diff --git a/.gitattributes b/.gitattributes index a6344aac8c09253b3b630fb776ae94478aa0275b..468e5d1770dc2f6d1504c62dc12d5d7014d47a05 100644 --- a/.gitattributes +++ b/.gitattributes @@ -33,3 +33,6 @@ 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 +BigVGAN/demo/examples/megalovania_24k.wav filter=lfs diff=lfs merge=lfs -text +BigVGAN/filelists/LibriTTS/train-full.txt filter=lfs diff=lfs merge=lfs -text +long_form_sample.ogg filter=lfs diff=lfs merge=lfs -text diff --git a/.gitignore b/.gitignore new file mode 100644 index 0000000000000000000000000000000000000000..f9f8fe3f7b4551ad42f5bacc3491221f56680ca3 --- /dev/null +++ b/.gitignore @@ -0,0 +1,165 @@ +*.pt +test_output.wav + +# Byte-compiled / optimized / DLL files +__pycache__/ +*.py[cod] +*$py.class + +# C extensions +*.so + +# Distribution / packaging +.Python +build/ +develop-eggs/ +dist/ +downloads/ +eggs/ +.eggs/ +lib/ +lib64/ +parts/ +sdist/ +var/ +wheels/ +share/python-wheels/ +*.egg-info/ +.installed.cfg +*.egg +MANIFEST + +# PyInstaller +# Usually these files are written by a python script from a template +# before PyInstaller builds the exe, so as to inject date/other infos into it. +*.manifest +*.spec + +# Installer logs +pip-log.txt +pip-delete-this-directory.txt + +# Unit test / coverage reports +htmlcov/ +.tox/ +.nox/ +.coverage +.coverage.* +.cache +nosetests.xml +coverage.xml +*.cover +*.py,cover +.hypothesis/ +.pytest_cache/ +cover/ + +# Translations +*.mo +*.pot + +# Django stuff: +*.log +local_settings.py +db.sqlite3 +db.sqlite3-journal + +# Flask stuff: +instance/ +.webassets-cache + +# Scrapy stuff: +.scrapy + +# Sphinx documentation +docs/_build/ + +# PyBuilder +.pybuilder/ +target/ + +# Jupyter Notebook +.ipynb_checkpoints + +# IPython +profile_default/ +ipython_config.py + +# pyenv +# For a library or package, you might want to ignore these files since the code is +# intended to run in multiple environments; otherwise, check them in: +# .python-version + +# pipenv +# According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control. +# However, in case of collaboration, if having platform-specific dependencies or dependencies +# having no cross-platform support, pipenv may install dependencies that don't work, or not +# install all needed dependencies. +#Pipfile.lock + +# poetry +# Similar to Pipfile.lock, it is generally recommended to include poetry.lock in version control. +# This is especially recommended for binary packages to ensure reproducibility, and is more +# commonly ignored for libraries. +# https://python-poetry.org/docs/basic-usage/#commit-your-poetrylock-file-to-version-control +#poetry.lock + +# pdm +# Similar to Pipfile.lock, it is generally recommended to include pdm.lock in version control. +#pdm.lock +# pdm stores project-wide configurations in .pdm.toml, but it is recommended to not include it +# in version control. +# https://pdm.fming.dev/latest/usage/project/#working-with-version-control +.pdm.toml +.pdm-python +.pdm-build/ + +# PEP 582; used by e.g. github.com/David-OConnor/pyflow and github.com/pdm-project/pdm +__pypackages__/ + +# Celery stuff +celerybeat-schedule +celerybeat.pid + +# SageMath parsed files +*.sage.py + +# Environments +.env +.venv +env/ +venv/ +ENV/ +env.bak/ +venv.bak/ + +# Spyder project settings +.spyderproject +.spyproject + +# Rope project settings +.ropeproject + +# mkdocs documentation +/site + +# mypy +.mypy_cache/ +.dmypy.json +dmypy.json + +# Pyre type checker +.pyre/ + +# pytype static type analyzer +.pytype/ + +# Cython debug symbols +cython_debug/ + +# PyCharm +# JetBrains specific template is maintained in a separate JetBrains.gitignore that can +# be found at https://github.com/github/gitignore/blob/main/Global/JetBrains.gitignore +# and can be added to the global gitignore or merged into this file. For a more nuclear +# option (not recommended) you can uncomment the following to ignore the entire idea folder. +#.idea/ diff --git a/.gitmodules b/.gitmodules new file mode 100644 index 0000000000000000000000000000000000000000..96a049863110fa1ea232eaeb214faf88b91b8267 --- /dev/null +++ b/.gitmodules @@ -0,0 +1,3 @@ +[submodule "BigVGAN"] + path = BigVGAN + url = https://github.com/NVIDIA/BigVGAN diff --git a/BigVGAN/.gitignore b/BigVGAN/.gitignore new file mode 100644 index 0000000000000000000000000000000000000000..d8f399ed4e81b88b0b6d61d2e99098f36503e80c --- /dev/null +++ b/BigVGAN/.gitignore @@ -0,0 +1,146 @@ +# BigVGAN +alias_free_activation/cuda/build/ +exp/ +tmp/ + +# Symlinks for bundled LibriTTS filelists +filelists/LibriTTS/train-clean-100 +filelists/LibriTTS/train-clean-360 +filelists/LibriTTS/train-other-500 +filelists/LibriTTS/dev-clean +filelists/LibriTTS/dev-other +filelists/LibriTTS/test-clean +filelists/LibriTTS/test-other + +# VSCode configs +.vscode/ + +# Byte-compiled / optimized / DLL files +__pycache__/ +*.py[cod] +*$py.class + +# C extensions +*.so + +# Distribution / packaging +.Python +build/ +develop-eggs/ +dist/ +downloads/ +eggs/ +.eggs/ +lib/ +lib64/ +parts/ +sdist/ +var/ +wheels/ +share/python-wheels/ +*.egg-info/ +.installed.cfg +*.egg +MANIFEST + +# PyInstaller +# Usually these files are written by a python script from a template +# before PyInstaller builds the exe, so as to inject date/other infos into it. +*.manifest +*.spec + +# Installer logs +pip-log.txt +pip-delete-this-directory.txt + +# Unit test / coverage reports +htmlcov/ +.tox/ +.nox/ +.coverage +.coverage.* +.cache +nosetests.xml +coverage.xml +*.cover +*.py,cover +.hypothesis/ +.pytest_cache/ +cover/ + +# Translations +*.mo +*.pot + +# Django stuff: +*.log +local_settings.py +db.sqlite3 +db.sqlite3-journal + +# Flask stuff: +instance/ +.webassets-cache + +# Scrapy stuff: +.scrapy + +# Sphinx documentation +docs/_build/ + +# PyBuilder +.pybuilder/ +target/ + +# Jupyter Notebook +.ipynb_checkpoints + +# IPython +profile_default/ +ipython_config.py + +# PEP 582; used by e.g. github.com/David-OConnor/pyflow and github.com/pdm-project/pdm +__pypackages__/ + +# Celery stuff +celerybeat-schedule +celerybeat.pid + +# SageMath parsed files +*.sage.py + +# Environments +.env +.venv +env/ +venv/ +ENV/ +env.bak/ +venv.bak/ + +# Spyder project settings +.spyderproject +.spyproject + +# Rope project settings +.ropeproject + +# mkdocs documentation +/site + +# mypy +.mypy_cache/ +.dmypy.json +dmypy.json + +# Pyre type checker +.pyre/ + +# pytype static type analyzer +.pytype/ + +# Cython debug symbols +cython_debug/ + +# PyCharm +.idea/ \ No newline at end of file diff --git a/BigVGAN/LICENSE b/BigVGAN/LICENSE new file mode 100644 index 0000000000000000000000000000000000000000..4c78361c86d4f685117d60d6623e2197fcfed706 --- /dev/null +++ b/BigVGAN/LICENSE @@ -0,0 +1,21 @@ +MIT License + +Copyright (c) 2024 NVIDIA CORPORATION. + +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/BigVGAN/README.md b/BigVGAN/README.md new file mode 100644 index 0000000000000000000000000000000000000000..2fa70ceea647053933b913b329041ee8c41526db --- /dev/null +++ b/BigVGAN/README.md @@ -0,0 +1,266 @@ +## BigVGAN: A Universal Neural Vocoder with Large-Scale Training + +#### Sang-gil Lee, Wei Ping, Boris Ginsburg, Bryan Catanzaro, Sungroh Yoon + +[[Paper]](https://arxiv.org/abs/2206.04658) - [[Code]](https://github.com/NVIDIA/BigVGAN) - [[Showcase]](https://bigvgan-demo.github.io/) - [[Project Page]](https://research.nvidia.com/labs/adlr/projects/bigvgan/) - [[Weights]](https://huggingface.co/collections/nvidia/bigvgan-66959df3d97fd7d98d97dc9a) - [[Demo]](https://huggingface.co/spaces/nvidia/BigVGAN) + +[![PWC](https://img.shields.io/endpoint.svg?url=https://paperswithcode.com/badge/bigvgan-a-universal-neural-vocoder-with-large/speech-synthesis-on-libritts)](https://paperswithcode.com/sota/speech-synthesis-on-libritts?p=bigvgan-a-universal-neural-vocoder-with-large) + +
