretrained, x) + + layer_1_rn = self.scratch.layer1_rn(layer_1) + layer_2_rn = self.scratch.layer2_rn(layer_2) + layer_3_rn = self.scratch.layer3_rn(layer_3) + layer_4_rn = self.scratch.layer4_rn(layer_4) + + path_4 = self.scratch.refinenet4(layer_4_rn) + path_3 = self.scratch.refinenet3(path_4, layer_3_rn) + path_2 = self.scratch.refinenet2(path_3, layer_2_rn) + path_1 = self.scratch.refinenet1(path_2, layer_1_rn) + + out = self.scratch.output_conv(path_1) + + return out + + +class DPTDepthModel(DPT): + def __init__(self, path=None, non_negative=True, **kwargs): + features = kwargs["features"] if "features" in kwargs else 256 + + head = nn.Sequential( + nn.Conv2d(features, features // 2, kernel_size=3, stride=1, padding=1), + Interpolate(scale_factor=2, mode="bilinear", align_corners=True), + nn.Conv2d(features // 2, 32, kernel_size=3, stride=1, padding=1), + nn.ReLU(True), + nn.Conv2d(32, 1, kernel_size=1, stride=1, padding=0), + nn.ReLU(True) if non_negative else nn.Identity(), + nn.Identity(), + ) + + super().__init__(head, **kwargs) + + if path is not None: + self.load(path) + + def forward(self, x): + return super().forward(x).squeeze(dim=1) + diff --git a/ldm/modules/midas/midas/midas_net.py b/ldm/modules/midas/midas/midas_net.py new file mode 100644 index 0000000000000000000000000000000000000000..8a954977800b0a0f48807e80fa63041910e33c1f --- /dev/null +++ b/ldm/modules/midas/midas/midas_net.py @@ -0,0 +1,76 @@ +"""MidashNet: Network for monocular depth estimation trained by mixing several datasets. +This file contains code that is adapted from +https://github.com/thomasjpfan/pytorch_refinenet/blob/master/pytorch_refinenet/refinenet/refinenet_4cascade.py +""" +import torch +import torch.nn as nn + +from .base_model import BaseModel +from .blocks import FeatureFusionBlock, Interpolate, _make_encoder + + +class MidasNet(BaseModel): + """Network for monocular depth estimation. + """ + + def __init__(self, path=None, features=256, non_negative=True): + """Init. + + Args: + path (str, optional): Path to saved model. Defaults to None. + features (int, optional): Number of features. Defaults to 256. + backbone (str, optional): Backbone network for encoder. Defaults to resnet50 + """ + print("Loading weights: ", path) + + super(MidasNet, self).__init__() + + use_pretrained = False if path is None else True + + self.pretrained, self.scratch = _make_encoder(backbone="resnext101_wsl", features=features, use_pretrained=use_pretrained) + + self.scratch.refinenet4 = FeatureFusionBlock(features) + self.scratch.refinenet3 = FeatureFusionBlock(features) + self.scratch.refinenet2 = FeatureFusionBlock(features) + self.scratch.refinenet1 = FeatureFusionBlock(features) + + self.scratch.output_conv = nn.Sequential( + nn.Conv2d(features, 128, kernel_size=3, stride=1, padding=1), + Interpolate(scale_factor=2, mode="bilinear"), + nn.Conv2d(128, 32, kernel_size=3, stride=1, padding=1), + nn.ReLU(True), + nn.Conv2d(32, 1, kernel_size=1, stride=1, padding=0), + nn.ReLU(True) if non_negative else nn.Identity(), + ) + + if path: + self.load(path) + + def forward(self, x): + """Forward pass. + + Args: + x (tensor): input data (image) + + Returns: + tensor: depth + """ + + layer_1 = self.pretrained.layer1(x) + layer_2 = self.pretrained.layer2(layer_1) + layer_3 = self.pretrained.layer3(layer_2) + layer_4 = self.pretrained.layer4(layer_3) + + layer_1_rn = self.scratch.layer1_rn(layer_1) + layer_2_rn = self.scratch.layer2_rn(layer_2) + layer_3_rn = self.scratch.layer3_rn(layer_3) + layer_4_rn = self.scratch.layer4_rn(layer_4) + + path_4 = self.scratch.refinenet4(layer_4_rn) + path_3 = self.scratch.refinenet3(path_4, layer_3_rn) + path_2 = self.scratch.refinenet2(path_3, layer_2_rn) + path_1 = self.scratch.refinenet1(path_2, layer_1_rn) + + out = self.scratch.output_conv(path_1) + + return torch.squeeze(out, dim=1) diff --git a/ldm/modules/midas/midas/midas_net_custom.py b/ldm/modules/midas/midas/midas_net_custom.py new file mode 100644 index 0000000000000000000000000000000000000000..50e4acb5e53d5fabefe3dde16ab49c33c2b7797c --- /dev/null +++ b/ldm/modules/midas/midas/midas_net_custom.py @@ -0,0 +1,128 @@ +"""MidashNet: Network for monocular depth estimation trained by mixing several datasets. +This file contains code that is adapted from +https://github.com/thomasjpfan/pytorch_refinenet/blob/master/pytorch_refinenet/refinenet/refinenet_4cascade.py +""" +import torch +import torch.nn as nn + +from .base_model import BaseModel +from .blocks import FeatureFusionBlock, FeatureFusionBlock_custom, Interpolate, _make_encoder + + +class MidasNet_small(BaseModel): + """Network for monocular depth estimation. + """ + + def __init__(self, path=None, features=64, backbone="efficientnet_lite3", non_negative=True, exportable=True, channels_last=False, align_corners=True, + blocks={'expand': True}): + """Init. + + Args: + path (str, optional): Path to saved model. Defaults to None. + features (int, optional): Number of features. Defaults to 256. + backbone (str, optional): Backbone network for encoder. Defaults to resnet50 + """ + print("Loading weights: ", path) + + super(MidasNet_small, self).__init__() + + use_pretrained = False if path else True + + self.channels_last = channels_last + self.blocks = blocks + self.backbone = backbone + + self.groups = 1 + + features1=features + features2=features + features3=features + features4=features + self.expand = False + if "expand" in self.blocks and self.blocks['expand'] == True: + self.expand = True + features1=features + features2=features*2 + features3=features*4 + features4=features*8 + + self.pretrained, self.scratch = _make_encoder(self.backbone, features, use_pretrained, groups=self.groups, expand=self.expand, exportable=exportable) + + self.scratch.activation = nn.ReLU(False) + + self.scratch.refinenet4 = FeatureFusionBlock_custom(features4, self.scratch.activation, deconv=False, bn=False, expand=self.expand, align_corners=align_corners) + self.scratch.refinenet3 = FeatureFusionBlock_custom(features3, self.scratch.activation, deconv=False, bn=False, expand=self.expand, align_corners=align_corners) + self.scratch.refinenet2 = FeatureFusionBlock_custom(features2, self.scratch.activation, deconv=False, bn=False, expand=self.expand, align_corners=align_corners) + self.scratch.refinenet1 = FeatureFusionBlock_custom(features1, self.scratch.activation, deconv=False, bn=False, align_corners=align_corners) + + + self.scratch.output_conv = nn.Sequential( + nn.Conv2d(features, features//2, kernel_size=3, stride=1, padding=1, groups=self.groups), + Interpolate(scale_factor=2, mode="bilinear"), + nn.Conv2d(features//2, 32, kernel_size=3, stride=1, padding=1), + self.scratch.activation, + nn.Conv2d(32, 1, kernel_size=1, stride=1, padding=0), + nn.ReLU(True) if non_negative else nn.Identity(), + nn.Identity(), + ) + + if path: + self.load(path) + + + def forward(self, x): + """Forward pass. + + Args: + x (tensor): input data (image) + + Returns: + tensor: depth + """ + if self.channels_last==True: + print("self.channels_last = ", self.channels_last) + x.contiguous(memory_format=torch.channels_last) + + + layer_1 = self.pretrained.layer1(x) + layer_2 = self.pretrained.layer2(layer_1) + layer_3 = self.pretrained.layer3(layer_2) + layer_4 = self.pretrained.layer4(layer_3) + + layer_1_rn = self.scratch.layer1_rn(layer_1) + layer_2_rn = self.scratch.layer2_rn(layer_2) + layer_3_rn = self.scratch.layer3_rn(layer_3) + layer_4_rn = self.scratch.layer4_rn(layer_4) + + + path_4 = self.scratch.refinenet4(layer_4_rn) + path_3 = self.scratch.refinenet3(path_4, layer_3_rn) + path_2 = self.scratch.refinenet2(path_3, layer_2_rn) + path_1 = self.scratch.refinenet1(path_2, layer_1_rn) + + out = self.scratch.output_conv(path_1) + + return torch.squeeze(out, dim=1) + + + +def fuse_model(m): + prev_previous_type = nn.Identity() + prev_previous_name = '' + previous_type = nn.Identity() + previous_name = '' + for name, module in m.named_modules(): + if prev_previous_type == nn.Conv2d and previous_type == nn.BatchNorm2d and type(module) == nn.ReLU: + # print("FUSED ", prev_previous_name, previous_name, name) + torch.quantization.fuse_modules(m, [prev_previous_name, previous_name, name], inplace=True) + elif prev_previous_type == nn.Conv2d and previous_type == nn.BatchNorm2d: + # print("FUSED ", prev_previous_name, previous_name) + torch.quantization.fuse_modules(m, [prev_previous_name, previous_name], inplace=True) + # elif previous_type == nn.Conv2d and type(module) == nn.ReLU: + # print("FUSED ", previous_name, name) + # torch.quantization.fuse_modules(m, [previous_name, name], inplace=True) + + prev_previous_type = previous_type + prev_previous_name = previous_name + previous_type = type(module) + previous_name = name \ No newline at end of file diff --git a/ldm/modules/midas/midas/transforms.py b/ldm/modules/midas/midas/transforms.py new file mode 100644 index 0000000000000000000000000000000000000000..350cbc11662633ad7f8968eb10be2e7de6e384e9 --- /dev/null +++ b/ldm/modules/midas/midas/transforms.py @@ -0,0 +1,234 @@ +import numpy as np +import cv2 +import math + + +def apply_min_size(sample, size, image_interpolation_method=cv2.INTER_AREA): + """Rezise the sample to ensure the given size. Keeps aspect ratio. + + Args: + sample (dict): sample + size (tuple): image size + + Returns: + tuple: new size + """ + shape = list(sample["disparity"].shape) + + if shape[0] >= size[0] and shape[1] >= size[1]: + return sample + + scale = [0, 0] + scale[0] = size[0] / shape[0] + scale[1] = size[1] / shape[1] + + scale = max(scale) + + shape[0] = math.ceil(scale * shape[0]) + shape[1] = math.ceil(scale * shape[1]) + + # resize + sample["image"] = cv2.resize( + sample["image"], tuple(shape[::-1]), interpolation=image_interpolation_method + ) + + sample["disparity"] = cv2.resize( + sample["disparity"], tuple(shape[::-1]), interpolation=cv2.INTER_NEAREST + ) + sample["mask"] = cv2.resize( + sample["mask"].astype(np.float32), + tuple(shape[::-1]), + interpolation=cv2.INTER_NEAREST, + ) + sample["mask"] = sample["mask"].astype(bool) + + return tuple(shape) + + +class Resize(object): + """Resize sample to given size (width, height). + """ + + def __init__( + self, + width, + height, + resize_target=True, + keep_aspect_ratio=False, + ensure_multiple_of=1, + resize_method="lower_bound", + image_interpolation_method=cv2.INTER_AREA, + ): + """Init. + + Args: + width (int): desired output width + height (int): desired output height + resize_target (bool, optional): + True: Resize the full sample (image, mask, target). + False: Resize image only. + Defaults to True. + keep_aspect_ratio (bool, optional): + True: Keep the aspect ratio of the input sample. + Output sample might not have the given width and height, and + resize behaviour depends on the parameter 'resize_method'. + Defaults to False. + ensure_multiple_of (int, optional): + Output width and height is constrained to be multiple of this parameter. + Defaults to 1. + resize_method (str, optional): + "lower_bound": Output will be at least as large as the given size. + "upper_bound": Output will be at max as large as the given size. (Output size might be smaller than given size.) + "minimal": Scale as least as possible. (Output size might be smaller than given size.) + Defaults to "lower_bound". + """ + self.__width = width + self.__height = height + + self.__resize_target = resize_target + self.__keep_aspect_ratio = keep_aspect_ratio + self.__multiple_of = ensure_multiple_of + self.__resize_method = resize_method + self.__image_interpolation_method = image_interpolation_method + + def constrain_to_multiple_of(self, x, min_val=0, max_val=None): + y = (np.round(x / self.__multiple_of) * self.__multiple_of).astype(int) + + if max_val is not None and y > max_val: + y = (np.floor(x / self.__multiple_of) * self.__multiple_of).astype(int) + + if y < min_val: + y = (np.ceil(x / self.__multiple_of) * self.__multiple_of).astype(int) + + return y + + def get_size(self, width, height): + # determine new height and width + scale_height = self.__height / height + scale_width = self.__width / width + + if self.__keep_aspect_ratio: + if self.__resize_method == "lower_bound": + # scale such that output size is lower bound + if scale_width > scale_height: + # fit width + scale_height = scale_width + else: + # fit height + scale_width = scale_height + elif self.__resize_method == "upper_bound": + # scale such that output size is upper bound + if scale_width < scale_height: + # fit width + scale_height = scale_width + else: + # fit height + scale_width = scale_height + elif self.__resize_method == "minimal": + # scale as least as possbile + if abs(1 - scale_width) < abs(1 - scale_height): + # fit width + scale_height = scale_width + else: + # fit height + scale_width = scale_height + else: + raise ValueError( + f"resize_method {self.__resize_method} not implemented" + ) + + if self.__resize_method == "lower_bound": + new_height = self.constrain_to_multiple_of( + scale_height * height, min_val=self.__height + ) + new_width = self.constrain_to_multiple_of( + scale_width * width, min_val=self.__width + ) + elif self.__resize_method == "upper_bound": + new_height = self.constrain_to_multiple_of( + scale_height * height, max_val=self.__height + ) + new_width = self.constrain_to_multiple_of( + scale_width * width, max_val=self.__width + ) + elif self.__resize_method == "minimal": + new_height = self.constrain_to_multiple_of(scale_height * height) + new_width = self.constrain_to_multiple_of(scale_width * width) + else: + raise ValueError(f"resize_method {self.__resize_method} not implemented") + + return (new_width, new_height) + + def __call__(self, sample): + width, height = self.get_size( + sample["image"].shape[1], sample["image"].shape[0] + ) + + # resize sample + sample["image"] = cv2.resize( + sample["image"], + (width, height), + interpolation=self.__image_interpolation_method, + ) + + if self.__resize_target: + if "disparity" in sample: + sample["disparity"] = cv2.resize( + sample["disparity"], + (width, height), + interpolation=cv2.INTER_NEAREST, + ) + + if "depth" in sample: + sample["depth"] = cv2.resize( + sample["depth"], (width, height), interpolation=cv2.INTER_NEAREST + ) + + sample["mask"] = cv2.resize( + sample["mask"].astype(np.float32), + (width, height), + interpolation=cv2.INTER_NEAREST, + ) + sample["mask"] = sample["mask"].astype(bool) + + return sample + + +class NormalizeImage(object): + """Normlize image by given mean and std. + """ + + def __init__(self, mean, std): + self.__mean = mean + self.__std = std + + def __call__(self, sample): + sample["image"] = (sample["image"] - self.__mean) / self.