Spaces:
Running
on
Zero
Running
on
Zero
File size: 9,920 Bytes
87d40d2 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 |
<!--Copyright 2024 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.
-->
# InstructPix2Pix
[InstructPix2Pix](https://arxiv.org/abs/2211.09800)λ text-conditioned diffusion λͺ¨λΈμ΄ ν μ΄λ―Έμ§μ νΈμ§μ λ°λ₯Ό μ μλλ‘ νμΈνλνλ λ°©λ²μ
λλ€. μ΄ λ°©λ²μ μ¬μ©νμ¬ νμΈνλλ λͺ¨λΈμ λ€μμ μ
λ ₯μΌλ‘ μ¬μ©ν©λλ€:
<p align="center">
<img src="https://huggingface.co/datasets/diffusers/docs-images/resolve/main/evaluation_diffusion_models/edit-instruction.png" alt="instructpix2pix-inputs" width=600/>
</p>
μΆλ ₯μ μ
λ ₯ μ΄λ―Έμ§μ νΈμ§ μ§μκ° λ°μλ "μμ λ" μ΄λ―Έμ§μ
λλ€:
<p align="center">
<img src="https://huggingface.co/datasets/diffusers/docs-images/resolve/main/output-gs%407-igs%401-steps%4050.png" alt="instructpix2pix-output" width=600/>
</p>
`train_instruct_pix2pix.py` μ€ν¬λ¦½νΈ([μ¬κΈ°](https://github.com/huggingface/diffusers/blob/main/examples/instruct_pix2pix/train_instruct_pix2pix.py)μμ μ°Ύμ μ μμ΅λλ€.)λ νμ΅ μ μ°¨λ₯Ό μ€λͺ
νκ³ Stable Diffusionμ μ μ©ν μ μλ λ°©λ²μ 보μ¬μ€λλ€.
*** `train_instruct_pix2pix.py`λ [μλ ꡬν](https://github.com/timothybrooks/instruct-pix2pix)μ μΆ©μ€νλ©΄μ InstructPix2Pix νμ΅ μ μ°¨λ₯Ό ꡬννκ³ μμ§λ§, [μκ·λͺ¨ λ°μ΄ν°μ
](https://huggingface.co/datasets/fusing/instructpix2pix-1000-samples)μμλ§ ν
μ€νΈλ₯Ό νμ΅λλ€. μ΄λ μ΅μ’
κ²°κ³Όμ μν₯μ λΌμΉ μ μμ΅λλ€. λ λμ κ²°κ³Όλ₯Ό μν΄, λ ν° λ°μ΄ν°μ
μμ λ κΈΈκ² νμ΅νλ κ²μ κΆμ₯ν©λλ€. [μ¬κΈ°](https://huggingface.co/datasets/timbrooks/instructpix2pix-clip-filtered)μμ InstructPix2Pix νμ΅μ μν΄ ν° λ°μ΄ν°μ
μ μ°Ύμ μ μμ΅λλ€.
***
## PyTorchλ‘ λ‘컬μμ μ€ννκΈ°
### μ’
μμ±(dependencies) μ€μΉνκΈ°
μ΄ μ€ν¬λ¦½νΈλ₯Ό μ€ννκΈ° μ μ, λΌμ΄λΈλ¬λ¦¬μ νμ΅ μ’
μμ±μ μ€μΉνμΈμ:
**μ€μ**
μ΅μ λ²μ μ μμ μ€ν¬λ¦½νΈλ₯Ό μ±κ³΅μ μΌλ‘ μ€ννκΈ° μν΄, **μλ³ΈμΌλ‘λΆν° μ€μΉ**νλ κ²κ³Ό μμ μ€ν¬λ¦½νΈλ₯Ό μμ£Ό μ
λ°μ΄νΈνκ³ μμ λ³ μꡬμ¬νμ μ€μΉνκΈ° λλ¬Έμ μ΅μ μνλ‘ μ μ§νλ κ²μ κΆμ₯ν©λλ€. μ΄λ₯Ό μν΄, μλ‘μ΄ κ°μ νκ²½μμ λ€μ μ€ν
μ μ€ννμΈμ:
```bash
git clone https://github.com/huggingface/diffusers
cd diffusers
pip install -e .
```
cd λͺ
λ Ήμ΄λ‘ μμ ν΄λλ‘ μ΄λνμΈμ.
```bash
cd examples/instruct_pix2pix
```
μ΄μ μ€ννμΈμ.
```bash
pip install -r requirements.txt
```
κ·Έλ¦¬κ³ [π€Accelerate](https://github.com/huggingface/accelerate/) νκ²½μμ μ΄κΈ°ννμΈμ:
```bash
accelerate config
```
νΉμ νκ²½μ λν μ§λ¬Έ μμ΄ κΈ°λ³Έμ μΈ accelerate ꡬμ±μ μ¬μ©νλ €λ©΄ λ€μμ μ€ννμΈμ.