+ +## News +- **Sep 2024 (v2.4):** + - We have updated the pretrained checkpoints trained for 5M steps. This is final release of the BigVGAN-v2 checkpoints. + +- **Jul 2024 (v2.3):** + - General refactor and code improvements for improved readability. + - Fully fused CUDA kernel of anti-alised activation (upsampling + activation + downsampling) with inference speed benchmark. + +- **Jul 2024 (v2.2):** The repository now includes an interactive local demo using gradio. + +- **Jul 2024 (v2.1):** BigVGAN is now integrated with 🤗 Hugging Face Hub with easy access to inference using pretrained checkpoints. We also provide an interactive demo on Hugging Face Spaces. + +- **Jul 2024 (v2):** We release BigVGAN-v2 along with pretrained checkpoints. Below are the highlights: + - Custom CUDA kernel for inference: we provide a fused upsampling + activation kernel written in CUDA for accelerated inference speed. Our test shows 1.5 - 3x faster speed on a single A100 GPU. + - Improved discriminator and loss: BigVGAN-v2 is trained using a [multi-scale sub-band CQT discriminator](https://arxiv.org/abs/2311.14957) and a [multi-scale mel spectrogram loss](https://arxiv.org/abs/2306.06546). + - Larger training data: BigVGAN-v2 is trained using datasets containing diverse audio types, including speech in multiple languages, environmental sounds, and instruments. + - We provide pretrained checkpoints of BigVGAN-v2 using diverse audio configurations, supporting up to 44 kHz sampling rate and 512x upsampling ratio. + +## Installation + +The codebase has been tested on Python `3.10` and PyTorch `2.3.1` conda packages with either `pytorch-cuda=12.1` or `pytorch-cuda=11.8`. Below is an example command to create the conda environment: + +```shell +conda create -n bigvgan python=3.10 pytorch torchvision torchaudio pytorch-cuda=12.1 -c pytorch -c nvidia +conda activate bigvgan +``` + +Clone the repository and install dependencies: + +```shell +git clone https://github.com/NVIDIA/BigVGAN +cd BigVGAN +pip install -r requirements.txt +``` + +## Inference Quickstart using 🤗 Hugging Face Hub + +Below example describes how you can use BigVGAN: load the pretrained BigVGAN generator from Hugging Face Hub, compute mel spectrogram from input waveform, and generate synthesized waveform using the mel spectrogram as the model's input. + +```python +device = 'cuda' + +import torch +import bigvgan +import librosa +from meldataset import get_mel_spectrogram + +# instantiate the model. You can optionally set use_cuda_kernel=True for faster inference. +model = bigvgan.BigVGAN.from_pretrained('nvidia/bigvgan_v2_24khz_100band_256x', use_cuda_kernel=False) + +# remove weight norm in the model and set to eval mode +model.remove_weight_norm() +model = model.eval().to(device) + +# load wav file and compute mel spectrogram +wav_path = '/path/to/your/audio.wav' +wav, sr = librosa.load(wav_path, sr=model.h.sampling_rate, mono=True) # wav is np.ndarray with shape [T_time] and values in [-1, 1] +wav = torch.FloatTensor(wav).unsqueeze(0) # wav is FloatTensor with shape [B(1), T_time] + +# compute mel spectrogram from the ground truth audio +mel = get_mel_spectrogram(wav, model.h).to(device) # mel is FloatTensor with shape [B(1), C_mel, T_frame] + +# generate waveform from mel +with torch.inference_mode(): + wav_gen = model(mel) # wav_gen is FloatTensor with shape [B(1), 1, T_time] and values in [-1, 1] +wav_gen_float = wav_gen.squeeze(0).cpu() # wav_gen is FloatTensor with shape [1, T_time] + +# you can convert the generated waveform to 16 bit linear PCM +wav_gen_int16 = (wav_gen_float * 32767.0).numpy().astype('int16') # wav_gen is now np.ndarray with shape [1, T_time] and int16 dtype +``` + +## Local gradio demo + +You can run a local gradio demo using below command: + +```python +pip install -r demo/requirements.txt +python demo/app.py +``` + +## Training + +Create symbolic link to the root of the dataset. The codebase uses filelist with the relative path from the dataset. Below are the example commands for LibriTTS dataset: + +```shell +cd filelists/LibriTTS && \ +ln -s /path/to/your/LibriTTS/train-clean-100 train-clean-100 && \ +ln -s /path/to/your/LibriTTS/train-clean-360 train-clean-360 && \ +ln -s /path/to/your/LibriTTS/train-other-500 train-other-500 && \ +ln -s /path/to/your/LibriTTS/dev-clean dev-clean && \ +ln -s /path/to/your/LibriTTS/dev-other dev-other && \ +ln -s /path/to/your/LibriTTS/test-clean test-clean && \ +ln -s /path/to/your/LibriTTS/test-other test-other && \ +cd ../.. +``` + +Train BigVGAN model. Below is an example command for training BigVGAN-v2 using LibriTTS dataset at 24kHz with a full 100-band mel spectrogram as input: + +```shell +python train.py \ +--config configs/bigvgan_v2_24khz_100band_256x.json \ +--input_wavs_dir filelists/LibriTTS \ +--input_training_file filelists/LibriTTS/train-full.txt \ +--input_validation_file filelists/LibriTTS/val-full.txt \ +--list_input_unseen_wavs_dir filelists/LibriTTS filelists/LibriTTS \ +--list_input_unseen_validation_file filelists/LibriTTS/dev-clean.txt filelists/LibriTTS/dev-other.txt \ +--checkpoint_path exp/bigvgan_v2_24khz_100band_256x +``` + +## Synthesis + +Synthesize from BigVGAN model. Below is an example command for generating audio from the model. +It computes mel spectrograms using wav files from `--input_wavs_dir` and saves the generated audio to `--output_dir`. + +```shell +python inference.py \ +--checkpoint_file /path/to/your/bigvgan_v2_24khz_100band_256x/bigvgan_generator.pt \ +--input_wavs_dir /path/to/your/input_wav \ +--output_dir /path/to/your/output_wav +``` + +`inference_e2e.py` supports synthesis directly from the mel spectrogram saved in `.npy` format, with shapes `[1, channel, frame]` or `[channel, frame]`. +It loads mel spectrograms from `--input_mels_dir` and saves the generated audio to `--output_dir`. + +Make sure that the STFT hyperparameters for mel spectrogram are the same as the model, which are defined in `config.json` of the corresponding model. + +```shell +python inference_e2e.py \ +--checkpoint_file /path/to/your/bigvgan_v2_24khz_100band_256x/bigvgan_generator.pt \ +--input_mels_dir /path/to/your/input_mel \ +--output_dir /path/to/your/output_wav +``` + +## Using Custom CUDA Kernel for Synthesis + +You can apply the fast CUDA inference kernel by using a parameter `use_cuda_kernel` when instantiating BigVGAN: + +```python +generator = BigVGAN(h, use_cuda_kernel=True) +``` + +You can also pass `--use_cuda_kernel` to `inference.py` and `inference_e2e.py` to enable this feature. + +When applied for the first time, it builds the kernel using `nvcc` and `ninja`. If the build succeeds, the kernel is saved to `alias_free_activation/cuda/build` and the model automatically loads the kernel. The codebase has been tested using CUDA `12.1`. + +Please make sure that both are installed in your system and `nvcc` installed in your system matches the version your PyTorch build is using. + +We recommend running `test_cuda_vs_torch_model.py` first to build and check the correctness of the CUDA kernel. See below example command and its output, where it returns `[Success] test CUDA fused vs. plain torch BigVGAN inference`: + +```python +python