__std + + return sample + + +class PrepareForNet(object): + """Prepare sample for usage as network input. + """ + + def __init__(self): + pass + + def __call__(self, sample): + image = np.transpose(sample["image"], (2, 0, 1)) + sample["image"] = np.ascontiguousarray(image).astype(np.float32) + + if "mask" in sample: + sample["mask"] = sample["mask"].astype(np.float32) + sample["mask"] = np.ascontiguousarray(sample["mask"]) + + if "disparity" in sample: + disparity = sample["disparity"].astype(np.float32) + sample["disparity"] = np.ascontiguousarray(disparity) + + if "depth" in sample: + depth = sample["depth"].astype(np.float32) + sample["depth"] = np.ascontiguousarray(depth) + + return sample diff --git a/ldm/modules/midas/midas/vit.py b/ldm/modules/midas/midas/vit.py new file mode 100644 index 0000000000000000000000000000000000000000..ea46b1be88b261b0dec04f3da0256f5f66f88a74 --- /dev/null +++ b/ldm/modules/midas/midas/vit.py @@ -0,0 +1,491 @@ +import torch +import torch.nn as nn +import timm +import types +import math +import torch.nn.functional as F + + +class Slice(nn.Module): + def __init__(self, start_index=1): + super(Slice, self).__init__() + self.start_index = start_index + + def forward(self, x): + return x[:, self.start_index :] + + +class AddReadout(nn.Module): + def __init__(self, start_index=1): + super(AddReadout, self).__init__() + self.start_index = start_index + + def forward(self, x): + if self.start_index == 2: + readout = (x[:, 0] + x[:, 1]) / 2 + else: + readout = x[:, 0] + return x[:, self.start_index :] + readout.unsqueeze(1) + + +class ProjectReadout(nn.Module): + def __init__(self, in_features, start_index=1): + super(ProjectReadout, self).__init__() + self.start_index = start_index + + self.project = nn.Sequential(nn.Linear(2 * in_features, in_features), nn.GELU()) + + def forward(self, x): + readout = x[:, 0].unsqueeze(1).expand_as(x[:, self.start_index :]) + features = torch.cat((x[:, self.start_index :], readout), -1) + + return self.project(features) + + +class Transpose(nn.Module): + def __init__(self, dim0, dim1): + super(Transpose, self).__init__() + self.dim0 = dim0 + self.dim1 = dim1 + + def forward(self, x): + x = x.transpose(self.dim0, self.dim1) + return x + + +def forward_vit(pretrained, x): + b, c, h, w = x.shape + + glob = pretrained.model.forward_flex(x) + + layer_1 = pretrained.activations["1"] + layer_2 = pretrained.activations["2"] + layer_3 = pretrained.activations["3"] + layer_4 = pretrained.activations["4"] + + layer_1 = pretrained.act_postprocess1[0:2](layer_1) + layer_2 = pretrained.act_postprocess2[0:2](layer_2) + layer_3 = pretrained.act_postprocess3[0:2](layer_3) + layer_4 = pretrained.act_postprocess4[0:2](layer_4) + + unflatten = nn.Sequential( + nn.Unflatten( + 2, + torch.Size( + [ + h // pretrained.model.patch_size[1], + w // pretrained.model.patch_size[0], + ] + ), + ) + ) + + if layer_1.ndim == 3: + layer_1 = unflatten(layer_1) + if layer_2.ndim == 3: + layer_2 = unflatten(layer_2) + if layer_3.ndim == 3: + layer_3 = unflatten(layer_3) + if layer_4.ndim == 3: + layer_4 = unflatten(layer_4) + + layer_1 = pretrained.act_postprocess1[3 : len(pretrained.act_postprocess1)](layer_1) + layer_2 = pretrained.act_postprocess2[3 : len(pretrained.act_postprocess2)](layer_2) + layer_3 = pretrained.act_postprocess3[3 : len(pretrained.act_postprocess3)](layer_3) + layer_4 = pretrained.act_postprocess4[3 : len(pretrained.act_postprocess4)](layer_4) + + return layer_1, layer_2, layer_3, layer_4 + + +def _resize_pos_embed(self, posemb, gs_h, gs_w): + posemb_tok, posemb_grid = ( + posemb[:, : self.start_index], + posemb[0, self.start_index :], + ) + + gs_old = int(math.sqrt(len(posemb_grid))) + + posemb_grid = posemb_grid.reshape(1, gs_old, gs_old, -1).permute(0, 3, 1, 2) + posemb_grid = F.interpolate(posemb_grid, size=(gs_h, gs_w), mode="bilinear") + posemb_grid = posemb_grid.permute(0, 2, 3, 1).reshape(1, gs_h * gs_w, -1) + + posemb = torch.cat([posemb_tok, posemb_grid], dim=1) + + return posemb + + +def forward_flex(self, x): + b, c, h, w = x.shape + + pos_embed = self._resize_pos_embed( + self.pos_embed, h // self.patch_size[1], w // self.patch_size[0] + ) + + B = x.shape[0] + + if hasattr(self.patch_embed, "backbone"): + x = self.patch_embed.backbone(x) + if isinstance(x, (list, tuple)): + x = x[-1] # last feature if backbone outputs list/tuple of features + + x = self.patch_embed.proj(x).flatten(2).transpose(1, 2) + + if getattr(self, "dist_token", None) is not None: + cls_tokens = self.cls_token.expand( + B, -1, -1 + ) # stole cls_tokens impl from Phil Wang, thanks + dist_token = self.dist_token.expand(B, -1, -1) + x = torch.cat((cls_tokens, dist_token, x), dim=1) + else: + cls_tokens = self.cls_token.expand( + B, -1, -1 + ) # stole cls_tokens impl from Phil Wang, thanks + x = torch.cat((cls_tokens, x), dim=1) + + x = x + pos_embed + x = self.pos_drop(x) + + for blk in self.blocks: + x = blk(x) + + x = self.norm(x) + + return x + + +activations = {} + + +def get_activation(name): + def hook(model, input, output): + activations[name] = output + + return hook + + +def get_readout_oper(vit_features, features, use_readout, start_index=1): + if use_readout == "ignore": + readout_oper = [Slice(start_index)] * len(features) + elif use_readout == "add": + readout_oper = [AddReadout(start_index)] * len(features) + elif use_readout == "project": + readout_oper = [ + ProjectReadout(vit_features, start_index) for out_feat in features + ] + else: + assert ( + False + ), "wrong operation for readout token, use_readout can be 'ignore', 'add', or 'project'" + + return readout_oper + + +def _make_vit_b16_backbone( + model, + features=[96, 192, 384, 768], + size=[384, 384], + hooks=[2, 5, 8, 11], + vit_features=768, + use_readout="ignore", + start_index=1, +): + pretrained = nn.Module() + + pretrained.model = model + pretrained.model.blocks[hooks[0]].register_forward_hook(get_activation("1")) + pretrained.model.blocks[hooks[1]].register_forward_hook(get_activation("2")) + pretrained.model.blocks[hooks[2]].register_forward_hook(get_activation("3")) + pretrained.model.blocks[hooks[3]].register_forward_hook(get_activation("4")) + + pretrained.activations = activations + + readout_oper = get_readout_oper(vit_features, features, use_readout, start_index) + + # 32, 48, 136, 384 + pretrained.act_postprocess1 = nn.Sequential( + readout_oper[0], + Transpose(1, 2), + nn.Unflatten(2, torch.Size([size[0] // 16, size[1] // 16])), + nn.Conv2d( + in_channels=vit_features, + out_channels=features[0], + kernel_size=1, + stride=1, + padding=0, + ), + nn.ConvTranspose2d( + in_channels=features[0], + out_channels=features[0], + kernel_size=4, + stride=4, + padding=0, + bias=True, + dilation=1, + groups=1, + ), + ) + + pretrained.act_postprocess2 = nn.Sequential( + readout_oper[1], + Transpose(1, 2), + nn.Unflatten(2, torch.Size([size[0] // 16, size[1] // 16])), + nn.Conv2d( + in_channels=vit_features, + out_channels=features[1], + kernel_size=1, + stride=1, + padding=0, + ), + nn.ConvTranspose2d( + in_channels=features[1], + out_channels=features[1], + kernel_size=2, + stride=2, + padding=0, + bias=True, + dilation=1, + groups=1, + ), + ) + + pretrained.act_postprocess3 = nn.Sequential( + readout_oper[2], + Transpose(1, 2), + nn.Unflatten(2, torch.Size([size[0] // 16, size[1] // 16])), + nn.Conv2d( + in_channels=vit_features, + out_channels=features[2], + kernel_size=1, + stride=1, + padding=0, + ), + ) + + pretrained.act_postprocess4 = nn.Sequential( + readout_oper[3], + Transpose(1, 2), + nn.Unflatten(2, torch.Size([size[0] // 16, size[1] // 16])), + nn.Conv2d( + in_channels=vit_features, + out_channels=features[3], + kernel_size=1, + stride=1, + padding=0, + ), + nn.Conv2d( + in_channels=features[3], + out_channels=features[3], + kernel_size=3, + stride=2, + padding=1, + ), + ) + + pretrained.model.start_index = start_index + pretrained.model.patch_size = [16, 16] + + # We inject this function into the VisionTransformer instances so that + # we can use it with interpolated position embeddings without modifying the library source. + pretrained.model.forward_flex = types.MethodType(forward_flex, pretrained.model) + pretrained.model._resize_pos_embed = types.MethodType( + _resize_pos_embed, pretrained.model + ) + + return pretrained + + +def _make_pretrained_vitl16_384(pretrained, use_readout="ignore", hooks=None): + model = timm.create_model("vit_large_patch16_384", pretrained=pretrained) + + hooks = [5, 11, 17, 23] if hooks == None else hooks + return _make_vit_b16_backbone( + model, + features=[256, 512, 1024, 1024], + hooks=hooks, + vit_features=1024, + use_readout=use_readout, + ) + + +def _make_pretrained_vitb16_384(pretrained, use_readout="ignore", hooks=None): + model = timm.create_model("vit_base_patch16_384", pretrained=pretrained) + + hooks = [2, 5, 8, 11] if hooks == None else hooks + return _make_vit_b16_backbone( + model, features=[96, 192, 384, 768], hooks=hooks, use_readout=use_readout + ) + + +def _make_pretrained_deitb16_384(pretrained, use_readout="ignore", hooks=None): + model = timm.create_model("vit_deit_base_patch16_384", pretrained=pretrained) + + hooks = [2, 5, 8, 11] if hooks == None else hooks + return _make_vit_b16_backbone( + model, features=[96, 192, 384, 768], hooks=hooks, use_readout=use_readout + ) + + +def _make_pretrained_deitb16_distil_384(pretrained, use_readout="ignore", hooks=None): + model = timm.create_model( + "vit_deit_base_distilled_patch16_384", pretrained=pretrained + ) + + hooks = [2, 5, 8, 11] if hooks == None else hooks + return _make_vit_b16_backbone( + model, + features=[96, 192, 384, 768], + hooks=hooks, + use_readout=use_readout, + start_index=2, + ) + + +def _make_vit_b_rn50_backbone( + model, + features=[256, 512, 768, 768], + size=[384, 384], + hooks=[0, 1, 8, 11], + vit_features=768, + use_vit_only=False, + use_readout="ignore", + start_index=1, +): + pretrained = nn.Module() + + pretrained.model = model + + if use_vit_only == True: + pretrained.model.blocks[hooks[0]].register_forward_hook(get_activation("1")) + pretrained.model.blocks[hooks[1]].register_forward_hook(get_activation("2")) + else: + pretrained.model.patch_embed.backbone.stages[0].register_forward_hook( + get_activation("1") + ) + pretrained.model.patch_embed.backbone.stages[1].register_forward_hook( + get_activation("2") + ) + + pretrained.model.blocks[hooks[2]].register_forward_hook(get_activation("3")) + pretrained.model.blocks[hooks[3]].register_forward_hook(get_activation("4")) + + pretrained.activations = activations + + readout_oper = get_readout_oper(vit_features, features, use_readout, start_index) + + if use_vit_only == True: + pretrained.act_postprocess1 = nn.Sequential( + readout_oper[0], + Transpose(1, 2), + nn.Unflatten(2, torch.Size([size[0] // 16, size[1] // 16])), + nn.Conv2d( + in_channels=vit_features, + out_channels=features[0], + kernel_size=1, + stride=1, + padding=0, + ), + nn.ConvTranspose2d( + in_channels=features[0], + out_channels=features[0], + kernel_size=4, + stride=4, + padding=0, + bias=True, + dilation=1, + groups=1, + ), + ) + + pretrained.act_postprocess2 = nn.Sequential( + readout_oper[1], + Transpose(1, 2), + nn.Unflatten(2, torch.Size([size[0] // 16, size[1] // 16])), + nn.Conv2d( + in_channels=vit_features, + out_channels=features[1], + kernel_size=1, + stride=1, + padding=0, + ), + nn.ConvTranspose2d( + in_channels=features[1], + out_channels=features[1], + kernel_size=2, + stride=2, + padding=0, + bias=True, + dilation=1, + groups=1, + ), + ) + else: + pretrained.act_postprocess1 = nn.Sequential( + nn.Identity(), nn.Identity(), nn.Identity() + ) + pretrained.act_postprocess2 = nn.Sequential( + nn.Identity(), nn.Identity(), nn.Identity() + ) + + pretrained.act_postprocess3 = nn.Sequential( + readout_oper[2], + Transpose(1, 2), + nn.Unflatten(2, torch.Size([size[0] // 16, size[1] // 16])), + nn.Conv2d( + in_channels=vit_features, + out_channels=features[2], + kernel_size=1, + stride=1, + padding=0, + ), + ) + + pretrained.act_postprocess4 = nn.Sequential( + readout_oper[3], + Transpose(1, 2), + nn.Unflatten(2, torch.Size([size[0] // 16, size[1] // 16])), + nn.Conv2d( + in_channels=vit_features, + out_channels=features[3], + kernel_size=1, + stride=1, + padding=0, + ), + nn.Conv2d( + in_channels=features[3], + out_channels=features[3], + kernel_size=3, + stride=2, + padding=1, + ), + ) + + pretrained.model.start_index = start_index + pretrained.model.patch_size = [16, 16] + + # We inject this function into the VisionTransformer instances so that + # we can use it with interpolated position embeddings without modifying the library source. + pretrained.model.forward_flex = types.MethodType(forward_flex, pretrained.model) + + # We inject this function into the VisionTransformer instances so that + # we can use it with interpolated position embeddings without modifying the library source. + pretrained.model._resize_pos_embed = types.MethodType( + _resize_pos_embed, pretrained.model + ) + + return pretrained + + +def _make_pretrained_vitb_rn50_384( + pretrained, use_readout="ignore", hooks=None, use_vit_only=False +): + model = timm.create_model("vit_base_resnet50_384", pretrained=pretrained) + + hooks = [0, 1, 8, 11] if hooks == None else hooks + return _make_vit_b_rn50_backbone( + model, + features=[256, 512, 768, 768], + size=[384, 384], + hooks=hooks, + use_vit_only=use_vit_only, + use_readout=use_readout, + ) diff --git a/ldm/modules/midas/utils.py b/ldm/modules/midas/utils.py new file mode 100644 index 0000000000000000000000000000000000000000..9a9d3b5b66370fa98da9e067ba53ead848ea9a59 --- /dev/null +++ b/ldm/modules/midas/utils.py @@ -0,0 +1,189 @@ +"""Utils for monoDepth.""" +import sys +import re +import numpy as np +import cv2 +import torch + + +def read_pfm(path): + """Read pfm file. + + Args: + path (str): path to file + + Returns: + tuple: (data, scale) + """ + with open(path, "rb") as file: + + color = None + width = None + height = None + scale = None + endian = None + + header = file.readline().rstrip() + if header.decode("ascii") == "PF": + color = True + elif header.decode("ascii") == "Pf": + color = False + else: + raise Exception("Not a PFM file: " + path) + + dim_match = re.match(r"^(\d+)\s(\d+)\s$", file.readline().decode("ascii")) + if dim_match: + width, height = list(map(int, dim_match.groups())) + else: + raise Exception("Malformed PFM header.") + + scale = float(file.readline().decode("ascii").rstrip()) + if scale < 0: + # little-endian + endian = "<" + scale = -scale + else: + # big-endian + endian = ">" + + data = np.fromfile(file, endian + "f") + shape = (height, width, 3) if color else (height, width) + + data = np.reshape(data, shape) + data = np.flipud(data) + + return data, scale + + +def write_pfm(path, image, scale=1): + """Write pfm file. + + Args: + path (str): pathto file + image (array): data + scale (int, optional): Scale. Defaults to 1. + """ + + with open(path, "wb") as file: + color = None + + if image.dtype.name != "float32": + raise Exception("Image dtype must be float32.") + + image = np.flipud(image) + + if len(image.shape) == 3 and image.shape[2] == 3: # color image + color = True + elif ( + len(image.shape) == 2 or len(image.shape) == 3 and image.shape[2] == 1 + ): # greyscale + color = False + else: + raise Exception("Image must have H x W x 3, H x W x 1 or H x W dimensions.") + + file.write("PF\n" if color else "Pf\n".encode()) + file.write("%d %d\n".encode() % (image.shape[1], image.shape[0])) + + endian = image.dtype.byteorder + + if endian == "<" or endian == "=" and sys.byteorder == "little": + scale = -scale + + file.write("%f\n".encode() % scale) + + image.tofile(file) + + +def read_image(path): + """Read image and output RGB image (0-1). + + Args: + path (str): path to file + + Returns: + array: RGB image (0-1) + """ + img = cv2.imread(path) + + if img.ndim == 2: + img = cv2.cvtColor(img, cv2.COLOR_GRAY2BGR) + + img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB) / 255.0 + + return img + + +def resize_image(img): + """Resize image and make it fit for network. + + Args: + img (array): image + + Returns: + tensor: data ready for network + """ + height_orig = img.shape[0] + width_orig = img.shape[1] + + if width_orig > height_orig: + scale = width_orig / 384 + else: + scale = height_orig / 384 + + height = (np.ceil(height_orig / scale / 32) * 32).astype(int) + width = (np.ceil(width_orig / scale / 32) * 32).astype(int) + + img_resized = cv2.resize(img, (width, height), interpolation=cv2.INTER_AREA) + + img_resized = ( + torch.from_numpy(np.transpose(img_resized, (2, 0, 1))).contiguous().float() + ) + img_resized = img_resized.unsqueeze(0) + + return img_resized + + +def resize_depth(depth, width, height): + """Resize depth map and bring to CPU (numpy). + + Args: + depth (tensor): depth + width (int): image width + height (int): image height + + Returns: + array: processed depth + """ + depth = torch.squeeze(depth[0, :, :, :]).to("cpu") + + depth_resized = cv2.resize( + depth.numpy(), (width, height), interpolation=cv2.INTER_CUBIC + ) + + return depth_resized + +def write_depth(path, depth, bits=1): + """Write depth map to pfm and png file. + + Args: + path (str): filepath without extension + depth (array): depth + """ + write_pfm(path + ".pfm", depth.astype(np.float32)) + + depth_min = depth.min() + depth_max = depth.max() + + max_val = (2**(8*bits))-1 + + if depth_max - depth_min > np.finfo("float").eps: + out = max_val * (depth - depth_min) / (depth_max - depth_min) + else: + out = np.zeros(depth.shape, dtype=depth.type) + + if bits == 1: + cv2.imwrite(path + ".png", out.astype("uint8")) + elif bits == 2: + cv2.imwrite(path + ".png", out.astype("uint16")) + + return diff --git a/ldm/util.py b/ldm/util.py new file mode 100644 index 0000000000000000000000000000000000000000..9ede259d5e7876912724582cc06a71916f6f9b62 --- /dev/null +++ b/ldm/util.py @@ -0,0 +1,207 @@ +import importlib + +import torch +from torch import optim +import numpy as np + +from inspect import isfunction +from PIL import Image, ImageDraw, ImageFont + + +def autocast(f): + def do_autocast(*args, **kwargs): + with torch.cuda.amp.autocast(enabled=True, + dtype=torch.get_autocast_gpu_dtype(), + cache_enabled=torch.is_autocast_cache_enabled()): + return f(*args, **kwargs) + + return do_autocast + + +def log_txt_as_img(wh, xc, size=10): + # wh a tuple of (width, height) + # xc a list of captions to plot + b = len(xc) + txts = list() + for bi in range(b): + txt = Image.new("RGB", wh, color="white") + draw = ImageDraw.Draw(txt) + font = ImageFont.truetype('data/DejaVuSans.ttf', size=size) + nc = int(40 * (wh[0] / 256)) + lines = "\n".join(xc[bi][start:start + nc] for start in range(0, len(xc[bi]), nc)) + + try: + draw.text((0, 0), lines, fill="black", font=font) + except UnicodeEncodeError: + print("Cant encode string for logging. Skipping.") + + txt = np.array(txt).transpose(2, 0, 1) / 127.5 - 1.0 + txts.append(txt) + txts = np.stack(txts) + txts = torch.tensor(txts) + return txts + + +def ismap(x): + if not isinstance(x, torch.Tensor): + return False + return (len(x.shape) == 4) and (x.shape[1] > 3) + + +def isimage(x): + if not isinstance(x,torch.Tensor): + return False + return (len(x.shape) == 4) and (x.shape[1] == 3 or x.shape[1] == 1) + + +def exists(x): + return x is not None + + +def default(val, d): + if exists(val): + return val + return d() if isfunction(d) else d + + +def mean_flat(tensor): + """ + https://github.com/openai/guided-diffusion/blob/27c20a8fab9cb472df5d6bdd6c8d11c8f430b924/guided_diffusion/nn.py#L86 + Take the mean over all non-batch dimensions. + """ + return tensor.mean(dim=list(range(1, len(tensor.shape)))) + + +def count_params(model, verbose=False): + total_params = sum(p.numel() for p in model.parameters()) + if verbose: + print(f"{model.__class__.__name__} has {total_params*1.e-6:.2f} M params.") + return total_params + + +def instantiate_from_config(config): + if not "target" in config: + if config == '__is_first_stage__': + return None + elif config == "__is_unconditional__": + return None + raise KeyError("Expected key `target` to instantiate.") + return get_obj_from_str(config["target"])(**config.get("params", dict())) + + +def get_obj_from_str(string, reload=False): + module, cls = string.rsplit(".", 1) + if reload: + module_imp = importlib.import_module(module) + importlib.reload(module_imp) + return getattr(importlib.import_module(module, package=None), cls) + + +class AdamWwithEMAandWings(optim.Optimizer): + # credit to https://gist.github.com/crowsonkb/65f7265353f403714fce3b2595e0b298 + def __init__(self, params, lr=1.e-3, betas=(0.9, 0.999), eps=1.e-8, # TODO: check hyperparameters before using + weight_decay=1.e-2, amsgrad=False, ema_decay=0.9999, # ema decay to match previous code + ema_power=1., param_names=()): + """AdamW that saves EMA versions of the parameters.""" + if not 0.0 <= lr: + raise ValueError("Invalid learning rate: {}".format(lr)) + if not 0.0 <= eps: + raise ValueError("Invalid epsilon value: {}".format(eps)) + if not 0.0 <= betas[0] < 1.0: + raise ValueError("Invalid beta parameter at index 0: {}".format(betas[0])) + if not 0.0 <= betas[1] < 1.0: + raise ValueError("Invalid beta parameter at index 1: {}".format(betas[1])) + if not 0.0 <= weight_decay: + raise ValueError("Invalid weight_decay value: {}".format(weight_decay)) + if not 0.0 <= ema_decay <= 1.0: + raise ValueError("Invalid ema_decay value: {}".format(ema_decay)) + defaults = dict(lr=lr, betas=betas, eps=eps, + weight_decay=weight_decay, amsgrad=amsgrad, ema_decay=ema_decay, + ema_power=ema_power, param_names=param_names) + super().__init__(params, defaults) + + def __setstate__(self, state): + super().__setstate__(state) + for group in self.param_groups: + group.setdefault('amsgrad', False) + + @torch.no_grad() + def step(self, closure=None): + """Performs a single optimization step. + Args: + closure (callable, optional): A closure that reevaluates the model + and returns the loss. + """ + loss = None + if closure is not None: + with torch.enable_grad(): + loss = closure() + + for group in self.param_groups: + params_with_grad = [] + grads = [] + exp_avgs = [] + exp_avg_sqs = [] + ema_params_with_grad = [] + state_sums = [] + max_exp_avg_sqs = [] + state_steps = [] + amsgrad = group['amsgrad'] + beta1, beta2 = group['betas'] + ema_decay = group['ema_decay'] + ema_power = group['ema_power'] + + for p in group['params']: + if p.grad is None: + continue + params_with_grad.append(p) + if p.grad.is_sparse: + raise RuntimeError('AdamW does not support sparse gradients') + grads.append(p.grad) + + state = self.state[p] + + # State initialization + if len(state) == 0: + state['step'] = 0 + # Exponential moving average of gradient values + state['exp_avg'] = torch.zeros_like(p, memory_format=torch.preserve_format) + # Exponential moving average of squared gradient values + state['exp_avg_sq'] = torch.zeros_like(p, memory_format=torch.preserve_format) + if amsgrad: + # Maintains max of all exp. moving avg. of sq. grad. values + state['max_exp_avg_sq'] = torch.zeros_like(p, memory_format=torch.preserve_format) + # Exponential moving average of parameter values + state['param_exp_avg'] = p.detach().float().clone() + + exp_avgs.append(state['exp_avg']) + exp_avg_sqs.append(state['exp_avg_sq']) + ema_params_with_grad.append(state['param_exp_avg']) + + if amsgrad: + max_exp_avg_sqs.append(state['max_exp_avg_sq']) + + # update the steps for each param group update + state['step'] += 1 + # record the step after step update + state_steps.append(state['step']) + + optim._functional.adamw(params_with_grad, + grads, + exp_avgs, + exp_avg_sqs, + max_exp_avg_sqs, + state_steps, + amsgrad=amsgrad, + beta1=beta1, + beta2=beta2, + lr=group['lr'], + weight_decay=group['weight_decay'], + eps=group['eps'], + maximize=False) + + cur_ema_decay = min(ema_decay, 1 - state['step'] ** -ema_power) + for param, ema_param in zip(params_with_grad, ema_params_with_grad): + ema_param.mul_(cur_ema_decay).add_(param.float(), alpha=1 - cur_ema_decay) + + return loss \ No newline at end of file diff --git a/modelcard.md b/modelcard.md new file mode 100644 index 0000000000000000000000000000000000000000..4b6190921e063f60e8089490d172e6ce8714a752 --- /dev/null +++ b/modelcard.md @@ -0,0 +1,153 @@ +# Stable Diffusion v2 Model Card +This model card focuses on the models associated with the Stable Diffusion v2, available [here](https://github.com/Stability-AI/stablediffusion/). + +## Model Details +- **Developed by:** Robin Rombach, Patrick Esser +- **Model type:** Diffusion-based text-to-image generation model +- **Language(s):** English +- **License:** CreativeML Open RAIL++-M License +- **Model Description:** This is a model that can be used to generate and modify images based on text prompts. It is a [Latent Diffusion Model](https://arxiv.org/abs/2112.10752) that uses a fixed, pretrained text encoder ([OpenCLIP-ViT/H](https://github.com/mlfoundations/open_clip)). +- **Resources for more information:** [GitHub Repository](https://github.com/Stability-AI/). +- **Cite as:** + + @InProceedings{Rombach_2022_CVPR, + author = {Rombach, Robin and Blattmann, Andreas and Lorenz, Dominik and Esser, Patrick and Ommer, Bj\"orn}, + title = {High-Resolution Image Synthesis With Latent Diffusion Models}, + booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)}, + month = {June}, + year = {2022}, + pages = {10684-10695} + } + +# Uses + +## Direct Use +The model is intended for research purposes only. Possible research areas and tasks include + +- Safe deployment of models which have the potential to generate harmful content. +- Probing and understanding the limitations and biases of generative models. +- Generation of artworks and use in design and other artistic processes. +- Applications in educational or creative tools. +- Research on generative models. + +Excluded uses are described below. + + ### Misuse, Malicious Use, and Out-of-Scope Use +_Note: This section is originally taken from the [DALLE-MINI model card](https://huggingface.co/dalle-mini/dalle-mini), was used for Stable Diffusion v1, but applies in the same way to Stable Diffusion v2_. + +The model should not be used to intentionally create or disseminate images that create hostile or alienating environments for people. This includes generating images that people would foreseeably find disturbing, distressing, or offensive; or content that propagates historical or current stereotypes. + +#### Out-of-Scope Use +The model was not trained to be factual or true representations of people or events, and therefore using the model to generate such content is out-of-scope for the abilities of this model. + +#### Misuse and Malicious Use +Using the model to generate content that is cruel to individuals is a misuse of this model. This includes, but is not limited to: + +- Generating demeaning, dehumanizing, or otherwise harmful representations of people or their environments, cultures, religions, etc. +- Intentionally promoting or propagating discriminatory content or harmful stereotypes. +- Impersonating individuals without their consent. +- Sexual content without consent of the people who might see it. +- Mis- and disinformation +- Representations of egregious violence and gore +- Sharing of copyrighted or licensed material in violation of its terms of use. +- Sharing content that is an alteration of copyrighted or licensed material in violation of its terms of use. + +## Limitations and Bias + +### Limitations + +- The model does not achieve perfect photorealism +- The model cannot render legible text +- The model does not perform well on more difficult tasks which involve compositionality, such as rendering an image corresponding to “A red cube on top of a blue sphere” +- Faces and people in general may not be generated properly. +- The model was trained mainly with English captions and will not work as well in other languages. +- The autoencoding part of the model is lossy +- The model was trained on a subset of the large-scale dataset + [LAION-5B](https://laion.ai/blog/laion-5b/), which contains adult, violent and sexual content. To partially mitigate this, we have filtered the dataset using LAION's NFSW detector (see Training section). + +### Bias +While the capabilities of image generation models are impressive, they can also reinforce or exacerbate social biases. +Stable Diffusion vw was primarily trained on subsets of [LAION-2B(en)](https://laion.ai/blog/laion-5b/), +which consists of images that are limited to English descriptions. +Texts and images from communities and cultures that use other languages are likely to be insufficiently accounted for. +This affects the overall output of the model, as white and western cultures are often set as the default. Further, the +ability of the model to generate content with non-English prompts is significantly worse than with English-language prompts. +Stable Diffusion v2 mirrors and exacerbates biases to such a degree that viewer discretion must be advised irrespective of the input or its intent. + + +## Training + +**Training Data** +The model developers used the following dataset for training the model: + +- LAION-5B and subsets (details below). The training data is further filtered using LAION's NSFW detector. For more details, please refer to LAION-5B's [NeurIPS 2022](https://openreview.net/forum?id=M3Y74vmsMcY) paper and reviewer discussions on the topic. + +**Training Procedure** +Stable Diffusion v2 is a latent diffusion model which combines an autoencoder with a diffusion model that is trained in the latent space of the autoencoder. During training, + +- Images are encoded through an encoder, which turns images into latent representations. The autoencoder uses a relative downsampling factor of 8 and maps images of shape H x W x 3 to latents of shape H/f x W/f x 4 +- Text prompts are encoded through the OpenCLIP-ViT/H text-encoder. +- The output of the text encoder is fed into the UNet backbone of the latent diffusion model via cross-attention. +- The loss is a reconstruction objective between the noise that was added to the latent and the prediction made by the UNet. We also use the so-called _v-objective_, see https://arxiv.org/abs/2202.00512. + +We currently provide the following checkpoints, for various versions: + +### Version 2.1 + +- `512-base-ema.ckpt`: Fine-tuned on `512-base-ema.ckpt` 2.0 with 220k extra steps taken, with `punsafe=0.98` on