```bash
accelerate config default
```
νΉμ μ¬μ© μ€μΈ νκ²½μ΄ notebookκ³Ό κ°μ λνν μμ μ§μνμ§ μλ κ²½μ°λ λ€μ μ μ°¨λ₯Ό λ°λΌμ£ΌμΈμ.
```python
from accelerate.utils import write_basic_config
write_basic_config()
```
### μμ
μ΄μ μ μΈκΈνλ―μ΄, νμ΅μ μν΄ [μμ λ°μ΄ν°μ
](https://huggingface.co/datasets/fusing/instructpix2pix-1000-samples)μ μ¬μ©ν κ²μ
λλ€. κ·Έ λ°μ΄ν°μ
μ InstructPix2Pix λ
Όλ¬Έμμ μ¬μ©λ [μλμ λ°μ΄ν°μ
](https://huggingface.co/datasets/timbrooks/instructpix2pix-clip-filtered)λ³΄λ€ μμ λ²μ μ
λλ€. μμ μ λ°μ΄ν°μ
μ μ¬μ©νκΈ° μν΄, [νμ΅μ μν λ°μ΄ν°μ
λ§λ€κΈ°](create_dataset) κ°μ΄λλ₯Ό μ°Έκ³ νμΈμ.
`MODEL_NAME` νκ²½ λ³μ(νλΈ λͺ¨λΈ λ ν¬μ§ν 리 λλ λͺ¨λΈ κ°μ€μΉκ° ν¬ν¨λ ν΄λ κ²½λ‘)λ₯Ό μ§μ νκ³ [`pretrained_model_name_or_path`](https://huggingface.co/docs/diffusers/en/api/diffusion_pipeline#diffusers.DiffusionPipeline.from_pretrained.pretrained_model_name_or_path) μΈμμ μ λ¬ν©λλ€. `DATASET_ID`μ λ°μ΄ν°μ
μ΄λ¦μ μ§μ ν΄μΌ ν©λλ€:
```bash
export MODEL_NAME="runwayml/stable-diffusion-v1-5"
export DATASET_ID="fusing/instructpix2pix-1000-samples"
```
μ§κΈ, νμ΅μ μ€νν μ μμ΅λλ€. μ€ν¬λ¦½νΈλ λ ν¬μ§ν 리μ νμ ν΄λμ λͺ¨λ ꡬμ±μμ(`feature_extractor`, `scheduler`, `text_encoder`, `unet` λ±)λ₯Ό μ μ₯ν©λλ€.
```bash
accelerate launch --mixed_precision="fp16" train_instruct_pix2pix.py \
--pretrained_model_name_or_path=$MODEL_NAME \
--dataset_name=$DATASET_ID \
--enable_xformers_memory_efficient_attention \
--resolution=256 --random_flip \
--train_batch_size=4 --gradient_accumulation_steps=4 --gradient_checkpointing \
--max_train_steps=15000 \
--checkpointing_steps=5000 --checkpoints_total_limit=1 \
--learning_rate=5e-05 --max_grad_norm=1 --lr_warmup_steps=0 \
--conditioning_dropout_prob=0.05 \
--mixed_precision=fp16 \
--seed=42 \
--push_to_hub
```
μΆκ°μ μΌλ‘, κ°μ€μΉμ λ°μ΄μ΄μ€λ₯Ό νμ΅ κ³Όμ μ λͺ¨λν°λ§νμ¬ κ²μ¦ μΆλ‘ μ μννλ κ²μ μ§μν©λλ€. `report_to="wandb"`μ μ΄ κΈ°λ₯μ μ¬μ©ν μ μμ΅λλ€:
```bash
accelerate launch --mixed_precision="fp16" train_instruct_pix2pix.py \
--pretrained_model_name_or_path=$MODEL_NAME \
--dataset_name=$DATASET_ID \
--enable_xformers_memory_efficient_attention \
--resolution=256 --random_flip \
--train_batch_size=4 --gradient_accumulation_steps=4 --gradient_checkpointing \
--max_train_steps=15000 \
--checkpointing_steps=5000 --checkpoints_total_limit=1 \
--learning_rate=5e-05 --max_grad_norm=1 --lr_warmup_steps=0 \
--conditioning_dropout_prob=0.05 \
--mixed_precision=fp16 \
--val_image_url="https://hf.co/datasets/diffusers/diffusers-images-docs/resolve/main/mountain.png" \
--validation_prompt="make the mountains snowy" \
--seed=42 \
--report_to=wandb \
--push_to_hub
```
λͺ¨λΈ λλ²κΉ
μ μ μ©ν μ΄ νκ° λ°©λ² κΆμ₯ν©λλ€. μ΄λ₯Ό μ¬μ©νκΈ° μν΄ `wandb`λ₯Ό μ€μΉνλ κ²μ μ£Όλͺ©ν΄μ£ΌμΈμ. `pip install wandb`λ‘ μ€νν΄ `wandb`λ₯Ό μ€μΉν μ μμ΅λλ€.