tests/test_cuda_vs_torch_model.py \ +--checkpoint_file /path/to/your/bigvgan_generator.pt +``` + +```shell +loading plain Pytorch BigVGAN +... +loading CUDA kernel BigVGAN with auto-build +Detected CUDA files, patching ldflags +Emitting ninja build file /path/to/your/BigVGAN/alias_free_activation/cuda/build/build.ninja.. +Building extension module anti_alias_activation_cuda... +... +Loading extension module anti_alias_activation_cuda... +... +Loading '/path/to/your/bigvgan_generator.pt' +... +[Success] test CUDA fused vs. plain torch BigVGAN inference + > mean_difference=0.0007238413265440613 +... +``` + +If you see `[Fail] test CUDA fused vs. plain torch BigVGAN inference`, it means that the CUDA kernel inference is incorrect. Please check if `nvcc` installed in your system is compatible with your PyTorch version. + +## Pretrained Models + +We provide the [pretrained models on Hugging Face Collections](https://huggingface.co/collections/nvidia/bigvgan-66959df3d97fd7d98d97dc9a). +One can download the checkpoints of the generator weight (named `bigvgan_generator.pt`) and its discriminator/optimizer states (named `bigvgan_discriminator_optimizer.pt`) within the listed model repositories. + +| Model Name | Sampling Rate | Mel band | fmax | Upsampling Ratio | Params | Dataset | Steps | Fine-Tuned | +|:--------------------------------------------------------------------------------------------------------:|:-------------:|:--------:|:-----:|:----------------:|:------:|:--------------------------:|:-----:|:----------:| +| [bigvgan_v2_44khz_128band_512x](https://huggingface.co/nvidia/bigvgan_v2_44khz_128band_512x) | 44 kHz | 128 | 22050 | 512 | 122M | Large-scale Compilation | 5M | No | +| [bigvgan_v2_44khz_128band_256x](https://huggingface.co/nvidia/bigvgan_v2_44khz_128band_256x) | 44 kHz | 128 | 22050 | 256 | 112M | Large-scale Compilation | 5M | No | +| [bigvgan_v2_24khz_100band_256x](https://huggingface.co/nvidia/bigvgan_v2_24khz_100band_256x) | 24 kHz | 100 | 12000 | 256 | 112M | Large-scale Compilation | 5M | No | +| [bigvgan_v2_22khz_80band_256x](https://huggingface.co/nvidia/bigvgan_v2_22khz_80band_256x) | 22 kHz | 80 | 11025 | 256 | 112M | Large-scale Compilation | 5M | No | +| [bigvgan_v2_22khz_80band_fmax8k_256x](https://huggingface.co/nvidia/bigvgan_v2_22khz_80band_fmax8k_256x) | 22 kHz | 80 | 8000 | 256 | 112M | Large-scale Compilation | 5M | No | +| [bigvgan_24khz_100band](https://huggingface.co/nvidia/bigvgan_24khz_100band) | 24 kHz | 100 | 12000 | 256 | 112M | LibriTTS | 5M | No | +| [bigvgan_base_24khz_100band](https://huggingface.co/nvidia/bigvgan_base_24khz_100band) | 24 kHz | 100 | 12000 | 256 | 14M | LibriTTS | 5M | No | +| [bigvgan_22khz_80band](https://huggingface.co/nvidia/bigvgan_22khz_80band) | 22 kHz | 80 | 8000 | 256 | 112M | LibriTTS + VCTK + LJSpeech | 5M | No | +| [bigvgan_base_22khz_80band](https://huggingface.co/nvidia/bigvgan_base_22khz_80band) | 22 kHz | 80 | 8000 | 256 | 14M | LibriTTS + VCTK + LJSpeech | 5M | No | + +The paper results are based on the original 24kHz BigVGAN models (`bigvgan_24khz_100band` and `bigvgan_base_24khz_100band`) trained on LibriTTS dataset. +We also provide 22kHz BigVGAN models with band-limited setup (i.e., fmax=8000) for TTS applications. +Note that the checkpoints use `snakebeta` activation with log scale parameterization, which have the best overall quality. + +You can fine-tune the models by: + +1. downloading the checkpoints (both the generator weight and its discriminator/optimizer states) +2. resuming training using your audio dataset by specifying `--checkpoint_path` that includes the checkpoints when launching `train.py` + +## Training Details of BigVGAN-v2 + +Comapred to the original BigVGAN, the pretrained checkpoints of BigVGAN-v2 used `batch_size=32` with a longer `segment_size=65536` and are trained using 8 A100 GPUs. + +Note that the BigVGAN-v2 `json` config files in `./configs` use `batch_size=4` as default to fit in a single A100 GPU for training. You can fine-tune the models adjusting `batch_size` depending on your GPUs. + +When training BigVGAN-v2 from scratch with small batch size, it can potentially encounter the early divergence problem mentioned in the paper. In such case, we recommend lowering the `clip_grad_norm` value (e.g. `100`) for the early training iterations (e.g. 20000 steps) and increase the value to the default `500`. + +## Evaluation Results of BigVGAN-v2 + +Below are the objective results of the 24kHz model (`bigvgan_v2_24khz_100band_256x`) obtained from the LibriTTS `dev` sets. BigVGAN-v2 shows noticeable improvements of the metrics. The model also exhibits reduced perceptual artifacts, especially for non-speech audio. + +| Model | Dataset | Steps | PESQ(↑) | M-STFT(↓) | MCD(↓) | Periodicity(↓) | V/UV F1(↑) | +|:----------:|:-----------------------:|:-----:|:---------:|:----------:|:----------:|:--------------:|:----------:| +| BigVGAN | LibriTTS | 1M | 4.027 | 0.7997 | 0.3745 | 0.1018 | 0.9598 | +| BigVGAN | LibriTTS | 5M | 4.256 | 0.7409 | 0.2988 | 0.0809 | 0.9698 | +| BigVGAN-v2 | Large-scale Compilation | 3M | 4.359 | 0.7134 | 0.3060 | 0.0621 | 0.9777 | +| BigVGAN-v2 | Large-scale Compilation | 5M | **4.362** | **0.7026** | **0.2903** | **0.0593** | **0.9793** | + +## Speed Benchmark + +Below are the speed and VRAM usage benchmark results of BigVGAN from `tests/test_cuda_vs_torch_model.py`, using `bigvgan_v2_24khz_100band_256x` as a reference model. + +| GPU | num_mel_frame | use_cuda_kernel | Speed (kHz) | Real-time Factor | VRAM (GB) | +|:--------------------------:|:-------------:|:---------------:|:-----------:|:----------------:|:---------:| +| NVIDIA A100 | 256 | False | 1672.1 | 69.7x | 1.3 | +| | | True | 3916.5 | 163.2x | 1.3 | +| | 2048 | False | 1899.6 | 79.2x | 1.7 | +| | | True | 5330.1 | 222.1x | 1.7 | +| | 16384 | False | 1973.8 | 82.2x | 5.0 | +| | | True | 5761.7 | 240.1x | 4.4 | +| NVIDIA GeForce RTX 3080 | 256 | False | 841.1 | 35.0x | 1.3 | +| | | True | 1598.1 | 66.6x | 1.3 | +| | 2048 | False | 929.9 | 38.7x | 1.7 | +| | | True | 1971.3 | 82.1x | 1.6 | +| | 16384 | False | 943.4 | 39.3x | 5.0 | +| | | True | 2026.5 | 84.4x | 3.9 | +| NVIDIA GeForce RTX 2080 Ti | 256 | False | 515.6 | 21.5x | 1.3 | +| | | True | 811.3 | 33.8x | 1.3 | +| | 2048 | False | 576.5 | 24.0x | 1.7 | +| | | True | 1023.0 | 42.6x | 1.5 | +| | 16384 | False | 589.4 | 24.6x | 5.0 | +| | | True | 1068.1 | 44.5x | 3.2 | + +## Acknowledgements + +We thank Vijay Anand Korthikanti and Kevin J. Shih for their generous support in implementing the CUDA kernel for inference. + +## References + +- [HiFi-GAN](https://github.com/jik876/hifi-gan) (for generator and multi-period discriminator) +- [Snake](https://github.com/EdwardDixon/snake) (for periodic activation) +- [Alias-free-torch](https://github.com/junjun3518/alias-free-torch) (for anti-aliasing) +- [Julius](https://github.com/adefossez/julius) (for low-pass filter) +- [UnivNet](https://github.com/mindslab-ai/univnet) (for multi-resolution discriminator) +- [descript-audio-codec](https://github.com/descriptinc/descript-audio-codec) and [vocos](https://github.com/gemelo-ai/vocos) (for multi-band multi-scale STFT discriminator and multi-scale mel spectrogram loss) +- [Amphion](https://github.com/open-mmlab/Amphion) (for multi-scale sub-band CQT discriminator) diff --git a/BigVGAN/activations.py b/BigVGAN/activations.py new file mode 100644 index 0000000000000000000000000000000000000000..4f08ddab5b55d6dcaf3e968af98889e0770c44f5 --- /dev/null +++ b/BigVGAN/activations.py @@ -0,0 +1,126 @@ +# Implementation adapted from https://github.com/EdwardDixon/snake under the MIT license. +# LICENSE is in incl_licenses directory. + +import torch +from torch import nn, sin, pow +from torch.nn import Parameter + + +class Snake(nn.Module): + """ + Implementation of a sine-based periodic activation function + Shape: + - Input: (B, C, T) + - Output: (B, C, T), same shape as the input + Parameters: + - alpha - trainable parameter + References: + - This activation function is from this paper by Liu Ziyin, Tilman Hartwig, Masahito Ueda: + https://arxiv.org/abs/2006.08195 + Examples: + >>> a1 = snake(256) + >>> x = torch.randn(256) + >>> x = a1(x) + """ + + def __init__( + self, in_features, alpha=1.0, alpha_trainable=True, alpha_logscale=False + ): + """ + Initialization. + INPUT: + - in_features: shape of the input + - alpha: trainable parameter + alpha is initialized to 1 by default, higher values = higher-frequency. + alpha will be trained along with the rest of your model. + """ + super(Snake, self).