the same dataset. +- `768-v-ema.ckpt`: Resumed from `768-v-ema.ckpt` 2.0 with an additional 55k steps on the same dataset (`punsafe=0.1`), and then fine-tuned for another 155k extra steps with `punsafe=0.98`. + +**SD-unCLIP 2.1** is a finetuned version of Stable Diffusion 2.1, modified to accept (noisy) CLIP image embedding in addition to the text prompt, and can be used to create image variations ([Examples](https://github.com/Stability-AI/stablediffusion/blob/main/doc/UNCLIP.MD)) or can be chained with text-to-image CLIP priors. The amount of noise added to the image embedding can be specified via the `noise_level` (0 means no noise, 1000 full noise). + +If you plan on building applications on top of the model that the general public may use, you are responsible for adding the guardrails to minimize or prevent misuse of the application, especially for use-cases highlighted in the earlier section, Misuse, Malicious Use, and Out-of-Scope Use. + +A public demo of SD-unCLIP is already available at [clipdrop.co/stable-diffusion-reimagine](https://clipdrop.co/stable-diffusion-reimagine) + +### Version 2.0 + +- `512-base-ema.ckpt`: 550k steps at resolution `256x256` on a subset of [LAION-5B](https://laion.ai/blog/laion-5b/) filtered for explicit pornographic material, using the [LAION-NSFW classifier](https://github.com/LAION-AI/CLIP-based-NSFW-Detector) with `punsafe=0.1` and an [aesthetic score](https://github.com/christophschuhmann/improved-aesthetic-predictor) >= `4.5`. + 850k steps at resolution `512x512` on the same dataset with resolution `>= 512x512`. +- `768-v-ema.ckpt`: Resumed from `512-base-ema.ckpt` and trained for 150k steps using a [v-objective](https://arxiv.org/abs/2202.00512) on the same dataset. Resumed for another 140k steps on a `768x768` subset of our dataset. +- `512-depth-ema.ckpt`: Resumed from `512-base-ema.ckpt` and finetuned for 200k steps. Added an extra input channel to process the (relative) depth prediction produced by [MiDaS](https://github.com/isl-org/MiDaS) (`dpt_hybrid`) which is used as an additional conditioning. +The additional input channels of the U-Net which process this extra information were zero-initialized. +- `512-inpainting-ema.ckpt`: Resumed from `512-base-ema.ckpt` and trained for another 200k steps. Follows the mask-generation strategy presented in [LAMA](https://github.com/saic-mdal/lama) which, in combination with the latent VAE representations of the masked image, are used as an additional conditioning. +The additional input channels of the U-Net which process this extra information were zero-initialized. The same strategy was used to train the [1.5-inpainting checkpoint](https://github.com/saic-mdal/lama). +- `x4-upscaling-ema.ckpt`: Trained for 1.25M steps on a 10M subset of LAION containing images `>2048x2048`. The model was trained on crops of size `512x512` and is a text-guided [latent upscaling diffusion model](https://arxiv.org/abs/2112.10752). +In addition to the textual input, it receives a `noise_level` as an input parameter, which can be used to add noise to the low-resolution input according to a [predefined diffusion schedule](configs/stable-diffusion/x4-upscaling.yaml). + +- **Hardware:** 32 x 8 x A100 GPUs +- **Optimizer:** AdamW +- **Gradient Accumulations**: 1 +- **Batch:** 32 x 8 x 2 x 4 = 2048 +- **Learning rate:** warmup to 0.0001 for 10,000 steps and then kept constant + +## Evaluation Results +Evaluations with different classifier-free guidance scales (1.5, 2.0, 3.0, 4.0, +5.0, 6.0, 7.0, 8.0) and 50 steps DDIM sampling steps show the relative improvements of the checkpoints: + +![pareto](assets/model-variants.jpg) + +Evaluated using 50 DDIM steps and 10000 random prompts from the COCO2017 validation set, evaluated at 512x512 resolution. Not optimized for FID scores. + +## Environmental Impact + +**Stable Diffusion v1** **Estimated Emissions** +Based on that information, we estimate the following CO2 emissions using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700). The hardware, runtime, cloud provider, and compute region were utilized to estimate the carbon impact. + +- **Hardware Type:** A100 PCIe 40GB +- **Hours used:** 200000 +- **Cloud Provider:** AWS +- **Compute Region:** US-east +- **Carbon Emitted (Power consumption x Time x Carbon produced based on location of power grid):** 15000 kg CO2 eq. + +## Citation + @InProceedings{Rombach_2022_CVPR, + author = {Rombach, Robin and Blattmann, Andreas and Lorenz, Dominik and Esser, Patrick and Ommer, Bj\"orn}, + title = {High-Resolution Image Synthesis With Latent Diffusion Models}, + booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)}, + month = {June}, + year = {2022}, + pages = {10684-10695} + } + +*This model card was written by: Robin Rombach, Patrick Esser and David Ha and is based on the [Stable Diffusion v1](https://github.com/CompVis/stable-diffusion/blob/main/Stable_Diffusion_v1_Model_Card.md) and [DALL-E Mini model card](https://huggingface.co/dalle-mini/dalle-mini).* diff --git a/requirements.txt b/requirements.txt new file mode 100644 index 0000000000000000000000000000000000000000..23b4acb7847ce43c24219bb4579d48056901299c --- /dev/null +++ b/requirements.txt @@ -0,0 +1,19 @@ +albumentations==0.4.3 +opencv-python +pudb==2019.2 +imageio==2.9.0 +imageio-ffmpeg==0.4.2 +pytorch-lightning==1.4.2 +torchmetrics==0.6 +omegaconf==2.1.1 +test-tube>=0.7.5 +streamlit>=0.73.1 +einops==0.3.0 +transformers==4.19.2 +webdataset==0.2.5 +open-clip-torch==2.7.0 +gradio==3.13.2 +kornia==0.6 +invisible-watermark>=0.1.5 +streamlit-drawable-canvas==0.8.0 +-e . diff --git a/scripts/gradio/depth2img.py b/scripts/gradio/depth2img.py new file mode 100644 index 0000000000000000000000000000000000000000..c791a4d0b2a510b3525658f4d852d14704ea9f1a --- /dev/null +++ b/scripts/gradio/depth2img.py @@ -0,0 +1,184 @@ +import sys +import torch +import numpy as np +import gradio as gr +from PIL import Image +from omegaconf import OmegaConf +from einops import repeat, rearrange +from pytorch_lightning import seed_everything +from imwatermark import WatermarkEncoder + +from scripts.txt2img import put_watermark +from ldm.util import instantiate_from_config +from ldm.models.diffusion.ddim import DDIMSampler +from ldm.data.util import AddMiDaS + +torch.set_grad_enabled(False) + + +def initialize_model(config, ckpt): + config = OmegaConf.load(config) + model = instantiate_from_config(config.model) + model.load_state_dict(torch.load(ckpt)["state_dict"], strict=False) + + device = torch.device( + "cuda") if torch.cuda.is_available() else torch.device("cpu") + model = model.to(device) + sampler = DDIMSampler(model) + return sampler + + +def make_batch_sd( + image, + txt, + device, + num_samples=1, + model_type="dpt_hybrid" +): + image = np.array(image.convert("RGB")) + image = torch.from_numpy(image).to(dtype=torch.float32) / 127.5 - 1.0 + # sample['jpg'] is tensor hwc in [-1, 1] at this point + midas_trafo = AddMiDaS(model_type=model_type) + batch = { + "jpg": image, + "txt": num_samples * [txt], + } + batch = midas_trafo(batch) + batch["jpg"] = rearrange(batch["jpg"], 'h w c -> 1 c h w') + batch["jpg"] = repeat(batch["jpg"].to(device=device), + "1 ... -> n ...", n=num_samples) + batch["midas_in"] = repeat(torch.from_numpy(batch["midas_in"][None, ...]).to( + device=device), "1 ... -> n ...", n=num_samples) + return batch + + +def paint(sampler, image, prompt, t_enc, seed, scale, num_samples=1, callback=None, + do_full_sample=False): + device = torch.device( + "cuda") if torch.cuda.is_available() else torch.device("cpu") + model = sampler.model + seed_everything(seed) + + print("Creating invisible watermark encoder (see https://github.com/ShieldMnt/invisible-watermark)...") + wm = "SDV2" + wm_encoder = WatermarkEncoder() + wm_encoder.set_watermark('bytes', wm.encode('utf-8')) + + with torch.no_grad(),\ + torch.autocast("cuda"): + batch = make_batch_sd( + image, txt=prompt, device=device, num_samples=num_samples) + z = model.get_first_stage_encoding(model.encode_first_stage( + batch[model.first_stage_key])) # move to latent space + c = model.cond_stage_model.encode(batch["txt"]) + c_cat = list() + for ck in model.concat_keys: + cc = batch[ck] + cc = model.depth_model(cc) + depth_min, depth_max = torch.amin(cc, dim=[1, 2, 3], keepdim=True), torch.amax(cc, dim=[1, 2, 3], + keepdim=True) + display_depth = (cc - depth_min) / (depth_max - depth_min) + depth_image = Image.fromarray( + (display_depth[0, 0, ...].cpu().numpy() * 255.).astype(np.uint8)) + cc = torch.nn.functional.interpolate( + cc, + size=z.shape[2:], + mode="bicubic", + align_corners=False, + ) + depth_min, depth_max = torch.amin(cc, dim=[1, 2, 3], keepdim=True), torch.amax(cc, dim=[1, 2, 3], + keepdim=True) + cc = 2. * (cc - depth_min) / (depth_max - depth_min) - 1. + c_cat.append(cc) + c_cat = torch.cat(c_cat, dim=1) + # cond + cond = {"c_concat": [c_cat], "c_crossattn": [c]} + + # uncond cond + uc_cross = model.get_unconditional_conditioning(num_samples, "") + uc_full = {"c_concat": [c_cat], "c_crossattn": [uc_cross]} + if not do_full_sample: + # encode (scaled latent) + z_enc = sampler.stochastic_encode( + z, torch.tensor([t_enc] * num_samples).to(model.device)) + else: + z_enc = torch.randn_like(z) + # decode it + samples = sampler.decode(z_enc, cond, t_enc, unconditional_guidance_scale=scale, + unconditional_conditioning=uc_full, callback=callback) + x_samples_ddim = model.decode_first_stage(samples) + result = torch.clamp((x_samples_ddim + 1.0) / 2.0, min=0.0, max=1.0) + result = result.cpu().numpy().transpose(0, 2, 3, 1) * 255 + return [depth_image] + [put_watermark(Image.fromarray(img.astype(np.uint8)), wm_encoder) for img in result] + + +def pad_image(input_image): + pad_w, pad_h = np.max(((2, 2), np.ceil( + np.array(input_image.size) / 64).astype(int)), axis=0) * 64 - input_image.size + im_padded = Image.fromarray( + np.pad(np.array(input_image), ((0, pad_h), (0, pad_w), (0, 0)), mode='edge')) + return im_padded + + +def predict(input_image, prompt, steps, num_samples, scale, seed, eta, strength): + init_image = input_image.convert("RGB") + image = pad_image(init_image) # resize to integer multiple of 32 + + sampler.make_schedule(steps, ddim_eta=eta, verbose=True) + assert 0. <= strength <= 1., 'can only work with strength in [0.0, 1.0]' + do_full_sample = strength == 1. + t_enc = min(int(strength * steps), steps-1) + result = paint( + sampler=sampler, + image=image, + prompt=prompt, + t_enc=t_enc, + seed=seed, + scale=scale, + num_samples=num_samples, + callback=None, + do_full_sample=do_full_sample + ) + return result + + +sampler = initialize_model(sys.argv[1], sys.argv[2]) + +block = gr.Blocks().queue() +with block: + with gr.Row(): + gr.Markdown("## Stable Diffusion Depth2Img") + + with gr.Row(): + with gr.Column(): + input_image = gr.Image(source='upload', type="pil") + prompt = gr.Textbox(label="Prompt") + run_button = gr.Button(label="Run") + with gr.Accordion("Advanced options", open=False): + num_samples = gr.Slider( + label="Images", minimum=1, maximum=4, value=1, step=1) + ddim_steps = gr.Slider(label="Steps", minimum=1, + maximum=50, value=50, step=1) + scale = gr.Slider( + label="Guidance Scale", minimum=0.1, maximum=30.0, value=9.0, step=0.1 + ) + strength = gr.Slider( + label="Strength", minimum=0.0, maximum=1.0, value=0.9, step=0.01 + ) + seed = gr.Slider( + label="Seed", + minimum=0, + maximum=2147483647, + step=1, + randomize=True, + ) + eta = gr.Number(label="eta (DDIM)", value=0.0) + with gr.Column(): + gallery = gr.Gallery(label="Generated images", show_label=False).style( + grid=[2], height="auto") + + run_button.click(fn=predict, inputs=[ + input_image, prompt, ddim_steps, num_samples, scale, seed, eta, strength], outputs=[gallery]) + + +block.launch() diff --git a/scripts/gradio/inpainting.py b/scripts/gradio/inpainting.py new file mode 100644 index 0000000000000000000000000000000000000000..09d44f3ddc528011d7421966915b93d0e2803ba5 --- /dev/null +++ b/scripts/gradio/inpainting.py @@ -0,0 +1,195 @@ +import sys +import cv2 +import torch +import numpy as np +import gradio as gr +from PIL import Image +from omegaconf import OmegaConf +from einops import repeat +from imwatermark import WatermarkEncoder +from pathlib import Path + +from ldm.models.diffusion.ddim import DDIMSampler +from ldm.util import instantiate_from_config + + +torch.set_grad_enabled(False) + + +def put_watermark(img, wm_encoder=None): + if wm_encoder is not None: + img = cv2.cvtColor(np.array(img), cv2.COLOR_RGB2BGR) + img = wm_encoder.encode(img, 'dwtDct') + img = Image.fromarray(img[:, :, ::-1]) + return img + + +def initialize_model(config, ckpt): + config = OmegaConf.load(config) + model = instantiate_from_config(config.model) + + model.load_state_dict(torch.load(ckpt)["state_dict"], strict=False) + + device = torch.device( + "cuda") if torch.cuda.is_available() else torch.device("cpu") + model = model.to(device) + sampler = DDIMSampler(model) + + return sampler + + +def make_batch_sd( + image, + mask, + txt, + device, + num_samples=1): + image = np.array(image.convert("RGB")) + image = image[None].transpose(0, 3, 1, 2) + image = torch.from_numpy(image).to(dtype=torch.float32) / 127.5 - 1.0 + + mask = np.array(mask.convert("L")) + mask = mask.astype(np.float32) / 255.0 + mask = mask[None, None] + mask[mask < 0.5] = 0 + mask[mask >= 0.5] = 1 + mask = torch.from_numpy(mask) + + masked_image = image * (mask < 0.5) + + batch = { + "image": repeat(image.to(device=device), "1 ... -> n ...", n=num_samples), + "txt": num_samples * [txt], + "mask": repeat(mask.to(device=device), "1 ... -> n ...", n=num_samples), + "masked_image": repeat(masked_image.to(device=device), "1 ... -> n ...", n=num_samples), + } + return batch + + +def inpaint(sampler, image, mask, prompt, seed, scale, ddim_steps, num_samples=1, w=512, h=512): + device = torch.device( + "cuda") if torch.cuda.is_available() else torch.device("cpu") + model = sampler.model + + print("Creating invisible watermark encoder (see https://github.com/ShieldMnt/invisible-watermark)...") + wm = "SDV2" + wm_encoder = WatermarkEncoder() + wm_encoder.set_watermark('bytes', wm.encode('utf-8')) + + prng = np.random.RandomState(seed) + start_code = prng.randn(num_samples, 4, h // 8, w // 8) + start_code = torch.from_numpy(start_code).to( + device=device, dtype=torch.float32) + + with torch.no_grad(), \ + torch.autocast("cuda"): + batch = make_batch_sd(image, mask, txt=prompt, + device=device, num_samples=num_samples) + + c = model.cond_stage_model.encode(batch["txt"]) + + c_cat = list() + for ck in model.concat_keys: + cc = batch[ck].float() + if ck != model.masked_image_key: + bchw = [num_samples, 4, h // 8, w // 8] + cc = torch.nn.functional.interpolate(cc, size=bchw[-2:]) + else: + cc = model.get_first_stage_encoding( + model.encode_first_stage(cc)) + c_cat.append(cc) + c_cat = torch.cat(c_cat, dim=1) + + # cond + cond = {"c_concat": [c_cat], "c_crossattn": [c]} + + # uncond cond + uc_cross = model.get_unconditional_conditioning(num_samples, "") + uc_full = {"c_concat": [c_cat], "c_crossattn": [uc_cross]} + + shape = [model.channels, h // 8, w // 8] + samples_cfg, intermediates = sampler.sample( + ddim_steps, + num_samples, + shape, + cond, + verbose=False, + eta=1.0, + unconditional_guidance_scale=scale, + unconditional_conditioning=uc_full, + x_T=start_code, + ) + x_samples_ddim = model.decode_first_stage(samples_cfg) + + result = torch.clamp((x_samples_ddim + 1.0) / 2.0, + min=0.0, max=1.0) + + result = result.cpu().numpy().transpose(0, 2, 3, 1) * 255 + return [put_watermark(Image.fromarray(img.astype(np.uint8)), wm_encoder) for img in result] + +def pad_image(input_image): + pad_w, pad_h = np.max(((2, 2), np.ceil( + np.array(input_image.size) / 64).astype(int)), axis=0) * 64 - input_image.size + im_padded = Image.fromarray( + np.pad(np.array(input_image), ((0, pad_h), (0, pad_w), (0, 0)), mode='edge')) + return im_padded + +def predict(input_image, prompt, ddim_steps, num_samples, scale, seed): + init_image = input_image["image"].convert("RGB") + init_mask = input_image["mask"].convert("RGB") + image = pad_image(init_image) # resize to integer multiple of 32 + mask = pad_image(init_mask) # resize to integer multiple of 32 + width, height = image.size + print("Inpainting...", width, height) + + result = inpaint( + sampler=sampler, + image=image, + mask=mask, + prompt=prompt, + seed=seed, + scale=scale, + ddim_steps=ddim_steps, + num_samples=num_samples, + h=height, w=width + ) + + return result + + +sampler = initialize_model(sys.argv[1], sys.argv[2]) + +block = gr.Blocks().queue() +with block: + with gr.Row(): + gr.Markdown("## Stable Diffusion Inpainting") + + with gr.Row(): + with gr.Column(): + input_image = gr.Image(source='upload', tool='sketch', type="pil") + prompt = gr.Textbox(label="Prompt") + run_button = gr.Button(label="Run") + with gr.Accordion("Advanced options", open=False): + num_samples = gr.Slider( + label="Images", minimum=1, maximum=4, value=4, step=1) + ddim_steps = gr.Slider(label="Steps", minimum=1, + maximum=50, value=45, step=1) + scale = gr.Slider( + label="Guidance Scale", minimum=0.1, maximum=30.0, value=10, step=0.1 + ) + seed = gr.Slider( + label="Seed", + minimum=0, + maximum=2147483647, + step=1, + randomize=True, + ) + with gr.Column(): + gallery = gr.Gallery(label="Generated images", show_label=False).style( + grid=[2], height="auto") + + run_button.click(fn=predict, inputs=[ + input_image, prompt, ddim_steps, num_samples, scale, seed], outputs=[gallery]) + + +block.launch() diff --git a/scripts/gradio/superresolution.py b/scripts/gradio/superresolution.py new file mode 100644 index 0000000000000000000000000000000000000000..3d08fbfae4f9639165e669f3c69c76763c5b32a8 --- /dev/null +++ b/scripts/gradio/superresolution.py @@ -0,0 +1,197 @@ +import sys +import torch +import numpy as np +import gradio as gr +from PIL import Image +from omegaconf import OmegaConf +from einops import repeat, rearrange +from pytorch_lightning import seed_everything +from imwatermark import WatermarkEncoder + +from scripts.txt2img import put_watermark +from ldm.models.diffusion.ddim import DDIMSampler +from ldm.models.diffusion.ddpm import LatentUpscaleDiffusion, LatentUpscaleFinetuneDiffusion +from ldm.util import exists, instantiate_from_config + + +torch.set_grad_enabled(False) + + +def initialize_model(config, ckpt): + config = OmegaConf.load(config) + model = instantiate_from_config(config.model) + model.load_state_dict(torch.load(ckpt)["state_dict"], strict=False) + + device = torch.device( + "cuda") if torch.cuda.is_available() else torch.device("cpu") + model = model.to(device) + sampler = DDIMSampler(model) + return sampler + + +def make_batch_sd( + image, + txt, + device, + num_samples=1, +): + image = np.array(image.convert("RGB")) + image = torch.from_numpy(image).to(dtype=torch.float32) / 127.5 - 1.0 + batch = { + "lr": rearrange(image, 'h w c -> 1 c h w'), + "txt": num_samples * [txt], + } + batch["lr"] = repeat(batch["lr"].to(device=device), + "1 ... -> n ...", n=num_samples) + return batch + + +def make_noise_augmentation(model, batch, noise_level=None): + x_low = batch[model.low_scale_key] + x_low = x_low.to(memory_format=torch.contiguous_format).float() + x_aug, noise_level = model.low_scale_model(x_low, noise_level) + return x_aug, noise_level + + +def paint(sampler, image, prompt, seed, scale, h, w, steps, num_samples=1, callback=None, eta=0., noise_level=None): + device = torch.device( + "cuda") if torch.cuda.is_available() else torch.device("cpu") + model = sampler.model + seed_everything(seed) + prng = np.random.RandomState(seed) + start_code = prng.randn(num_samples, model.channels, h, w) + start_code = torch.from_numpy(start_code).to( + device=device, dtype=torch.float32) + + print("Creating invisible watermark encoder (see https://github.com/ShieldMnt/invisible-watermark)...") + wm = "SDV2" + wm_encoder = WatermarkEncoder() + wm_encoder.set_watermark('bytes', wm.encode('utf-8')) + with torch.no_grad(),\ + torch.autocast("cuda"): + batch = make_batch_sd( + image, txt=prompt, device=device, num_samples=num_samples) + c = model.cond_stage_model.encode(batch["txt"]) + c_cat = list() + if isinstance(model, LatentUpscaleFinetuneDiffusion): + for ck in model.concat_keys: + cc = batch[ck] + if exists(model.reshuffle_patch_size): + assert isinstance(model.reshuffle_patch_size, int) + cc = rearrange(cc, 'b c (p1 h) (p2 w) -> b (p1 p2 c) h w', + p1=model.reshuffle_patch_size, p2=model.reshuffle_patch_size) + c_cat.append(cc) + c_cat = torch.cat(c_cat, dim=1) + # cond + cond = {"c_concat": [c_cat], "c_crossattn": [c]} + # uncond cond + uc_cross = model.get_unconditional_conditioning(num_samples, "") + uc_full = {"c_concat": [c_cat], "c_crossattn": [uc_cross]} + elif isinstance(model, LatentUpscaleDiffusion): + x_augment, noise_level = make_noise_augmentation( + model, batch, noise_level) + cond = {"c_concat": [x_augment], + "c_crossattn": [c], "c_adm": noise_level} + # uncond cond + uc_cross = model.get_unconditional_conditioning(num_samples, "") + uc_full = {"c_concat": [x_augment], "c_crossattn": [ + uc_cross], "c_adm": noise_level} + else: + raise NotImplementedError() + + shape = [model.channels, h, w] + samples, intermediates = sampler.sample( + steps, + num_samples, + shape, + cond, + verbose=False, + eta=eta, + unconditional_guidance_scale=scale, + unconditional_conditioning=uc_full, + x_T=start_code, + callback=callback + ) + with torch.no_grad(): + x_samples_ddim = model.decode_first_stage(samples) + result = torch.clamp((x_samples_ddim + 1.0) / 2.0, min=0.0, max=1.0) + result = result.cpu().numpy().transpose(0, 2, 3, 1) * 255 + return [put_watermark(Image.fromarray(img.astype(np.uint8)), wm_encoder) for img in result] + + +def pad_image(input_image): + pad_w, pad_h = np.max(((2, 2), np.ceil( + np.array(input_image.size) / 64).astype(int)), axis=0) * 64 - input_image.size + im_padded = Image.fromarray( + np.pad(np.array(input_image), ((0, pad_h), (0, pad_w), (0, 0)), mode='edge')) + return im_padded + + +def predict(input_image, prompt, steps, num_samples, scale, seed, eta, noise_level): + init_image = input_image.convert("RGB") + image = pad_image(init_image) # resize to integer multiple of 32 + width, height = image.size + + noise_level = torch.Tensor( + num_samples * [noise_level]).to(sampler.model.device).long() + sampler.make_schedule(steps, ddim_eta=eta, verbose=True) + result = paint( + sampler=sampler, + image=image, + prompt=prompt, + seed=seed, + scale=scale, + h=height, w=width, steps=steps, + num_samples=num_samples, + callback=None, + noise_level=noise_level + ) + return result + + +sampler = initialize_model(sys.argv[1], sys.argv[2]) + +block = gr.Blocks().queue() +with block: + with gr.Row(): + gr.Markdown("## Stable Diffusion Upscaling") + + with gr.Row(): + with gr.Column(): + input_image = gr.Image(source='upload', type="pil") + gr.Markdown( + "Tip: Add a description of the object that should be upscaled, e.g.: 'a professional photograph of a cat") + prompt = gr.Textbox(label="Prompt") + run_button = gr.Button(label="Run") + with gr.Accordion("Advanced options", open=False): + num_samples = gr.Slider( + label="Number of Samples", minimum=1, maximum=4, value=1, step=1) + steps = gr.Slider(label="DDIM Steps", minimum=2, + maximum=200, value=75, step=1) + scale = gr.Slider( + label="Scale", minimum=0.1, maximum=30.0, value=10, step=0.1 + ) + seed = gr.Slider( + label="Seed", + minimum=0, + maximum=2147483647, + step=1, + randomize=True, + ) + eta = gr.Number(label="eta (DDIM)", + value=0.0, min=0.0, max=1.0) + noise_level = None + if isinstance(sampler.model, LatentUpscaleDiffusion): + # TODO: make this work for all models + noise_level = gr.Number( + label="Noise Augmentation", min=0, max=350, value=20, step=1) + + with gr.Column(): + gallery = gr.Gallery(label="Generated images", show_label=False).style( + grid=[2], height="auto") + + run_button.click(fn=predict, inputs=[ + input_image, prompt, steps, num_samples, scale, seed, eta, noise_level], outputs=[gallery]) + + +block.launch() diff --git a/scripts/img2img.py b/scripts/img2img.py new file mode 100644 index 0000000000000000000000000000000000000000..9085ba9d37ea6402b9ee543e82f7d8c56a1c273a --- /dev/null +++ b/scripts/img2img.py @@ -0,0 +1,279 @@ +"""make variations of input image""" + +import argparse, os +import PIL +import torch +import numpy as np +from omegaconf import OmegaConf +from PIL import Image +from tqdm import tqdm, trange +from itertools import islice +from einops import rearrange, repeat +from torchvision.utils import make_grid +from torch import autocast +from contextlib import nullcontext +from pytorch_lightning import seed_everything +from imwatermark import WatermarkEncoder + + +from scripts.txt2img import put_watermark +from ldm.util import instantiate_from_config +from ldm.models.diffusion.ddim import DDIMSampler + + +def chunk(it, size): + it = iter(it) + return iter(lambda: tuple(islice(it, size)), ()) + + +def load_model_from_config(config, ckpt, verbose=False): + print(f"Loading model from {ckpt}") + pl_sd = torch.load(ckpt, map_location="cpu") + if "global_step" in pl_sd: + print(f"Global Step: {pl_sd['global_step']}") + sd = pl_sd["state_dict"] + model = instantiate_from_config(config.model) + m, u = model.load_state_dict(sd, strict=False) + if len(m) > 0 and verbose: + print("missing keys:") + print(m) + if len(u) > 0 and verbose: + print("unexpected keys:") + print(u) + + model.cuda() + model.eval() + return model + + +def load_img(path): + image = Image.open(path).convert("RGB") + w, h = image.size + print(f"loaded input image of size ({w}, {h}) from {path}") + w, h = map(lambda x: x - x % 64, (w, h)) # resize to integer multiple of 64 + image = image.resize((w, h), resample=PIL.Image.LANCZOS) + image = np.array(image).astype(np.float32) / 255.0 + image = image[None].transpose(0, 3, 1, 2) + image = torch.from_numpy(image) + return 2. * image - 1. + + +def main(): + parser = argparse.ArgumentParser() + + parser.add_argument( + "--prompt", + type=str, + nargs="?", + default="a painting of a virus monster playing guitar", + help="the prompt to render" + ) + + parser.add_argument( + "--init-img", + type=str, + nargs="?", + help="path to the input image" + ) + + parser.add_argument( + "--outdir", + type=str, + nargs="?", + help="dir to write results to", + default="outputs/img2img-samples" + ) + + parser.add_argument( + "--ddim_steps", + type=int, + default=50, + help="number of ddim sampling steps", + ) + + parser.add_argument( + "--fixed_code", + action='store_true', + help="if enabled, uses the same starting code across all samples ", + ) + + parser.add_argument( + "--ddim_eta", + type=float, + default=0.0, + help="ddim eta (eta=0.0 corresponds to deterministic sampling", + ) + parser.add_argument( + "--n_iter", + type=int, + default=1, + help="sample this often", + ) + + parser.add_argument( + "--C", + type=int, + default=4, + help="latent channels", + ) + parser.add_argument( + "--f", + type=int, + default=8, + help="downsampling factor, most often 8 or 16", + ) + + parser.add_argument( + "--n_samples", + type=int, + default=2, + help="how many samples to produce for each given prompt. A.k.a batch size", + ) + + parser.add_argument( + "--n_rows", + type=int, + default=0, + help="rows in the grid (default: n_samples)", + ) + + parser.add_argument( + "--scale", + type=float, + default=9.0, + help="unconditional guidance scale: eps = eps(x, empty) + scale * (eps(x, cond) - eps(x, empty))", + ) + + parser.add_argument( + "--strength", + type=float, + default=0.8, + help="strength for noising/unnoising. 1.0 corresponds to full destruction of information in init image", + ) + + parser.add_argument( + "--from-file", + type=str, + help="if specified, load prompts from this file", + ) + parser.add_argument( + "--config", + type=str, + default="configs/stable-diffusion/v2-inference.yaml", + help="path to config which constructs model", + ) + parser.add_argument( + "--ckpt", + type=str, + help="path to checkpoint of model", + ) + parser.add_argument( + "--seed", + type=int, + default=42, + help="the seed (for reproducible sampling)", + ) + parser.add_argument( + "--precision", + type=str, + help="evaluate at this precision", + choices=["full", "autocast"], + default="autocast" + ) + + opt = parser.parse_args() + seed_everything(opt.seed) + + config = OmegaConf.load(f"{opt.config}") + model = load_model_from_config(config, f"{opt.ckpt}") + + device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu") + model = model.to(device) + + sampler = DDIMSampler(model) + + os.makedirs(opt.outdir, exist_ok=True) + outpath = opt.outdir + + print("Creating invisible watermark encoder (see https://github.com/ShieldMnt/invisible-watermark)...") + wm = "SDV2" + wm_encoder = WatermarkEncoder() + wm_encoder.set_watermark('bytes', wm.encode('utf-8')) + + batch_size = opt.n_samples + n_rows = opt.n_rows if opt.n_rows > 0 else batch_size + if not opt.from_file: + prompt = opt.prompt + assert prompt is not None + data = [batch_size * [prompt]] + + else: + print(f"reading prompts from {opt.from_file}") + with open(opt.from_file, "r") as f: + data = f.read().splitlines() + data = list(chunk(data, batch_size)) + + sample_path = os.path.join(outpath, "samples") + os.makedirs(sample_path, exist_ok=True) + base_count = len(os.listdir(sample_path)) + grid_count = len(os.listdir(outpath)) - 1 + + assert os.path.isfile(opt.init_img) + init_image = load_img(opt.init_img).to(device) + init_image = repeat(init_image, '1 ... -> b ...', b=batch_size) + init_latent = model.get_first_stage_encoding(model.encode_first_stage(init_image)) # move to latent space + + sampler.make_schedule(ddim_num_steps=opt.ddim_steps, ddim_eta=opt.ddim_eta, verbose=False) + + assert 0. <= opt.strength <= 1., 'can only work with strength in [0.0, 1.0]' + t_enc = int(opt.strength * opt.ddim_steps) + print(f"target t_enc is {t_enc} steps") + + precision_scope = autocast if opt.precision == "autocast" else nullcontext + with torch.no_grad(): + with precision_scope("cuda"): + with