[μ¬κΈ°](https://wandb.ai/sayakpaul/instruct-pix2pix/runs/ctr3kovq), λͺ κ°μ§ νκ° λ°©λ²κ³Ό νμ΅ νλΌλ―Έν°λ₯Ό ν¬ν¨νλ μμλ₯Ό λ³Ό μ μμ΅λλ€.
***μ°Έκ³ : μλ³Έ λ
Όλ¬Έμμ, μ μλ€μ 256x256 μ΄λ―Έμ§ ν΄μλλ‘ νμ΅ν λͺ¨λΈλ‘ 512x512μ κ°μ λ ν° ν΄μλλ‘ μ μΌλ°νλλ κ²μ λ³Ό μ μμμ΅λλ€. μ΄λ νμ΅μ μ¬μ©ν ν° λ°μ΄ν°μ
μ μ¬μ©νκΈ° λλ¬Έμ
λλ€.***
## λ€μμ GPUλ‘ νμ΅νκΈ°
`accelerate`λ μνν λ€μμ GPUλ‘ νμ΅μ κ°λ₯νκ² ν©λλ€. `accelerate`λ‘ λΆμ° νμ΅μ μ€ννλ [μ¬κΈ°](https://huggingface.co/docs/accelerate/basic_tutorials/launch) μ€λͺ
μ λ°λΌ ν΄ μ£ΌμκΈ° λ°λλλ€. μμμ λͺ
λ Ήμ΄ μ
λλ€:
```bash
accelerate launch --mixed_precision="fp16" --multi_gpu train_instruct_pix2pix.py \
--pretrained_model_name_or_path=runwayml/stable-diffusion-v1-5 \
--dataset_name=sayakpaul/instructpix2pix-1000-samples \
--use_ema \
--enable_xformers_memory_efficient_attention \
--resolution=512 --random_flip \
--train_batch_size=4 --gradient_accumulation_steps=4 --gradient_checkpointing \
--max_train_steps=15000 \
--checkpointing_steps=5000 --checkpoints_total_limit=1 \
--learning_rate=5e-05 --lr_warmup_steps=0 \
--conditioning_dropout_prob=0.05 \
--mixed_precision=fp16 \
--seed=42 \
--push_to_hub
```
## μΆλ‘ νκΈ°
μΌλ¨ νμ΅μ΄ μλ£λλ©΄, μΆλ‘ ν μ μμ΅λλ€:
```python
import PIL
import requests
import torch
from diffusers import StableDiffusionInstructPix2PixPipeline
model_id = "your_model_id" # <- μ΄λ₯Ό μμ νμΈμ.
pipe = StableDiffusionInstructPix2PixPipeline.from_pretrained(model_id, torch_dtype=torch.float16).to("cuda")
generator = torch.Generator("cuda").manual_seed(0)
url = "https://huggingface.co/datasets/sayakpaul/sample-datasets/resolve/main/test_pix2pix_4.png"
def download_image(url):
image = PIL.Image.open(requests.get(url, stream=True).raw)
image = PIL.ImageOps.exif_transpose(image)
image = image.convert("RGB")
return image
image = download_image(url)
prompt = "wipe out the lake"
num_inference_steps = 20
image_guidance_scale = 1.5
guidance_scale = 10
edited_image = pipe(
prompt,
image=image,
num_inference_steps=num_inference_steps,
image_guidance_scale=image_guidance_scale,
guidance_scale=guidance_scale,
generator=generator,
).images[0]
edited_image.save("edited_image.png")
```
νμ΅ μ€ν¬λ¦½νΈλ₯Ό μ¬μ©ν΄ μ»μ μμμ λͺ¨λΈ λ ν¬μ§ν 리λ μ¬κΈ° [sayakpaul/instruct-pix2pix](https://huggingface.co/sayakpaul/instruct-pix2pix)μμ νμΈν μ μμ΅λλ€.
μ±λ₯μ μν μλμ νμ§μ μ μ΄νκΈ° μν΄ μΈ κ°μ§ νλΌλ―Έν°λ₯Ό μ¬μ©νλ κ²μ΄ μ’μ΅λλ€:
* `num_inference_steps`
* `image_guidance_scale`
* `guidance_scale`
νΉν, `image_guidance_scale`μ `guidance_scale`λ μμ±λ("μμ λ") μ΄λ―Έμ§μμ ν° μν₯μ λ―ΈμΉ μ μμ΅λλ€.([μ¬κΈ°](https://twitter.com/RisingSayak/status/1628392199196151808?s=20)μμλ₯Ό μ°Έκ³ ν΄μ£ΌμΈμ.)
λ§μ½ InstructPix2Pix νμ΅ λ°©λ²μ μ¬μ©ν΄ λͺ κ°μ§ ν₯λ―Έλ‘μ΄ λ°©λ²μ μ°Ύκ³ μλ€λ©΄, μ΄ λΈλ‘κ·Έ κ²μλ¬Ό[Instruction-tuning Stable Diffusion with InstructPix2Pix](https://huggingface.co/blog/instruction-tuning-sd)μ νμΈν΄μ£ΌμΈμ. |