__init__() + self.in_features = in_features + + # Initialize alpha + self.alpha_logscale = alpha_logscale + if self.alpha_logscale: # Log scale alphas initialized to zeros + self.alpha = Parameter(torch.zeros(in_features) * alpha) + else: # Linear scale alphas initialized to ones + self.alpha = Parameter(torch.ones(in_features) * alpha) + + self.alpha.requires_grad = alpha_trainable + + self.no_div_by_zero = 0.000000001 + + def forward(self, x): + """ + Forward pass of the function. + Applies the function to the input elementwise. + Snake ∶= x + 1/a * sin^2 (xa) + """ + alpha = self.alpha.unsqueeze(0).unsqueeze(-1) # Line up with x to [B, C, T] + if self.alpha_logscale: + alpha = torch.exp(alpha) + x = x + (1.0 / (alpha + self.no_div_by_zero)) * pow(sin(x * alpha), 2) + + return x + + +class SnakeBeta(nn.Module): + """ + A modified Snake function which uses separate parameters for the magnitude of the periodic components + Shape: + - Input: (B, C, T) + - Output: (B, C, T), same shape as the input + Parameters: + - alpha - trainable parameter that controls frequency + - beta - trainable parameter that controls magnitude + References: + - This activation function is a modified version based on this paper by Liu Ziyin, Tilman Hartwig, Masahito Ueda: + https://arxiv.org/abs/2006.08195 + Examples: + >>> a1 = snakebeta(256) + >>> x = torch.randn(256) + >>> x = a1(x) + """ + + def __init__( + self, in_features, alpha=1.0, alpha_trainable=True, alpha_logscale=False + ): + """ + Initialization. + INPUT: + - in_features: shape of the input + - alpha - trainable parameter that controls frequency + - beta - trainable parameter that controls magnitude + alpha is initialized to 1 by default, higher values = higher-frequency. + beta is initialized to 1 by default, higher values = higher-magnitude. + alpha will be trained along with the rest of your model. + """ + super(SnakeBeta, self).__init__() + self.in_features = in_features + + # Initialize alpha + self.alpha_logscale = alpha_logscale + if self.alpha_logscale: # Log scale alphas initialized to zeros + self.alpha = Parameter(torch.zeros(in_features) * alpha) + self.beta = Parameter(torch.zeros(in_features) * alpha) + else: # Linear scale alphas initialized to ones + self.alpha = Parameter(torch.ones(in_features) * alpha) + self.beta = Parameter(torch.ones(in_features) * alpha) + + self.alpha.requires_grad = alpha_trainable + self.beta.requires_grad = alpha_trainable + + self.no_div_by_zero = 0.000000001 + + def forward(self, x): + """ + Forward pass of the function. + Applies the function to the input elementwise. + SnakeBeta ∶= x + 1/b * sin^2 (xa) + """ + alpha = self.alpha.unsqueeze(0).unsqueeze(-1) # Line up with x to [B, C, T] + beta = self.beta.unsqueeze(0).unsqueeze(-1) + if self.alpha_logscale: + alpha = torch.exp(alpha) + beta = torch.exp(beta) + x = x + (1.0 / (beta + self.no_div_by_zero)) * pow(sin(x * alpha), 2) + + return x diff --git a/BigVGAN/alias_free_activation/cuda/__init__.py b/BigVGAN/alias_free_activation/cuda/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/BigVGAN/alias_free_activation/cuda/activation1d.py b/BigVGAN/alias_free_activation/cuda/activation1d.py new file mode 100644 index 0000000000000000000000000000000000000000..fa4cc2bd5e4457624dfabee5e33741a70fe3a2be --- /dev/null +++ b/BigVGAN/alias_free_activation/cuda/activation1d.py @@ -0,0 +1,77 @@ +# Copyright (c) 2024 NVIDIA CORPORATION. +# Licensed under the MIT license. + +import torch +import torch.nn as nn +from voicerestore.BigVGAN.alias_free_activation.torch.resample import UpSample1d, DownSample1d + +# load fused CUDA kernel: this enables importing anti_alias_activation_cuda +from voicerestore.BigVGAN.alias_free_activation.cuda import load + +anti_alias_activation_cuda = load.load() + + +class FusedAntiAliasActivation(torch.autograd.Function): + """ + Assumes filter size 12, replication padding on upsampling/downsampling, and logscale alpha/beta parameters as inputs. + The hyperparameters are hard-coded in the kernel to maximize speed. + NOTE: The fused kenrel is incorrect for Activation1d with different hyperparameters. + """ + + @staticmethod + def forward(ctx, inputs, up_ftr, down_ftr, alpha, beta): + activation_results = anti_alias_activation_cuda.forward( + inputs, up_ftr, down_ftr, alpha, beta + ) + + return activation_results + + @staticmethod + def backward(ctx, output_grads): + raise NotImplementedError + return output_grads, None, None + + +class Activation1d(nn.Module): + def __init__( + self, + activation, + up_ratio: int = 2, + down_ratio: int = 2, + up_kernel_size: int = 12, + down_kernel_size: int = 12, + fused: bool = True, + ): + super().__init__() + self.up_ratio = up_ratio + self.down_ratio = down_ratio + self.act = activation + self.upsample = UpSample1d(up_ratio, up_kernel_size) + self.downsample = DownSample1d(down_ratio, down_kernel_size) + + self.fused = fused # Whether to use fused CUDA kernel or not + + def forward(self, x): + if not self.fused: + x = self.upsample(x) + x = self.act(x) + x = self.downsample(x) + return x + else: + if self.act.__class__.__name__ == "Snake": + beta = self.act.alpha.data # Snake uses same params for alpha and beta + else: + beta = ( + self.act.beta.data + ) # Snakebeta uses different params for alpha and beta + alpha = self.act.alpha.data + if ( + not self.act.alpha_logscale + ): # Exp baked into cuda kernel, cancel it out with a log + alpha = torch.log(alpha) + beta = torch.log(beta) + + x = FusedAntiAliasActivation.apply( + x, self.upsample.filter, self.downsample.lowpass.filter, alpha, beta + ) + return x diff --git a/BigVGAN/alias_free_activation/cuda/anti_alias_activation.cpp b/BigVGAN/alias_free_activation/cuda/anti_alias_activation.cpp new file mode 100644 index 0000000000000000000000000000000000000000..c5651f77143bd678169eb11564a7cf7a7969a59e --- /dev/null +++ b/BigVGAN/alias_free_activation/cuda/anti_alias_activation.cpp @@ -0,0 +1,23 @@ +/* coding=utf-8 + * Copyright (c) 2024, NVIDIA CORPORATION. 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. + */ + + #include + +extern "C" torch::Tensor fwd_cuda(torch::Tensor const &input, torch::Tensor const &up_filter, torch::Tensor const &down_filter, torch::Tensor const &alpha, torch::Tensor const &beta); + +PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) { + m.def("forward", &fwd_cuda, "Anti-Alias Activation forward (CUDA)"); +} \ No newline at end of file diff --git a/BigVGAN/alias_free_activation/cuda/anti_alias_activation_cuda.cu b/BigVGAN/alias_free_activation/cuda/anti_alias_activation_cuda.cu new file mode 100644 index 0000000000000000000000000000000000000000..8c442334869fe72d639ec203fa4fac07f96a0ee1 --- /dev/null +++ b/BigVGAN/alias_free_activation/cuda/anti_alias_activation_cuda.cu @@ -0,0 +1,246 @@ +/* coding=utf-8 + * Copyright (c) 2024, NVIDIA CORPORATION. 