model.ema_scope(): + all_samples = list() + for n in trange(opt.n_iter, desc="Sampling"): + for prompts in tqdm(data, desc="data"): + uc = None + if opt.scale != 1.0: + uc = model.get_learned_conditioning(batch_size * [""]) + if isinstance(prompts, tuple): + prompts = list(prompts) + c = model.get_learned_conditioning(prompts) + + # encode (scaled latent) + z_enc = sampler.stochastic_encode(init_latent, torch.tensor([t_enc] * batch_size).to(device)) + # decode it + samples = sampler.decode(z_enc, c, t_enc, unconditional_guidance_scale=opt.scale, + unconditional_conditioning=uc, ) + + x_samples = model.decode_first_stage(samples) + x_samples = torch.clamp((x_samples + 1.0) / 2.0, min=0.0, max=1.0) + + for x_sample in x_samples: + x_sample = 255. * rearrange(x_sample.cpu().numpy(), 'c h w -> h w c') + img = Image.fromarray(x_sample.astype(np.uint8)) + img = put_watermark(img, wm_encoder) + img.save(os.path.join(sample_path, f"{base_count:05}.png")) + base_count += 1 + all_samples.append(x_samples) + + # additionally, save as grid + grid = torch.stack(all_samples, 0) + grid = rearrange(grid, 'n b c h w -> (n b) c h w') + grid = make_grid(grid, nrow=n_rows) + + # to image + grid = 255. * rearrange(grid, 'c h w -> h w c').cpu().numpy() + grid = Image.fromarray(grid.astype(np.uint8)) + grid = put_watermark(grid, wm_encoder) + grid.save(os.path.join(outpath, f'grid-{grid_count:04}.png')) + grid_count += 1 + + print(f"Your samples are ready and waiting for you here: \n{outpath} \nEnjoy.") + + +if __name__ == "__main__": + main() diff --git a/scripts/streamlit/depth2img.py b/scripts/streamlit/depth2img.py new file mode 100644 index 0000000000000000000000000000000000000000..7f80223405a26ded02964513293b8b3316c34344 --- /dev/null +++ b/scripts/streamlit/depth2img.py @@ -0,0 +1,157 @@ +import sys +import torch +import numpy as np +import streamlit as st +from PIL import Image +from omegaconf import OmegaConf +from einops import repeat, rearrange +from pytorch_lightning import seed_everything +from imwatermark import WatermarkEncoder + +from scripts.txt2img import put_watermark +from ldm.util import instantiate_from_config +from ldm.models.diffusion.ddim import DDIMSampler +from ldm.data.util import AddMiDaS + +torch.set_grad_enabled(False) + + +@st.cache(allow_output_mutation=True) +def initialize_model(config, ckpt): + config = OmegaConf.load(config) + model = instantiate_from_config(config.model) + model.load_state_dict(torch.load(ckpt)["state_dict"], strict=False) + + device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu") + model = model.to(device) + sampler = DDIMSampler(model) + return sampler + + +def make_batch_sd( + image, + txt, + device, + num_samples=1, + model_type="dpt_hybrid" +): + image = np.array(image.convert("RGB")) + image = torch.from_numpy(image).to(dtype=torch.float32) / 127.5 - 1.0 + # sample['jpg'] is tensor hwc in [-1, 1] at this point + midas_trafo = AddMiDaS(model_type=model_type) + batch = { + "jpg": image, + "txt": num_samples * [txt], + } + batch = midas_trafo(batch) + batch["jpg"] = rearrange(batch["jpg"], 'h w c -> 1 c h w') + batch["jpg"] = repeat(batch["jpg"].to(device=device), "1 ... -> n ...", n=num_samples) + batch["midas_in"] = repeat(torch.from_numpy(batch["midas_in"][None, ...]).to(device=device), "1 ... -> n ...", n=num_samples) + return batch + + +def paint(sampler, image, prompt, t_enc, seed, scale, num_samples=1, callback=None, + do_full_sample=False): + device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu") + model = sampler.model + seed_everything(seed) + + print("Creating invisible watermark encoder (see https://github.com/ShieldMnt/invisible-watermark)...") + wm = "SDV2" + wm_encoder = WatermarkEncoder() + wm_encoder.set_watermark('bytes', wm.encode('utf-8')) + + with torch.no_grad(),\ + torch.autocast("cuda"): + batch = make_batch_sd(image, txt=prompt, device=device, num_samples=num_samples) + z = model.get_first_stage_encoding(model.encode_first_stage(batch[model.first_stage_key])) # move to latent space + c = model.cond_stage_model.encode(batch["txt"]) + c_cat = list() + for ck in model.concat_keys: + cc = batch[ck] + cc = model.depth_model(cc) + depth_min, depth_max = torch.amin(cc, dim=[1, 2, 3], keepdim=True), torch.amax(cc, dim=[1, 2, 3], + keepdim=True) + display_depth = (cc - depth_min) / (depth_max - depth_min) + st.image(Image.fromarray((display_depth[0, 0, ...].cpu().numpy() * 255.).astype(np.uint8))) + cc = torch.nn.functional.interpolate( + cc, + size=z.shape[2:], + mode="bicubic", + align_corners=False, + ) + depth_min, depth_max = torch.amin(cc, dim=[1, 2, 3], keepdim=True), torch.amax(cc, dim=[1, 2, 3], + keepdim=True) + cc = 2. * (cc - depth_min) / (depth_max - depth_min) - 1. + c_cat.append(cc) + c_cat = torch.cat(c_cat, dim=1) + # cond + cond = {"c_concat": [c_cat], "c_crossattn": [c]} + + # uncond cond + uc_cross = model.get_unconditional_conditioning(num_samples, "") + uc_full = {"c_concat": [c_cat], "c_crossattn": [uc_cross]} + if not do_full_sample: + # encode (scaled latent) + z_enc = sampler.stochastic_encode(z, torch.tensor([t_enc] * num_samples).to(model.device)) + else: + z_enc = torch.randn_like(z) + # decode it + samples = sampler.decode(z_enc, cond, t_enc, unconditional_guidance_scale=scale, + unconditional_conditioning=uc_full, callback=callback) + x_samples_ddim = model.decode_first_stage(samples) + result = torch.clamp((x_samples_ddim + 1.0) / 2.0, min=0.0, max=1.0) + result = result.cpu().numpy().transpose(0, 2, 3, 1) * 255 + return [put_watermark(Image.fromarray(img.astype(np.uint8)), wm_encoder) for img in result] + + +def run(): + st.title("Stable Diffusion Depth2Img") + # run via streamlit run scripts/demo/depth2img.py + sampler = initialize_model(sys.argv[1], sys.argv[2]) + + image = st.file_uploader("Image", ["jpg", "png"]) + if image: + image = Image.open(image) + w, h = image.size + st.text(f"loaded input image of size ({w}, {h})") + width, height = map(lambda x: x - x % 64, (w, h)) # resize to integer multiple of 64 + image = image.resize((width, height)) + st.text(f"resized input image to size ({width}, {height} (w, h))") + st.image(image) + + prompt = st.text_input("Prompt") + + seed = st.number_input("Seed", min_value=0, max_value=1000000, value=0) + num_samples = st.number_input("Number of Samples", min_value=1, max_value=64, value=1) + scale = st.slider("Scale", min_value=0.1, max_value=30.0, value=9.0, step=0.1) + steps = st.slider("DDIM Steps", min_value=0, max_value=50, value=50, step=1) + strength = st.slider("Strength", min_value=0., max_value=1., value=0.9) + + t_progress = st.progress(0) + def t_callback(t): + t_progress.progress(min((t + 1) / t_enc, 1.)) + + assert 0. <= strength <= 1., 'can only work with strength in [0.0, 1.0]' + do_full_sample = strength == 1. + t_enc = min(int(strength * steps), steps-1) + sampler.make_schedule(steps, ddim_eta=0., verbose=True) + if st.button("Sample"): + result = paint( + sampler=sampler, + image=image, + prompt=prompt, + t_enc=t_enc, + seed=seed, + scale=scale, + num_samples=num_samples, + callback=t_callback, + do_full_sample=do_full_sample, + ) + st.write("Result") + for image in result: + st.image(image, output_format='PNG') + + +if __name__ == "__main__": + run() diff --git a/scripts/streamlit/inpainting.py b/scripts/streamlit/inpainting.py new file mode 100644 index 0000000000000000000000000000000000000000..c35772f063da8bdf0091e3931dbe12b1869dd11a --- /dev/null +++ b/scripts/streamlit/inpainting.py @@ -0,0 +1,195 @@ +import sys +import cv2 +import torch +import numpy as np +import streamlit as st +from PIL import Image +from omegaconf import OmegaConf +from einops import repeat +from streamlit_drawable_canvas import st_canvas +from imwatermark import WatermarkEncoder + +from ldm.models.diffusion.ddim import DDIMSampler +from ldm.util import instantiate_from_config + + +torch.set_grad_enabled(False) + + +def put_watermark(img, wm_encoder=None): + if wm_encoder is not None: + img = cv2.cvtColor(np.array(img), cv2.COLOR_RGB2BGR) + img = wm_encoder.encode(img, 'dwtDct') + img = Image.fromarray(img[:, :, ::-1]) + return img + + +@st.cache(allow_output_mutation=True) +def initialize_model(config, ckpt): + config = OmegaConf.load(config) + model = instantiate_from_config(config.model) + + model.load_state_dict(torch.load(ckpt)["state_dict"], strict=False) + + device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu") + model = model.to(device) + sampler = DDIMSampler(model) + + return sampler + + +def make_batch_sd( + image, + mask, + txt, + device, + num_samples=1): + image = np.array(image.convert("RGB")) + image = image[None].transpose(0, 3, 1, 2) + image = torch.from_numpy(image).to(dtype=torch.float32) / 127.5 - 1.0 + + mask = np.array(mask.convert("L")) + mask = mask.astype(np.float32) / 255.0 + mask = mask[None, None] + mask[mask < 0.5] = 0 + mask[mask >= 0.5] = 1 + mask = torch.from_numpy(mask) + + masked_image = image * (mask < 0.5) + + batch = { + "image": repeat(image.to(device=device), "1 ... -> n ...", n=num_samples), + "txt": num_samples * [txt], + "mask": repeat(mask.to(device=device), "1 ... -> n ...", n=num_samples), + "masked_image": repeat(masked_image.to(device=device), "1 ... -> n ...", n=num_samples), + } + return batch + + +def inpaint(sampler, image, mask, prompt, seed, scale, ddim_steps, num_samples=1, w=512, h=512, eta=1.): + device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu") + model = sampler.model + + print("Creating invisible watermark encoder (see https://github.com/ShieldMnt/invisible-watermark)...") + wm = "SDV2" + wm_encoder = WatermarkEncoder() + wm_encoder.set_watermark('bytes', wm.encode('utf-8')) + + prng = np.random.RandomState(seed) + start_code = prng.randn(num_samples, 4, h // 8, w // 8) + start_code = torch.from_numpy(start_code).to(device=device, dtype=torch.float32) + + with torch.no_grad(), \ + torch.autocast("cuda"): + batch = make_batch_sd(image, mask, txt=prompt, device=device, num_samples=num_samples) + + c = model.cond_stage_model.encode(batch["txt"]) + + c_cat = list() + for ck in model.concat_keys: + cc = batch[ck].float() + if ck != model.masked_image_key: + bchw = [num_samples, 4, h // 8, w // 8] + cc = torch.nn.functional.interpolate(cc, size=bchw[-2:]) + else: + cc = model.get_first_stage_encoding(model.encode_first_stage(cc)) + c_cat.append(cc) + c_cat = torch.cat(c_cat, dim=1) + + # cond + cond = {"c_concat": [c_cat], "c_crossattn": [c]} + + # uncond cond + uc_cross = model.get_unconditional_conditioning(num_samples, "") + uc_full = {"c_concat": [c_cat], "c_crossattn": [uc_cross]} + + shape = [model.channels, h // 8, w // 8] + samples_cfg, intermediates = sampler.sample( + ddim_steps, + num_samples, + shape, + cond, + verbose=False, + eta=eta, + unconditional_guidance_scale=scale, + unconditional_conditioning=uc_full, + x_T=start_code, + ) + x_samples_ddim = model.decode_first_stage(samples_cfg) + + result = torch.clamp((x_samples_ddim + 1.0) / 2.0, + min=0.0, max=1.0) + + result = result.cpu().numpy().transpose(0, 2, 3, 1) * 255 + return [put_watermark(Image.fromarray(img.astype(np.uint8)), wm_encoder) for img in result] + + +def run(): + st.title("Stable Diffusion Inpainting") + + sampler = initialize_model(sys.argv[1], sys.argv[2]) + + image = st.file_uploader("Image", ["jpg", "png"]) + if image: + image = Image.open(image) + w, h = image.size + print(f"loaded input image of size ({w}, {h})") + width, height = map(lambda x: x - x % 64, (w, h)) # resize to integer multiple of 32 + image = image.resize((width, height)) + + prompt = st.text_input("Prompt") + + seed = st.number_input("Seed", min_value=0, max_value=1000000, value=0) + num_samples = st.number_input("Number of Samples", min_value=1, max_value=64, value=1) + scale = st.slider("Scale", min_value=0.1, max_value=30.0, value=10., step=0.1) + ddim_steps = st.slider("DDIM Steps", min_value=0, max_value=50, value=50, step=1) + eta = st.sidebar.number_input("eta (DDIM)", value=0., min_value=0., max_value=1.) + + fill_color = "rgba(255, 255, 255, 0.0)" + stroke_width = st.number_input("Brush Size", + value=64, + min_value=1, + max_value=100) + stroke_color = "rgba(255, 255, 255, 1.0)" + bg_color = "rgba(0, 0, 0, 1.0)" + drawing_mode = "freedraw" + + st.write("Canvas") + st.caption( + "Draw a mask to inpaint, then click the 'Send to Streamlit' button (bottom left, with an arrow on it).") + canvas_result = st_canvas( + fill_color=fill_color, + stroke_width=stroke_width, + stroke_color=stroke_color, + background_color=bg_color, + background_image=image, + update_streamlit=False, + height=height, + width=width, + drawing_mode=drawing_mode, + key="canvas", + ) + if canvas_result: + mask = canvas_result.image_data + mask = mask[:, :, -1] > 0 + if mask.sum() > 0: + mask = Image.fromarray(mask) + + result = inpaint( + sampler=sampler, + image=image, + mask=mask, + prompt=prompt, + seed=seed, + scale=scale, + ddim_steps=ddim_steps, + num_samples=num_samples, + h=height, w=width, eta=eta + ) + st.write("Inpainted") + for image in result: + st.image(image, output_format='PNG') + + +if __name__ == "__main__": + run() \ No newline at end of file diff --git a/scripts/streamlit/stableunclip.py b/scripts/streamlit/stableunclip.py new file mode 100644 index 0000000000000000000000000000000000000000..122fa9a584b1961681883709bc6672bf90339cc5 --- /dev/null +++ b/scripts/streamlit/stableunclip.py @@ -0,0 +1,416 @@ +import importlib +import streamlit as st +import torch +import cv2 +import numpy as np +import PIL +from omegaconf import OmegaConf +from PIL import Image +from tqdm import trange +import io, os +from torch import autocast +from einops import rearrange, repeat +from torchvision.utils import make_grid +from pytorch_lightning import seed_everything +from contextlib import nullcontext + +from ldm.models.diffusion.ddim import DDIMSampler +from ldm.models.diffusion.plms import PLMSSampler +from ldm.models.diffusion.dpm_solver import DPMSolverSampler + +torch.set_grad_enabled(False) + +PROMPTS_ROOT = "scripts/prompts/" +SAVE_PATH = "outputs/demo/stable-unclip/" + +VERSION2SPECS = { + "Stable unCLIP-L": {"H": 768, "W": 768, "C": 4, "f": 8}, + "Stable unOpenCLIP-H": {"H": 768, "W": 768, "C": 4, "f": 8}, + "Full Karlo": {} +} + + +def get_obj_from_str(string, reload=False): + module, cls = string.rsplit(".", 1) + importlib.invalidate_caches() + if reload: + module_imp = importlib.import_module(module) + importlib.reload(module_imp) + return getattr(importlib.import_module(module, package=None), cls) + + +def instantiate_from_config(config): + if not "target" in config: + raise KeyError("Expected key `target` to instantiate.") + return get_obj_from_str(config["target"])(**config.get("params", dict())) + + +def get_interactive_image(key=None): + image = st.file_uploader("Input", type=["jpg", "JPEG", "png"], key=key) + if image is not None: + image = Image.open(image) + if not image.mode == "RGB": + image = image.convert("RGB") + return image + + +def load_img(display=True, key=None): + image = get_interactive_image(key=key) + if display: + st.image(image) + w, h = image.size + print(f"loaded input image of size ({w}, {h})") + w, h = map(lambda x: x - x % 64, (w, h)) + image = image.resize((w, h), resample=PIL.Image.LANCZOS) + image = np.array(image).astype(np.float32) / 255.0 + image = image[None].transpose(0, 3, 1, 2) + image = torch.from_numpy(image) + return 2. * image - 1. + + +def get_init_img(batch_size=1, key=None): + init_image = load_img(key=key).cuda() + init_image = repeat(init_image, '1 ... -> b ...', b=batch_size) + return init_image + + +def sample( + model, + prompt, + n_runs=3, + n_samples=2, + H=512, + W=512, + C=4, + f=8, + scale=10.0, + ddim_steps=50, + ddim_eta=0.0, + callback=None, + skip_single_save=False, + save_grid=True, + ucg_schedule=None, + negative_prompt="", + adm_cond=None, + adm_uc=None, + use_full_precision=False, + only_adm_cond=False +): + batch_size = n_samples + precision_scope = autocast if not use_full_precision else nullcontext + # decoderscope = autocast if not use_full_precision else nullcontext + if use_full_precision: st.warning(f"Running {model.