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. + */ + +#include +#include +#include +#include +#include +#include +#include +#include "type_shim.h" +#include +#include +#include +#include +#include + +namespace +{ + // Hard-coded hyperparameters + // WARP_SIZE and WARP_BATCH must match the return values batches_per_warp and + constexpr int ELEMENTS_PER_LDG_STG = 1; //(WARP_ITERATIONS < 4) ? 1 : 4; + constexpr int BUFFER_SIZE = 32; + constexpr int FILTER_SIZE = 12; + constexpr int HALF_FILTER_SIZE = 6; + constexpr int UPSAMPLE_REPLICATION_PAD = 5; // 5 on each side, matching torch impl + constexpr int DOWNSAMPLE_REPLICATION_PAD_LEFT = 5; // matching torch impl + constexpr int DOWNSAMPLE_REPLICATION_PAD_RIGHT = 6; // matching torch impl + + template + __global__ void anti_alias_activation_forward( + output_t *dst, + const input_t *src, + const input_t *up_ftr, + const input_t *down_ftr, + const input_t *alpha, + const input_t *beta, + int batch_size, + int channels, + int seq_len) + { + // Up and downsample filters + input_t up_filter[FILTER_SIZE]; + input_t down_filter[FILTER_SIZE]; + + // Load data from global memory including extra indices reserved for replication paddings + input_t elements[2 * FILTER_SIZE + 2 * BUFFER_SIZE + 2 * UPSAMPLE_REPLICATION_PAD] = {0}; + input_t intermediates[2 * FILTER_SIZE + 2 * BUFFER_SIZE + DOWNSAMPLE_REPLICATION_PAD_LEFT + DOWNSAMPLE_REPLICATION_PAD_RIGHT] = {0}; + + // Output stores downsampled output before writing to dst + output_t output[BUFFER_SIZE]; + + // blockDim/threadIdx = (128, 1, 1) + // gridDim/blockIdx = (seq_blocks, channels, batches) + int block_offset = (blockIdx.x * 128 * BUFFER_SIZE + seq_len * (blockIdx.y + gridDim.y * blockIdx.z)); + int local_offset = threadIdx.x * BUFFER_SIZE; + int seq_offset = blockIdx.x * 128 * BUFFER_SIZE + local_offset; + + // intermediate have double the seq_len + int intermediate_local_offset = threadIdx.x * BUFFER_SIZE * 2; + int intermediate_seq_offset = blockIdx.x * 128 * BUFFER_SIZE * 2 + intermediate_local_offset; + + // Get values needed for replication padding before moving pointer + const input_t *right_most_pntr = src + (seq_len * (blockIdx.y + gridDim.y * blockIdx.z)); + input_t seq_left_most_value = right_most_pntr[0]; + input_t seq_right_most_value = right_most_pntr[seq_len - 1]; + + // Move src and dst pointers + src += block_offset + local_offset; + dst += block_offset + local_offset; + + // Alpha and beta values for snake activatons. Applies exp by default + alpha = alpha + blockIdx.y; + input_t alpha_val = expf(alpha[0]); + beta = beta + blockIdx.y; + input_t beta_val = expf(beta[0]); + + #pragma unroll + for (int it = 0; it < FILTER_SIZE; it += 1) + { + up_filter[it] = up_ftr[it]; + down_filter[it] = down_ftr[it]; + } + + // Apply replication padding for upsampling, matching torch impl + #pragma unroll + for (int it = -HALF_FILTER_SIZE; it < BUFFER_SIZE + HALF_FILTER_SIZE; it += 1) + { + int element_index = seq_offset + it; // index for element + if ((element_index < 0) && (element_index >= -UPSAMPLE_REPLICATION_PAD)) + { + elements[2 * (HALF_FILTER_SIZE + it)] = 2 * seq_left_most_value; + } + if ((element_index >= seq_len) && (element_index < seq_len + UPSAMPLE_REPLICATION_PAD)) + { + elements[2 * (HALF_FILTER_SIZE + it)] = 2 * seq_right_most_value; + } + if ((element_index >= 0) && (element_index < seq_len)) + { + elements[2 * (HALF_FILTER_SIZE + it)] = 2 * src[it]; + } + } + + // Apply upsampling strided convolution and write to intermediates. It reserves DOWNSAMPLE_REPLICATION_PAD_LEFT for replication padding of the downsampilng conv later + #pragma unroll + for (int it = 0; it < (2 * BUFFER_SIZE + 2 * FILTER_SIZE); it += 1) + { + input_t acc = 0.0; + int element_index = intermediate_seq_offset + it; // index for intermediate + #pragma unroll + for (int f_idx = 0; f_idx < FILTER_SIZE; f_idx += 1) + { + if ((element_index + f_idx) >= 0) + { + acc += up_filter[f_idx] * elements[it + f_idx]; + } + } + intermediates[it + DOWNSAMPLE_REPLICATION_PAD_LEFT] = acc; + } + + // Apply activation function. It reserves DOWNSAMPLE_REPLICATION_PAD_LEFT and DOWNSAMPLE_REPLICATION_PAD_RIGHT for replication padding of the downsampilng conv later + double no_div_by_zero = 0.000000001; + #pragma unroll + for (int it = 0; it < 2 * BUFFER_SIZE + 2 * FILTER_SIZE; it += 1) + { + intermediates[it + DOWNSAMPLE_REPLICATION_PAD_LEFT] += (1.0 / (beta_val + no_div_by_zero)) * sinf(intermediates[it + DOWNSAMPLE_REPLICATION_PAD_LEFT] * alpha_val) * sinf(intermediates[it + DOWNSAMPLE_REPLICATION_PAD_LEFT] * alpha_val); + } + + // Apply replication padding before downsampling conv from intermediates + #pragma unroll + for (int it = 0; it < DOWNSAMPLE_REPLICATION_PAD_LEFT; it += 1) + { + intermediates[it] = intermediates[DOWNSAMPLE_REPLICATION_PAD_LEFT]; + } + #pragma unroll + for (int it = DOWNSAMPLE_REPLICATION_PAD_LEFT + 2 * BUFFER_SIZE + 2 * FILTER_SIZE; it < DOWNSAMPLE_REPLICATION_PAD_LEFT + 2 * BUFFER_SIZE + 2 * FILTER_SIZE + DOWNSAMPLE_REPLICATION_PAD_RIGHT; it += 1) + { + intermediates[it] = intermediates[DOWNSAMPLE_REPLICATION_PAD_LEFT + 2 * BUFFER_SIZE + 2 * FILTER_SIZE - 1]; + } + + // Apply downsample strided convolution (assuming stride=2) from intermediates + #pragma unroll + for (int it = 0; it < BUFFER_SIZE; it += 1) + { + input_t acc = 0.0; + #pragma unroll + for (int f_idx = 0; f_idx < FILTER_SIZE; f_idx += 1) + { + // Add constant DOWNSAMPLE_REPLICATION_PAD_RIGHT to match torch implementation + acc += down_filter[f_idx] * intermediates[it * 2 + f_idx + DOWNSAMPLE_REPLICATION_PAD_RIGHT]; + } + output[it] = acc; + } + + // Write output to dst + #pragma unroll + for (int it = 0; it < BUFFER_SIZE; it += ELEMENTS_PER_LDG_STG) + { + int element_index = seq_offset + it; + if (element_index < seq_len) + { + dst[it] = output[it]; + } + } + + } + + template + void dispatch_anti_alias_activation_forward( + output_t *dst, + const input_t *src, + const input_t *up_ftr, + const input_t *down_ftr, + const input_t *alpha, + const input_t *beta, + int batch_size, + int channels, + int seq_len) + { + if (seq_len == 0) + { + return; + } + else + { + // Use 128 threads per block to maximimize gpu utilization + constexpr int threads_per_block = 128; + constexpr int seq_len_per_block = 4096; + int blocks_per_seq_len = (seq_len + seq_len_per_block - 1) / seq_len_per_block; + dim3 blocks(blocks_per_seq_len, channels, batch_size); + dim3 threads(threads_per_block, 1, 1); + + anti_alias_activation_forward + <<>>(dst, src, up_ftr, down_ftr, alpha, beta, batch_size, channels, seq_len); + } + } +} + +extern "C" torch::Tensor fwd_cuda(torch::Tensor const &input, torch::Tensor const &up_filter, torch::Tensor const &down_filter, torch::Tensor const &alpha, torch::Tensor const &beta) +{ + // Input is a 3d tensor with dimensions [batches, channels, seq_len] + const int batches = input.size(0); + const int channels = input.size(1); + const int seq_len = input.size(2); + + // Output + auto act_options = input.options().requires_grad(false); + + torch::Tensor anti_alias_activation_results = + torch::empty({batches, channels, seq_len}, act_options); + + void *input_ptr = static_cast(input.data_ptr()); + void *up_filter_ptr = static_cast(up_filter.data_ptr()); + void *down_filter_ptr = static_cast(down_filter.data_ptr()); + void *alpha_ptr = static_cast(alpha.data_ptr()); + void *beta_ptr = static_cast(beta.data_ptr()); + void *anti_alias_activation_results_ptr = static_cast(anti_alias_activation_results.data_ptr()); + + DISPATCH_FLOAT_HALF_AND_BFLOAT( + input.scalar_type(), + "dispatch anti alias activation_forward", + dispatch_anti_alias_activation_forward( + reinterpret_cast(anti_alias_activation_results_ptr), + reinterpret_cast(input_ptr), + reinterpret_cast(up_filter_ptr), + reinterpret_cast(down_filter_ptr), + reinterpret_cast(alpha_ptr), + reinterpret_cast(beta_ptr), + batches, + channels, + seq_len);); + return anti_alias_activation_results; +} \ No newline at end of file diff --git a/BigVGAN/alias_free_activation/cuda/compat.h b/BigVGAN/alias_free_activation/cuda/compat.h new file mode 100644 index 0000000000000000000000000000000000000000..25818b2edf4cb0dc9130e62c7c4de8d16a01baa5 --- /dev/null +++ b/BigVGAN/alias_free_activation/cuda/compat.h @@ -0,0 +1,29 @@ +/* coding=utf-8 + * Copyright (c) 2020, NVIDIA CORPORATION. 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. + */ + +/*This code is copied fron NVIDIA apex: + * https://github.com/NVIDIA/apex + * with minor changes. */ + +#ifndef TORCH_CHECK +#define TORCH_CHECK AT_CHECK +#endif + +#ifdef VERSION_GE_1_3 +#define DATA_PTR data_ptr +#else +#define DATA_PTR data +#endif diff --git a/BigVGAN/alias_free_activation/cuda/load.py b/BigVGAN/alias_free_activation/cuda/load.py new file mode 100644 index 0000000000000000000000000000000000000000..ca5d01de398249e75e9e2298958764acb436edba --- /dev/null +++ b/BigVGAN/alias_free_activation/cuda/load.py @@ -0,0 +1,86 @@ +# Copyright (c) 2024 NVIDIA CORPORATION. +# Licensed under the MIT license. + +import os +import pathlib +import subprocess + +from torch.utils import cpp_extension + +""" +Setting this param to a list has a problem of generating different compilation commands (with diferent order of architectures) and leading to recompilation of fused kernels. +Set it to empty stringo avoid recompilation and assign arch flags explicity in extra_cuda_cflags below +""" +os.environ["TORCH_CUDA_ARCH_LIST"] = "" + + +def load(): + # Check if cuda 11 is installed for compute capability 8.0 + cc_flag = [] + _, bare_metal_major, _ = _get_cuda_bare_metal_version(cpp_extension.CUDA_HOME) + if int(bare_metal_major) >= 11: + cc_flag.append("-gencode") + cc_flag.append("arch=compute_80,code=sm_80") + + # Build path + srcpath = pathlib.Path(__file__).parent.absolute() + buildpath = srcpath / "build" + _create_build_dir(buildpath) + + # Helper function to build the kernels. + def _cpp_extention_load_helper(name, sources, extra_cuda_flags): + return cpp_extension.load( + name=name, + sources=sources, + build_directory=buildpath, + extra_cflags=[ + "-O3", + ], + extra_cuda_cflags=[ + "-O3", + "-gencode", + "arch=compute_70,code=sm_70", + "--use_fast_math", + ] + + extra_cuda_flags + + cc_flag, + verbose=True, + ) + + extra_cuda_flags = [ + "-U__CUDA_NO_HALF_OPERATORS__", + "-U__CUDA_NO_HALF_CONVERSIONS__", + "--expt-relaxed-constexpr", + "--expt-extended-lambda", + ] + + sources = [ + srcpath / "anti_alias_activation.cpp", + srcpath / "anti_alias_activation_cuda.cu", + ] + anti_alias_activation_cuda = _cpp_extention_load_helper( + "anti_alias_activation_cuda", sources, extra_cuda_flags + ) + + return anti_alias_activation_cuda + + +def _get_cuda_bare_metal_version(cuda_dir): + raw_output = subprocess.check_output( + [cuda_dir + "/bin/nvcc", "-V"], universal_newlines=True + ) + output = raw_output.split() + release_idx = output.index("release") + 1 + release = output[release_idx].split(".") + bare_metal_major = release[0] + bare_metal_minor = release[1][0] + + return raw_output, bare_metal_major, bare_metal_minor + + +def _create_build_dir(buildpath): + try: + os.mkdir(buildpath) + except OSError: + if not os.path.isdir(buildpath): + print(f"Creation of the build directory {buildpath} failed") diff --git a/BigVGAN/alias_free_activation/cuda/type_shim.h b/BigVGAN/alias_free_activation/cuda/type_shim.h new file mode 100644 index 0000000000000000000000000000000000000000..5db7e8a397e982d4d30d16ab6060814b98b7ab83 --- /dev/null +++ b/BigVGAN/alias_free_activation/cuda/type_shim.h @@ -0,0 +1,92 @@ +/* coding=utf-8 + * Copyright (c) 2020, NVIDIA CORPORATION. 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. + */ + +#include +#include "compat.h" + +#define DISPATCH_FLOAT_HALF_AND_BFLOAT(TYPE, NAME, ...) \ + switch (TYPE) \ + { \ + case at::ScalarType::Float: \ + { \ + using scalar_t = float; \ + __VA_ARGS__; \ + break; \ + } \ + case at::ScalarType::Half: \ + { \ + using scalar_t = at::Half; \ + __VA_ARGS__; \ + break; \ + } \ + case at::ScalarType::BFloat16: \ + { \ + using scalar_t = at::BFloat16; \ + __VA_ARGS__; \ + break; \ + } \ + default: \ + AT_ERROR(#NAME, " not implemented for '", toString(TYPE), "'"); \ + } + +#define DISPATCH_FLOAT_HALF_AND_BFLOAT_INOUT_TYPES(TYPEIN, TYPEOUT, NAME, ...) \ + switch (TYPEIN) \ + { \ + case at::ScalarType::Float: \ + { \ + using scalar_t_in = float; \ + switch (TYPEOUT) \ + { \ + case at::ScalarType::Float: \ + { \ + using scalar_t_out = float; \ + __VA_ARGS__; \ + break; \ + } \ + case at::ScalarType::Half: \ + { \ + using scalar_t_out = at::Half; \ + __VA_ARGS__; \ + break; \ + } \ + case at::ScalarType::BFloat16: \ + { \ + using scalar_t_out = at::BFloat16; \ + __VA_ARGS__; \ + break; \ + } \ + default: \ + AT_ERROR(#NAME, " not implemented for '", toString(TYPEOUT), "'"); \ + } \ + break; \ + } \ + case at::ScalarType::Half: \ + { \ + using scalar_t_in = at::Half; \ + using scalar_t_out = at::Half; \ + __VA_ARGS__; \ + break; \ + } \ + case at::ScalarType::BFloat16: \ + { \ + using scalar_t_in = at::BFloat16; \ + using scalar_t_out = at::BFloat16; \ + __VA_ARGS__; \ + break; \ + } \ + default: \ + AT_ERROR(#NAME, " not implemented for '", toString(TYPEIN), "'"); \ + } diff --git a/BigVGAN/alias_free_activation/torch/__init__.py b/BigVGAN/alias_free_activation/torch/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..8f756ed83f87f9839e457b240f60469bc187707d --- /dev/null +++ b/BigVGAN/alias_free_activation/torch/__init__.py @@ -0,0 +1,6 @@ +# Adapted from https://github.com/junjun3518/alias-free-torch under the Apache License 2.0 +# LICENSE is in incl_licenses directory. + +from .filter import * +from .resample import * +from .act import * diff --git a/BigVGAN/alias_free_activation/torch/act.py b/BigVGAN/alias_free_activation/torch/act.py new file mode 100644 index 0000000000000000000000000000000000000000..e19af3eb362c73dfd2e12f027fb77dd85b1729e2 --- /dev/null +++ b/BigVGAN/alias_free_activation/torch/act.py @@ -0,0 +1,30 @@ +# Adapted from https://github.com/junjun3518/alias-free-torch under the Apache License 2.0 +# LICENSE is in incl_licenses directory. + +import torch.nn as nn +from voicerestore.BigVGAN.alias_free_activation.torch.resample import UpSample1d, DownSample1d + + +class Activation1d(nn.Module): + def __init__( + self, + activation, + up_ratio: int = 2, + down_ratio: int = 2, + up_kernel_size: int = 12, + down_kernel_size: int = 12, + ): + super().__init__() + self.up_ratio = up_ratio + self.down_ratio = down_ratio + self.act = activation + self.upsample = UpSample1d(up_ratio, up_kernel_size) + self.downsample = DownSample1d(down_ratio, down_kernel_size) + + # x: [B,C,T] + def forward(self, x): + x = self.upsample(x) + x = self.act(x) + x = self.downsample(x) + + return x diff --git a/BigVGAN/alias_free_activation/torch/filter.py b/BigVGAN/alias_free_activation/torch/filter.py new file mode 100644 index 0000000000000000000000000000000000000000..0fa35b0d5ddf8d6cb04cd9d47364ca033cebcd32 --- /dev/null +++ b/BigVGAN/alias_free_activation/torch/filter.py @@ -0,0 +1,101 @@ +# Adapted from https://github.com/junjun3518/alias-free-torch under the Apache License 2.0 +# LICENSE is in incl_licenses directory. + +import torch +import torch.nn as nn +import torch.nn.functional as F +import math + +if "sinc" in dir(torch): + sinc = torch.sinc +else: + # This code is adopted from adefossez's julius.core.sinc under the MIT License + # https://adefossez.github.io/julius/julius/core.html + # LICENSE is in incl_licenses directory. + def sinc(x: torch.Tensor): + """ + Implementation of sinc, i.e. sin(pi * x) / (pi * x) + __Warning__: Different to julius.sinc, the input is multiplied by `pi`! + """ + return torch.where( + x == 0, + torch.tensor(1.0, device=x.device, dtype=x.dtype), + torch.sin(math.pi * x) / math.pi / x, + ) + + +# This code is adopted from adefossez's julius.lowpass.LowPassFilters under the MIT License +# https://adefossez.github.io/julius/julius/lowpass.html +# LICENSE is in incl_licenses directory. +def kaiser_sinc_filter1d( + cutoff, half_width, kernel_size +): # return filter [1,1,kernel_size] + even = kernel_size % 2 == 0 + half_size = kernel_size // 2 + + # For kaiser window + delta_f = 4 * half_width + A = 2.285 * (half_size - 1) * math.pi * delta_f + 7.95 + if A > 50.0: + beta = 0.1102 * (A - 8.7) + elif A >= 21.0: + beta = 0.5842 * (A - 21) ** 0.4 + 0.07886 * (A - 21.0) + else: + beta = 0.0 + window = torch.kaiser_window(kernel_size, beta=beta, periodic=False) + + # ratio = 0.5/cutoff -> 2 * cutoff = 1 / ratio + if even: + time = torch.arange(-half_size, half_size) + 0.5 + else: + time = torch.arange(kernel_size) - half_size + if cutoff == 0: + filter_ = torch.zeros_like(time) + else: + filter_ = 2 * cutoff * window * sinc(2 * cutoff * time) + """ + Normalize filter to have sum = 1, otherwise we will have a small leakage of the constant component in the input signal. + """ + filter_ /= filter_.sum() + filter = filter_.view(1, 1, kernel_size) + + return filter + + +class LowPassFilter1d(nn.Module): + def __init__( + self, + cutoff=0.5, + half_width=0.6, + stride: int = 1, + padding: bool = True, + padding_mode: str = "replicate", + kernel_size: int = 12, + ): + """ + kernel_size should be even number for stylegan3 setup, in this implementation, odd number is also possible. + """ + super().__init__() + if cutoff < -0.0: + raise ValueError("Minimum cutoff must be larger than zero.") + if cutoff > 0.5: + raise ValueError("A cutoff above 0.5 does not make sense.") + self.kernel_size = kernel_size + self.even = kernel_size % 2 == 0 + self.pad_left = kernel_size // 2 - int(self.even) + self.pad_right = kernel_size // 2 + self.stride = stride + self.padding = padding + self.padding_mode = padding_mode + filter = kaiser_sinc_filter1d(cutoff, half_width, kernel_size) + self.register_buffer("filter", filter) + + # Input [B, C, T] + def forward(self, x): + _, C, _ = x.shape + + if self.padding: + x = F.pad(x, (self.pad_left, self.pad_right), mode=self.padding_mode) + out = F.conv1d(x, self.filter.expand(C, -1, -1), stride=self.stride, groups=C) + + return out diff --git a/BigVGAN/alias_free_activation/torch/resample.py b/BigVGAN/alias_free_activation/torch/resample.py new file mode 100644 index 0000000000000000000000000000000000000000..dfc01fba9d4342fa0abffdfd0cafb59e2143c033 --- /dev/null +++ b/BigVGAN/alias_free_activation/torch/resample.py @@ -0,0 +1,58 @@ +# Adapted from https://github.com/junjun3518/alias-free-torch under the Apache License 2.0 +# LICENSE is in incl_licenses directory. + +import torch.nn as nn +from torch.nn import functional as F +from voicerestore.BigVGAN.alias_free_activation.torch.filter import LowPassFilter1d +from voicerestore.BigVGAN.alias_free_activation.torch.filter import kaiser_sinc_filter1d + + +class UpSample1d(nn.Module): + def __init__(self, ratio=2, kernel_size=None): + super().__init__() + self.ratio = ratio + self.kernel_size = ( + int(6 * ratio // 2) * 2 if kernel_size is None else kernel_size + ) + self.stride = ratio + self.pad = self.kernel_size // ratio - 1 + self.pad_left = self.pad * self.stride + (self.kernel_size - self.stride) // 2 + self.pad_right = ( + self.pad * self.stride + (self.kernel_size - self.stride + 1) // 2 + ) + filter = kaiser_sinc_filter1d( + cutoff=0.5 / ratio, half_width=0.6 / ratio, kernel_size=self.kernel_size + ) + self.register_buffer("filter", filter) + + # x: [B, C, T] + def forward(self, x): + _, C, _ = x.shape + + x = F.pad(x, (self.pad, self.pad), mode="replicate") + x = self.ratio * F.conv_transpose1d( + x, self.filter.expand(C, -1, -1), stride=self.stride, groups=C + ) + x = x[..., self.pad_left : -self.pad_right] + + return x + + +class DownSample1d(nn.Module): + def __init__(self, ratio=2, kernel_size=None): + super().__init__() + self.ratio = ratio + self.kernel_size = ( + int(6 * ratio // 2) * 2 if kernel_size is None else kernel_size + ) + self.lowpass = LowPassFilter1d( + cutoff=0.5 / ratio, + half_width=0.6 / ratio, + stride=ratio, + kernel_size=self.kernel_size, + ) + + def forward(self, x): + xx = self.lowpass(x) + + return xx diff --git a/BigVGAN/bigvgan.py b/BigVGAN/bigvgan.py new file mode 100644 index 0000000000000000000000000000000000000000..bbd01bfdc6481211f9ac913fb1b59f21470c99c1 --- /dev/null +++ b/BigVGAN/bigvgan.py @@ -0,0 +1,494 @@ +# Copyright (c) 2024 NVIDIA CORPORATION. +# Licensed under the MIT license. + +# Adapted from https://github.com/jik876/hifi-gan under the MIT license. +# LICENSE is in incl_licenses directory. + +import os +import json +from pathlib import Path +from typing import Optional, Union, Dict + +import torch +import torch.nn as nn +from torch.nn import Conv1d, ConvTranspose1d +from torch.nn.utils import weight_norm, remove_weight_norm + +# import activations +from . import activations +from .utils import init_weights, get_padding +from .alias_free_activation.torch.act import Activation1d as TorchActivation1d +from .env import AttrDict + +from huggingface_hub import PyTorchModelHubMixin, hf_hub_download + + +def load_hparams_from_json(path) -> AttrDict: + with open(path) as f: + data = f.read() + return AttrDict(json.loads(data)) + + +class AMPBlock1(torch.nn.Module): + """ + AMPBlock applies Snake / SnakeBeta activation functions with trainable parameters that control periodicity, defined for each layer. + AMPBlock1 has additional self.convs2 that contains additional Conv1d layers with a fixed dilation=1 followed by each layer in self.convs1 + + Args: + h (AttrDict): Hyperparameters. + channels (int): Number of convolution channels. + kernel_size (int): Size of the convolution kernel. Default is 3. + dilation (tuple): Dilation rates for the convolutions. Each dilation layer has two convolutions. Default is (1, 3, 5). + activation (str): Activation function type. Should be either 'snake' or 'snakebeta'. Default is None. + """ + + def __init__( + self, + h: AttrDict, + channels: int, + kernel_size: int = 3, + dilation: tuple = (1, 3, 5), + activation: str = None, + ): + super().