__class__.__name__} at full precision.") + if isinstance(prompt, str): + prompt = [prompt] + prompts = batch_size * prompt + + outputs = st.empty() + + with precision_scope("cuda"): + with model.ema_scope(): + all_samples = list() + for n in trange(n_runs, desc="Sampling"): + shape = [C, H // f, W // f] + if not only_adm_cond: + uc = None + if scale != 1.0: + uc = model.get_learned_conditioning(batch_size * [negative_prompt]) + if isinstance(prompts, tuple): + prompts = list(prompts) + c = model.get_learned_conditioning(prompts) + + if adm_cond is not None: + if adm_cond.shape[0] == 1: + adm_cond = repeat(adm_cond, '1 ... -> b ...', b=batch_size) + if adm_uc is None: + st.warning("Not guiding via c_adm") + adm_uc = adm_cond + else: + if adm_uc.shape[0] == 1: + adm_uc = repeat(adm_uc, '1 ... -> b ...', b=batch_size) + if not only_adm_cond: + c = {"c_crossattn": [c], "c_adm": adm_cond} + uc = {"c_crossattn": [uc], "c_adm": adm_uc} + else: + c = adm_cond + uc = adm_uc + samples_ddim, _ = sampler.sample(S=ddim_steps, + conditioning=c, + batch_size=batch_size, + shape=shape, + verbose=False, + unconditional_guidance_scale=scale, + unconditional_conditioning=uc, + eta=ddim_eta, + x_T=None, + callback=callback, + ucg_schedule=ucg_schedule + ) + x_samples = model.decode_first_stage(samples_ddim) + x_samples = torch.clamp((x_samples + 1.0) / 2.0, min=0.0, max=1.0) + + if not skip_single_save: + base_count = len(os.listdir(os.path.join(SAVE_PATH, "samples"))) + for x_sample in x_samples: + x_sample = 255. * rearrange(x_sample.cpu().numpy(), 'c h w -> h w c') + Image.fromarray(x_sample.astype(np.uint8)).save( + os.path.join(SAVE_PATH, "samples", f"{base_count:09}.png")) + base_count += 1 + + all_samples.append(x_samples) + + # get grid of all samples + grid = torch.stack(all_samples, 0) + grid = rearrange(grid, 'n b c h w -> (n h) (b w) c') + outputs.image(grid.cpu().numpy()) + + # additionally, save grid + grid = Image.fromarray((255. * grid.cpu().numpy()).astype(np.uint8)) + if save_grid: + grid_count = len(os.listdir(SAVE_PATH)) - 1 + grid.save(os.path.join(SAVE_PATH, f'grid-{grid_count:06}.png')) + + return x_samples + + +def make_oscillating_guidance_schedule(num_steps, max_weight=15., min_weight=1.): + schedule = list() + for i in range(num_steps): + if float(i / num_steps) < 0.1: + schedule.append(max_weight) + elif i % 2 == 0: + schedule.append(min_weight) + else: + schedule.append(max_weight) + print(f"OSCILLATING GUIDANCE SCHEDULE: \n {schedule}") + return schedule + + +def torch2np(x): + x = ((x + 1.0) * 127.5).clamp(0, 255).to(dtype=torch.uint8) + x = x.permute(0, 2, 3, 1).detach().cpu().numpy() + return x + + +@st.cache(allow_output_mutation=True, suppress_st_warning=True) +def init(version="Stable unCLIP-L", load_karlo_prior=False): + state = dict() + if not "model" in state: + if version == "Stable unCLIP-L": + config = "configs/stable-diffusion/v2-1-stable-unclip-l-inference.yaml" + ckpt = "checkpoints/sd21-unclip-l.ckpt" + + elif version == "Stable unOpenCLIP-H": + config = "configs/stable-diffusion/v2-1-stable-unclip-h-inference.yaml" + ckpt = "checkpoints/sd21-unclip-h.ckpt" + + elif version == "Full Karlo": + from ldm.modules.karlo.kakao.sampler import T2ISampler + st.info("Loading full KARLO..") + karlo = T2ISampler.from_pretrained( + root_dir="checkpoints/karlo_models", + clip_model_path="ViT-L-14.pt", + clip_stat_path="ViT-L-14_stats.th", + sampling_type="default", + ) + state["karlo_prior"] = karlo + state["msg"] = "loaded full Karlo" + return state + else: + raise ValueError(f"version {version} unknown!") + + config = OmegaConf.load(config) + model, msg = load_model_from_config(config, ckpt, vae_sd=None) + state["msg"] = msg + + if load_karlo_prior: + from ldm.modules.karlo.kakao.sampler import PriorSampler + st.info("Loading KARLO CLIP prior...") + karlo_prior = PriorSampler.from_pretrained( + root_dir="checkpoints/karlo_models", + clip_model_path="ViT-L-14.pt", + clip_stat_path="ViT-L-14_stats.th", + sampling_type="default", + ) + state["karlo_prior"] = karlo_prior + state["model"] = model + state["ckpt"] = ckpt + state["config"] = config + return state + + +def load_model_from_config(config, ckpt, verbose=False, vae_sd=None): + print(f"Loading model from {ckpt}") + pl_sd = torch.load(ckpt, map_location="cpu") + msg = None + if "global_step" in pl_sd: + msg = f"This is global step {pl_sd['global_step']}. " + if "model_ema.num_updates" in pl_sd["state_dict"]: + msg += f"And we got {pl_sd['state_dict']['model_ema.num_updates']} EMA updates." + global_step = pl_sd.get("global_step", "?") + sd = pl_sd["state_dict"] + if vae_sd is not None: + for k in sd.keys(): + if "first_stage" in k: + sd[k] = vae_sd[k[len("first_stage_model."):]] + + model = instantiate_from_config(config.model) + m, u = model.load_state_dict(sd, strict=False) + if len(m) > 0 and verbose: + print("missing keys:") + print(m) + if len(u) > 0 and verbose: + print("unexpected keys:") + print(u) + + model.cuda() + model.eval() + print(f"Loaded global step {global_step}") + return model, msg + + +if __name__ == "__main__": + st.title("Stable unCLIP") + mode = "txt2img" + version = st.selectbox("Model Version", list(VERSION2SPECS.keys()), 0) + use_karlo_prior = version in ["Stable unCLIP-L"] and st.checkbox("Use KARLO prior", False) + state = init(version=version, load_karlo_prior=use_karlo_prior) + prompt = st.text_input("Prompt", "a professional photograph") + negative_prompt = st.text_input("Negative Prompt", "") + scale = st.number_input("cfg-scale", value=10., min_value=-100., max_value=100.) + number_rows = st.number_input("num rows", value=2, min_value=1, max_value=10) + number_cols = st.number_input("num cols", value=2, min_value=1, max_value=10) + steps = st.sidebar.number_input("steps", value=20, min_value=1, max_value=1000) + eta = st.sidebar.number_input("eta (DDIM)", value=0., min_value=0., max_value=1.) + force_full_precision = st.sidebar.checkbox("Force FP32", False) # TODO: check if/where things break. + if version != "Full Karlo": + H = st.sidebar.number_input("H", value=VERSION2SPECS[version]["H"], min_value=64, max_value=2048) + W = st.sidebar.number_input("W", value=VERSION2SPECS[version]["W"], min_value=64, max_value=2048) + C = VERSION2SPECS[version]["C"] + f = VERSION2SPECS[version]["f"] + + SAVE_PATH = os.path.join(SAVE_PATH, version) + os.makedirs(os.path.join(SAVE_PATH, "samples"), exist_ok=True) + + seed = st.sidebar.number_input("seed", value=42, min_value=0, max_value=int(1e9)) + seed_everything(seed) + + ucg_schedule = None + sampler = st.sidebar.selectbox("Sampler", ["DDIM", "DPM"], 0) + if version == "Full Karlo": + pass + else: + if sampler == "DPM": + sampler = DPMSolverSampler(state["model"]) + elif sampler == "DDIM": + sampler = DDIMSampler(state["model"]) + else: + raise ValueError(f"unknown sampler {sampler}!") + + adm_cond, adm_uc = None, None + if use_karlo_prior: + # uses the prior + karlo_sampler = state["karlo_prior"] + noise_level = None + if state["model"].noise_augmentor is not None: + noise_level = st.number_input("Noise Augmentation for CLIP embeddings", min_value=0, + max_value=state["model"].noise_augmentor.max_noise_level - 1, value=0) + with torch.no_grad(): + karlo_prediction = iter( + karlo_sampler( + prompt=prompt, + bsz=number_cols, + progressive_mode="final", + ) + ).__next__() + adm_cond = karlo_prediction + if noise_level is not None: + c_adm, noise_level_emb = state["model"].noise_augmentor(adm_cond, noise_level=repeat( + torch.tensor([noise_level]).to(state["model"].device), '1 -> b', b=number_cols)) + adm_cond = torch.cat((c_adm, noise_level_emb), 1) + adm_uc = torch.zeros_like(adm_cond) + elif version == "Full Karlo": + pass + else: + num_inputs = st.number_input("Number of Input Images", 1) + + + def make_conditionings_from_input(num=1, key=None): + init_img = get_init_img(batch_size=number_cols, key=key) + with torch.no_grad(): + adm_cond = state["model"].embedder(init_img) + weight = st.slider(f"Weight for Input {num}", min_value=-10., max_value=10., value=1.) + if state["model"].noise_augmentor is not None: + noise_level = st.number_input(f"Noise Augmentation for CLIP embedding of input #{num}", min_value=0, + max_value=state["model"].noise_augmentor.max_noise_level - 1, + value=0, ) + c_adm, noise_level_emb = state["model"].noise_augmentor(adm_cond, noise_level=repeat( + torch.tensor([noise_level]).to(state["model"].device), '1 -> b', b=number_cols)) + adm_cond = torch.cat((c_adm, noise_level_emb), 1) * weight + adm_uc = torch.zeros_like(adm_cond) + return adm_cond, adm_uc, weight + + + adm_inputs = list() + weights = list() + for n in range(num_inputs): + adm_cond, adm_uc, w = make_conditionings_from_input(num=n + 1, key=n) + weights.append(w) + adm_inputs.append(adm_cond) + adm_cond = torch.stack(adm_inputs).sum(0) / sum(weights) + if num_inputs > 1: + if st.checkbox("Apply Noise to Embedding Mix", True): + noise_level = st.number_input(f"Noise Augmentation for averaged CLIP embeddings", min_value=0, + max_value=state["model"].noise_augmentor.max_noise_level - 1, value=50, ) + c_adm, noise_level_emb = state["model"].noise_augmentor( + adm_cond[:, :state["model"].noise_augmentor.time_embed.dim], + noise_level=repeat( + torch.tensor([noise_level]).to(state["model"].device), '1 -> b', b=number_cols)) + adm_cond = torch.cat((c_adm, noise_level_emb), 1) + + if st.button("Sample"): + print("running prompt:", prompt) + st.text("Sampling") + t_progress = st.progress(0) + result = st.empty() + + + def t_callback(t): + t_progress.progress(min((t + 1) / steps, 1.)) + + + if version == "Full Karlo": + outputs = st.empty() + karlo_sampler = state["karlo_prior"] + all_samples = list() + with torch.no_grad(): + for _ in range(number_rows): + karlo_prediction = iter( + karlo_sampler( + prompt=prompt, + bsz=number_cols, + progressive_mode="final", + ) + ).__next__() + all_samples.append(karlo_prediction) + grid = torch.stack(all_samples, 0) + grid = rearrange(grid, 'n b c h w -> (n h) (b w) c') + outputs.image(grid.cpu().numpy()) + + else: + samples = sample( + state["model"], + prompt, + n_runs=number_rows, + n_samples=number_cols, + H=H, W=W, C=C, f=f, + scale=scale, + ddim_steps=steps, + ddim_eta=eta, + callback=t_callback, + ucg_schedule=ucg_schedule, + negative_prompt=negative_prompt, + adm_cond=adm_cond, adm_uc=adm_uc, + use_full_precision=force_full_precision, + only_adm_cond=False + ) diff --git a/scripts/streamlit/superresolution.py b/scripts/streamlit/superresolution.py new file mode 100644 index 0000000000000000000000000000000000000000..c1172b02ea8141781118d4536ba28de0f24404a1 --- /dev/null +++ b/scripts/streamlit/superresolution.py @@ -0,0 +1,170 @@ +import sys +import torch +import numpy as np +import streamlit as st +from PIL import Image +from omegaconf import OmegaConf +from einops import repeat, rearrange +from pytorch_lightning import seed_everything +from imwatermark import WatermarkEncoder + +from scripts.txt2img import put_watermark +from ldm.models.diffusion.ddim import DDIMSampler +from ldm.models.diffusion.ddpm import LatentUpscaleDiffusion, LatentUpscaleFinetuneDiffusion +from ldm.util import exists, instantiate_from_config + + +torch.set_grad_enabled(False) + + +@st.cache(allow_output_mutation=True) +def initialize_model(config, ckpt): + config = OmegaConf.load(config) + model = instantiate_from_config(config.model) + model.load_state_dict(torch.load(ckpt)["state_dict"], strict=False) + + device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu") + model = model.to(device) + sampler = DDIMSampler(model) + return sampler + + +def make_batch_sd( + image, + txt, + device, + num_samples=1, +): + image = np.array(image.convert("RGB")) + image = torch.from_numpy(image).to(dtype=torch.float32) / 127.5 - 1.0 + batch = { + "lr": rearrange(image, 'h w c -> 1 c h w'), + "txt": num_samples * [txt], + } + batch["lr"] = repeat(batch["lr"].to(device=device), "1 ... -> n ...", n=num_samples) + return batch + + +def make_noise_augmentation(model, batch, noise_level=None): + x_low = batch[model.low_scale_key] + x_low = x_low.to(memory_format=torch.contiguous_format).float() + x_aug, noise_level = model.low_scale_model(x_low, noise_level) + return x_aug, noise_level + + +def paint(sampler, image, prompt, seed, scale, h, w, steps, num_samples=1, callback=None, eta=0., noise_level=None): + device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu") + model = sampler.model + seed_everything(seed) + prng = np.random.RandomState(seed) + start_code = prng.randn(num_samples, model.channels, h , w) + start_code = torch.from_numpy(start_code).to(device=device, dtype=torch.float32) + + print("Creating invisible watermark encoder (see https://github.com/ShieldMnt/invisible-watermark)...") + wm = "SDV2" + wm_encoder = WatermarkEncoder() + wm_encoder.set_watermark('bytes', wm.encode('utf-8')) + with torch.no_grad(),\ + torch.autocast("cuda"): + batch = make_batch_sd(image, txt=prompt, device=device, num_samples=num_samples) + c = model.cond_stage_model.encode(batch["txt"]) + c_cat = list() + if isinstance(model, LatentUpscaleFinetuneDiffusion): + for ck in model.concat_keys: + cc = batch[ck] + if exists(model.reshuffle_patch_size): + assert isinstance(model.reshuffle_patch_size, int) + cc = rearrange(cc, 'b c (p1 h) (p2 w) -> b (p1 p2 c) h w', + p1=model.reshuffle_patch_size, p2=model.reshuffle_patch_size) + c_cat.append(cc) + c_cat = torch.cat(c_cat, dim=1) + # cond + cond = {"c_concat": [c_cat], "c_crossattn": [c]} + # uncond cond + uc_cross = model.get_unconditional_conditioning(num_samples, "") + uc_full = {"c_concat": [c_cat], "c_crossattn": [uc_cross]} + elif isinstance(model, LatentUpscaleDiffusion): + x_augment, noise_level = make_noise_augmentation(model, batch, noise_level) + cond = {"c_concat": [x_augment], "c_crossattn": [c], "c_adm": noise_level} + # uncond cond + uc_cross = model.get_unconditional_conditioning(num_samples, "") + uc_full = {"c_concat": [x_augment], "c_crossattn": [uc_cross], "c_adm": noise_level} + else: + raise NotImplementedError() + + shape = [model.channels, h, w] + samples, intermediates = sampler.sample( + steps, + num_samples, + shape, + cond, + verbose=False, + eta=eta, + unconditional_guidance_scale=scale, + unconditional_conditioning=uc_full, + x_T=start_code, + callback=callback + ) + with torch.no_grad(): + x_samples_ddim = model.decode_first_stage(samples) + result = torch.clamp((x_samples_ddim + 1.0) / 2.0, min=0.0, max=1.0) + result = result.cpu().numpy().transpose(0, 2, 3, 1) * 255 + st.text(f"upscaled image shape: {result.shape}") + return [put_watermark(Image.fromarray(img.astype(np.uint8)), wm_encoder) for img in result] + + +def run(): + st.title("Stable Diffusion Upscaling") + # run via streamlit run scripts/demo/depth2img.py + sampler = initialize_model(sys.argv[1], sys.argv[2]) + + image = st.file_uploader("Image", ["jpg", "png"]) + if image: + image = Image.open(image) + w, h = image.size + st.text(f"loaded input image of size ({w}, {h})") + width, height = map(lambda x: x - x % 64, (w, h)) # resize to integer multiple of 64 + image = image.resize((width, height)) + st.text(f"resized input image to size ({width}, {height} (w, h))") + st.image(image) + + st.write(f"\n Tip: Add a description of the object that should be upscaled, e.g.: 'a professional photograph of a cat'") + prompt = st.text_input("Prompt", "a high quality professional photograph") + + seed = st.number_input("Seed", min_value=0, max_value=1000000, value=0) + num_samples = st.number_input("Number of Samples", min_value=1, max_value=64, value=1) + scale = st.slider("Scale", min_value=0.1, max_value=30.0, value=9.0, step=0.1) + steps = st.slider("DDIM Steps", min_value=2, max_value=250, value=50, step=1) + eta = st.sidebar.number_input("eta (DDIM)", value=0., min_value=0., max_value=1.) + + noise_level = None + if isinstance(sampler.model, LatentUpscaleDiffusion): + # TODO: make this work for all models + noise_level = st.sidebar.number_input("Noise Augmentation", min_value=0, max_value=350, value=20) + noise_level = torch.Tensor(num_samples * [noise_level]).to(sampler.model.device).long() + + t_progress = st.progress(0) + def t_callback(t): + t_progress.progress(min((t + 1) / steps, 1.)) + + sampler.make_schedule(steps, ddim_eta=eta, verbose=True) + if st.button("Sample"): + result = paint( + sampler=sampler, + image=image, + prompt=prompt, + seed=seed, + scale=scale, + h=height, w=width, steps=steps, + num_samples=num_samples, + callback=t_callback, + noise_level=noise_level, + eta=eta + ) + st.write("Result") + for image in result: + st.image(image, output_format='PNG') + + +if __name__ == "__main__": + run() diff --git a/scripts/tests/test_watermark.py b/scripts/tests/test_watermark.py new file mode 100644 index 0000000000000000000000000000000000000000..f93f8a6e70763c0e284157bc8225827520b2f5ef --- /dev/null +++ b/scripts/tests/test_watermark.py @@ -0,0 +1,18 @@ +import cv2 +import fire +from imwatermark import WatermarkDecoder + + +def testit(img_path): + bgr = cv2.imread(img_path) + decoder = WatermarkDecoder('bytes', 136) + watermark = decoder.decode(bgr, 'dwtDct') + try: + dec = watermark.decode('utf-8') + except: + dec = "null" + print(dec) + + +if __name__ == "__main__": + fire.Fire(testit) \ No newline at end of file diff --git a/scripts/txt2img.py b/scripts/txt2img.py new file mode 100644 index 0000000000000000000000000000000000000000..9d955e3dc7d562f0809fd3d89480dab2ee2d8ea0 --- /dev/null +++ b/scripts/txt2img.py @@ -0,0 +1,388 @@ +import argparse, os +import cv2 +import torch +import numpy as np +from omegaconf import OmegaConf +from PIL import Image +from tqdm import tqdm, trange +from itertools import islice +from einops import rearrange +from torchvision.utils import make_grid +from pytorch_lightning import seed_everything +from torch import autocast +from contextlib import nullcontext +from imwatermark import WatermarkEncoder + +from ldm.util import instantiate_from_config +from ldm.models.diffusion.ddim import DDIMSampler +from ldm.models.diffusion.plms import PLMSSampler +from ldm.models.diffusion.dpm_solver import DPMSolverSampler + +torch.set_grad_enabled(False) + +def chunk(it, size): + it = iter(it) + return iter(lambda: tuple(islice(it, size)), ()) + + +def load_model_from_config(config, ckpt, device=torch.device("cuda"), verbose=False): + print(f"Loading model from {ckpt}") + pl_sd = torch.load(ckpt, map_location="cpu") + if "global_step" in pl_sd: + print(f"Global Step: {pl_sd['global_step']}") + sd = pl_sd["state_dict"] + model = instantiate_from_config(config.model) + m, u = model.load_state_dict(sd, strict=False) + if len(m) > 0 and verbose: + print("missing keys:") + print(m) + if len(u) > 0 and verbose: + print("unexpected keys:") + print(u) + + if device == torch.device("cuda"): + model.cuda() + elif device == torch.device("cpu"): + model.cpu() + model.cond_stage_model.device = "cpu" + else: + raise ValueError(f"Incorrect device name. Received: {device}") + model.eval() + return model + + +def parse_args(): + parser = argparse.ArgumentParser() + parser.add_argument( + "--prompt", + type=str, + nargs="?", + default="a professional photograph of an astronaut riding a triceratops", + help="the prompt to render" + ) + parser.add_argument( + "--outdir", + type=str, + nargs="?", + help="dir to write results to", + default="outputs/txt2img-samples" + ) + parser.add_argument( + "--steps", + type=int, + default=50, + help="number of ddim sampling steps", + ) + parser.add_argument( + "--plms", + action='store_true', + help="use plms sampling", + ) + parser.add_argument( + "--dpm", + action='store_true', + help="use DPM (2) sampler", + ) + parser.add_argument( + "--fixed_code", + action='store_true', + help="if enabled, uses the same starting code across all samples ", + ) + parser.add_argument( + "--ddim_eta", + type=float, + default=0.0, + help="ddim eta (eta=0.0 corresponds to deterministic sampling", + ) + parser.add_argument( + "--n_iter", + type=int, + default=3, + help="sample this often", + ) + parser.add_argument( + "--H", + type=int, + default=512, + help="image height, in pixel space", + ) + parser.add_argument( + "--W", + type=int, + default=512, + help="image width, in pixel space", + ) + parser.add_argument( + "--C", + type=int, + default=4, + help="latent channels", + ) + parser.add_argument( + "--f", + type=int, + default=8, + help="downsampling factor, most often 8 or 16", + ) + parser.add_argument( + "--n_samples", + type=int, + default=3, + help="how many samples to produce for each given prompt. A.k.a batch size", + ) + parser.add_argument( + "--n_rows", + type=int, + default=0, + help="rows in the grid (default: n_samples)", + ) + parser.add_argument( + "--scale", + type=float, + default=9.0, + help="unconditional guidance scale: eps = eps(x, empty) + scale * (eps(x, cond) - eps(x, empty))", + ) + parser.add_argument( + "--from-file", + type=str, + help="if specified, load prompts from this file, separated by newlines", + ) + parser.add_argument( + "--config", + type=str, + default="configs/stable-diffusion/v2-inference.yaml", + help="path to config which constructs model", + ) + parser.add_argument( + "--ckpt", + type=str, + help="path to checkpoint of model", + ) + parser.add_argument( + "--seed", + type=int, + default=42, + help="the seed (for reproducible sampling)", + ) + parser.add_argument( + "--precision", + type=str, + help="evaluate at this precision", + choices=["full", "autocast"], + default="autocast" + ) + parser.add_argument( + "--repeat", + type=int, + default=1, + help="repeat each prompt in file this often", + ) + parser.add_argument( + "--device", + type=str, + help="Device on which Stable Diffusion will be run", + choices=["cpu", "cuda"], + default="cpu" + ) + parser.add_argument( + "--torchscript", + action='store_true', + help="Use TorchScript", + ) + parser.add_argument( + "--ipex", + action='store_true', + help="Use Intel® Extension for PyTorch*", + ) + parser.add_argument( + "--bf16", + action='store_true', + help="Use bfloat16", + ) + opt = parser.parse_args() + return opt + + +def put_watermark(img, wm_encoder=None): + if wm_encoder is not None: + img = cv2.cvtColor(np.array(img), cv2.COLOR_RGB2BGR) + img = wm_encoder.encode(img, 'dwtDct') + img = Image.fromarray(img[:, :, ::-1]) + return img + + +def main(opt): + seed_everything(opt.seed) + + config = OmegaConf.load(f"{opt.config}") + device = torch.device("cuda") if opt.device == "cuda" else torch.device("cpu") + model = load_model_from_config(config, f"{opt.ckpt}", device) + + if opt.plms: + sampler = PLMSSampler(model, device=device) + elif opt.dpm: + sampler = DPMSolverSampler(model, device=device) + else: + sampler = DDIMSampler(model, device=device) + + os.makedirs(opt.outdir, exist_ok=True) + outpath = opt.outdir + + print("Creating invisible watermark encoder (see https://github.com/ShieldMnt/invisible-watermark)...") + wm = "SDV2" + wm_encoder = WatermarkEncoder() + wm_encoder.set_watermark('bytes', wm.encode('utf-8')) + + batch_size = opt.n_samples + n_rows = opt.n_rows if opt.n_rows > 0 else batch_size + if not opt.from_file: + prompt = opt.prompt + assert prompt is not None + data = [batch_size * [prompt]] + + else: + print(f"reading prompts from {opt.from_file}") + with open(opt.from_file, "r") as f: + data = f.read().splitlines() + data = [p for p in data for i in range(opt.repeat)] + data = list(chunk(data, batch_size)) + + sample_path = os.path.join(outpath, "samples") + os.makedirs(sample_path, exist_ok=True) + sample_count = 0 + base_count = len(os.listdir(sample_path)) + grid_count = len(os.listdir(outpath)) - 1 + + start_code = None + if opt.fixed_code: + start_code = torch.randn([opt.n_samples, opt.C, opt.H // opt.f, opt.W // opt.f], device=device) + + if opt.torchscript or opt.ipex: + transformer = model.cond_stage_model.model + unet = model.model.diffusion_model + decoder = model.first_stage_model.decoder + additional_context = torch.cpu.amp.autocast() if opt.bf16 else nullcontext() + shape = [opt.C, opt.H // opt.f, opt.W // opt.f] + + if opt.bf16 and not opt.torchscript and not opt.ipex: + raise ValueError('Bfloat16 is supported only for torchscript+ipex') + if opt.bf16 and unet.dtype != torch.bfloat16: + raise ValueError("Use configs/stable-diffusion/intel/ configs with bf16 enabled if " + + "you'd like to use bfloat16 with CPU.") + if unet.dtype == torch.float16 and device == torch.device("cpu"): + raise ValueError("Use configs/stable-diffusion/intel/ configs for your model if you'd like to run it on CPU.") + + if opt.ipex: + import intel_extension_for_pytorch as ipex + bf16_dtype = torch.bfloat16 if opt.bf16 else None + transformer = transformer.to(memory_format=torch.channels_last) + transformer = ipex.optimize(transformer, level="O1", inplace=True) + + unet = unet.to(memory_format=torch.channels_last) + unet = ipex.optimize(unet, level="O1", auto_kernel_selection=True, inplace=True, dtype=bf16_dtype) + + decoder = decoder.to(memory_format=torch.channels_last) + decoder = ipex.optimize(decoder, level="O1", auto_kernel_selection=True, inplace=True, dtype=bf16_dtype) + + if opt.torchscript: + with torch.no_grad(), additional_context: + # get UNET scripted + if unet.use_checkpoint: + raise ValueError("Gradient checkpoint won't work with tracing. " + + "Use configs/stable-diffusion/intel/ configs for your model or disable checkpoint in your config.") + + img_in = torch.ones(2, 4, 96, 96, dtype=torch.float32) + t_in = torch.ones(2, dtype=torch.int64) + context = torch.ones(2, 77, 1024, dtype=torch.float32) + scripted_unet = torch.jit.trace(unet, (img_in, t_in, context)) + scripted_unet = torch.jit.optimize_for_inference(scripted_unet) + print(type(scripted_unet)) + model.model.scripted_diffusion_model = scripted_unet + + # get Decoder for first stage model scripted + samples_ddim = torch.ones(1, 4, 96, 96, dtype=torch.float32) + scripted_decoder = torch.jit.trace(decoder, (samples_ddim)) + scripted_decoder = torch.jit.optimize_for_inference(scripted_decoder) + print(type(scripted_decoder)) + model.first_stage_model.decoder = scripted_decoder + + prompts = data[0] + print("Running a forward pass to initialize optimizations") + uc = None + if opt.scale != 1.0: + uc = model.get_learned_conditioning(batch_size * [""]) + if isinstance(prompts, tuple): + prompts = list(prompts) + + with torch.no_grad(), additional_context: + for _ in range(3): + c = model.get_learned_conditioning(prompts) + samples_ddim, _ = sampler.sample(S=5, + conditioning=c, + batch_size=batch_size, + shape=shape, + verbose=False, + unconditional_guidance_scale=opt.scale, + unconditional_conditioning=uc, + eta=opt.ddim_eta, + x_T=start_code) + print("Running a forward pass for decoder") + for _ in range(3): + x_samples_ddim = model.decode_first_stage(samples_ddim) + + precision_scope = autocast if opt.precision=="autocast" or opt.bf16 else nullcontext + with torch.no_grad(), \ + precision_scope(opt.device), \ + model.ema_scope(): + all_samples = list() + for n in trange(opt.n_iter, desc="Sampling"): + for prompts in tqdm(data, desc="data"): + uc = None + if opt.scale != 1.0: + uc = model.get_learned_conditioning(batch_size * [""]) + if isinstance(prompts, tuple): + prompts = list(prompts) + c = model.get_learned_conditioning(prompts) + shape = [opt.C, opt.H // opt.f, opt.W // opt.f] + samples, _ = sampler.sample(S=opt.steps, + conditioning=c, + batch_size=opt.n_samples, + shape=shape, + verbose=False, + unconditional_guidance_scale=opt.scale, + unconditional_conditioning=uc, + eta=opt.ddim_eta, + x_T=start_code) + + x_samples = model.decode_first_stage(samples) + x_samples = torch.clamp((x_samples + 1.0) / 2.0, min=0.0, max=1.0) + + for x_sample in x_samples: + x_sample = 255. * rearrange(x_sample.cpu().numpy(), 'c h w -> h w c') + img = Image.fromarray(x_sample.astype(np.uint8)) + img = put_watermark(img, wm_encoder) + img.save(os.path.join(sample_path, f"{base_count:05}.png")) + base_count += 1 + sample_count += 1 + + all_samples.append(x_samples) + + # additionally, save as grid + grid = torch.stack(all_samples, 0) + grid = rearrange(grid, 'n b c h w -> (n b) c h w') + grid = make_grid(grid, nrow=n_rows) + + # to image + grid = 255. * rearrange(grid, 'c h w -> h w c').cpu().numpy() + grid = Image.fromarray(grid.astype(np.uint8)) + grid = put_watermark(grid, wm_encoder) + grid.save(os.path.join(outpath, f'grid-{grid_count:04}.png')) + grid_count += 1 + + print(f"Your samples are ready and waiting for you here: \n{outpath} \n" + f" \nEnjoy.") + + +if __name__ == "__main__": + opt = parse_args() + main(opt) diff --git a/setup.py b/setup.py new file mode 100644 index 0000000000000000000000000000000000000000..00f5b4d874f0f19ece54fac2dd50b39774b86c5b --- /dev/null +++ b/setup.py @@ -0,0 +1,13 @@ +from setuptools import setup, find_packages + +setup( + name='stable-diffusion', + version='0.0.1', + description='', + packages=find_packages(), + install_requires=[ + 'torch', + 'numpy', + 'tqdm', + ], +) \ No newline at end of file