__init__() + + self.h = h + + self.convs1 = nn.ModuleList( + [ + weight_norm( + Conv1d( + channels, + channels, + kernel_size, + stride=1, + dilation=d, + padding=get_padding(kernel_size, d), + ) + ) + for d in dilation + ] + ) + self.convs1.apply(init_weights) + + self.convs2 = nn.ModuleList( + [ + weight_norm( + Conv1d( + channels, + channels, + kernel_size, + stride=1, + dilation=1, + padding=get_padding(kernel_size, 1), + ) + ) + for _ in range(len(dilation)) + ] + ) + self.convs2.apply(init_weights) + + self.num_layers = len(self.convs1) + len( + self.convs2 + ) # Total number of conv layers + + # Select which Activation1d, lazy-load cuda version to ensure backward compatibility + if self.h.get("use_cuda_kernel", False): + from alias_free_activation.cuda.activation1d import ( + Activation1d as CudaActivation1d, + ) + + Activation1d = CudaActivation1d + else: + Activation1d = TorchActivation1d + + # Activation functions + if activation == "snake": + self.activations = nn.ModuleList( + [ + Activation1d( + activation=activations.Snake( + channels, alpha_logscale=h.snake_logscale + ) + ) + for _ in range(self.num_layers) + ] + ) + elif activation == "snakebeta": + self.activations = nn.ModuleList( + [ + Activation1d( + activation=activations.SnakeBeta( + channels, alpha_logscale=h.snake_logscale + ) + ) + for _ in range(self.num_layers) + ] + ) + else: + raise NotImplementedError( + "activation incorrectly specified. check the config file and look for 'activation'." + ) + + def forward(self, x): + acts1, acts2 = self.activations[::2], self.activations[1::2] + for c1, c2, a1, a2 in zip(self.convs1, self.convs2, acts1, acts2): + xt = a1(x) + xt = c1(xt) + xt = a2(xt) + xt = c2(xt) + x = xt + x + + return x + + def remove_weight_norm(self): + for l in self.convs1: + remove_weight_norm(l) + for l in self.convs2: + remove_weight_norm(l) + + +class AMPBlock2(torch.nn.Module): + """ + AMPBlock applies Snake / SnakeBeta activation functions with trainable parameters that control periodicity, defined for each layer. + Unlike AMPBlock1, AMPBlock2 does not contain extra Conv1d layers with fixed dilation=1 + + Args: + h (AttrDict): Hyperparameters. + channels (int): Number of convolution channels. + kernel_size (int): Size of the convolution kernel. Default is 3. + dilation (tuple): Dilation rates for the convolutions. Each dilation layer has two convolutions. Default is (1, 3, 5). + activation (str): Activation function type. Should be either 'snake' or 'snakebeta'. Default is None. + """ + + def __init__( + self, + h: AttrDict, + channels: int, + kernel_size: int = 3, + dilation: tuple = (1, 3, 5), + activation: str = None, + ): + super().__init__() + + self.h = h + + self.convs = nn.ModuleList( + [ + weight_norm( + Conv1d( + channels, + channels, + kernel_size, + stride=1, + dilation=d, + padding=get_padding(kernel_size, d), + ) + ) + for d in dilation + ] + ) + self.convs.apply(init_weights) + + self.num_layers = len(self.convs) # Total number of conv layers + + # Select which Activation1d, lazy-load cuda version to ensure backward compatibility + if self.h.get("use_cuda_kernel", False): + from alias_free_activation.cuda.activation1d import ( + Activation1d as CudaActivation1d, + ) + + Activation1d = CudaActivation1d + else: + Activation1d = TorchActivation1d + + # Activation functions + if activation == "snake": + self.activations = nn.ModuleList( + [ + Activation1d( + activation=activations.Snake( + channels, alpha_logscale=h.snake_logscale + ) + ) + for _ in range(self.num_layers) + ] + ) + elif activation == "snakebeta": + self.activations = nn.ModuleList( + [ + Activation1d( + activation=activations.SnakeBeta( + channels, alpha_logscale=h.snake_logscale + ) + ) + for _ in range(self.num_layers) + ] + ) + else: + raise NotImplementedError( + "activation incorrectly specified. check the config file and look for 'activation'." + ) + + def forward(self, x): + for c, a in zip(self.convs, self.activations): + xt = a(x) + xt = c(xt) + x = xt + x + return x + + def remove_weight_norm(self): + for l in self.convs: + remove_weight_norm(l) + + +class BigVGAN( + torch.nn.Module, + PyTorchModelHubMixin, + library_name="bigvgan", + repo_url="https://github.com/NVIDIA/BigVGAN", + docs_url="https://github.com/NVIDIA/BigVGAN/blob/main/README.md", + pipeline_tag="audio-to-audio", + license="mit", + tags=["neural-vocoder", "audio-generation", "arxiv:2206.04658"], +): + """ + BigVGAN is a neural vocoder model that applies anti-aliased periodic activation for residual blocks (resblocks). + New in BigVGAN-v2: it can optionally use optimized CUDA kernels for AMP (anti-aliased multi-periodicity) blocks. + + Args: + h (AttrDict): Hyperparameters. + use_cuda_kernel (bool): If set to True, loads optimized CUDA kernels for AMP. This should be used for inference only, as training is not supported with CUDA kernels. + + Note: + - The `use_cuda_kernel` parameter should be used for inference only, as training with CUDA kernels is not supported. + - Ensure that the activation function is correctly specified in the hyperparameters (h.activation). + """ + + def __init__(self, h: AttrDict, use_cuda_kernel: bool = False): + super().__init__() + self.h = h + self.h["use_cuda_kernel"] = use_cuda_kernel + + # Select which Activation1d, lazy-load cuda version to ensure backward compatibility + if self.h.get("use_cuda_kernel", False): + from alias_free_activation.cuda.activation1d import ( + Activation1d as CudaActivation1d, + ) + + Activation1d = CudaActivation1d + else: + Activation1d = TorchActivation1d + + self.num_kernels = len(h.resblock_kernel_sizes) + self.num_upsamples = len(h.upsample_rates) + + # Pre-conv + self.conv_pre = weight_norm( + Conv1d(h.num_mels, h.upsample_initial_channel, 7, 1, padding=3) + ) + + # Define which AMPBlock to use. BigVGAN uses AMPBlock1 as default + if h.resblock == "1": + resblock_class = AMPBlock1 + elif h.resblock == "2": + resblock_class = AMPBlock2 + else: + raise ValueError( + f"Incorrect resblock class specified in hyperparameters. Got {h.resblock}" + ) + + # Transposed conv-based upsamplers. does not apply anti-aliasing + self.ups = nn.ModuleList() + for i, (u, k) in enumerate(zip(h.upsample_rates, h.upsample_kernel_sizes)): + self.ups.append( + nn.ModuleList( + [ + weight_norm( + ConvTranspose1d( + h.upsample_initial_channel // (2**i), + h.upsample_initial_channel // (2 ** (i + 1)), + k, + u, + padding=(k - u) // 2, + ) + ) + ] + ) + ) + + # Residual blocks using anti-aliased multi-periodicity composition modules (AMP) + self.resblocks = nn.ModuleList() + for i in range(len(self.ups)): + ch = h.upsample_initial_channel // (2 ** (i + 1)) + for j, (k, d) in enumerate( + zip(h.resblock_kernel_sizes, h.resblock_dilation_sizes) + ): + self.resblocks.append( + resblock_class(h, ch, k, d, activation=h.activation) + ) + + # Post-conv + activation_post = ( + activations.Snake(ch, alpha_logscale=h.snake_logscale) + if h.activation == "snake" + else ( + activations.SnakeBeta(ch, alpha_logscale=h.snake_logscale) + if h.activation == "snakebeta" + else None + ) + ) + if activation_post is None: + raise NotImplementedError( + "activation incorrectly specified. check the config file and look for 'activation'." + ) + + self.activation_post = Activation1d(activation=activation_post) + + # Whether to use bias for the final conv_post. Default to True for backward compatibility + self.use_bias_at_final = h.get("use_bias_at_final", True) + self.conv_post = weight_norm( + Conv1d(ch, 1, 7, 1, padding=3, bias=self.use_bias_at_final) + ) + + # Weight initialization + for i in range(len(self.ups)): + self.ups[i].apply(init_weights) + self.conv_post.apply(init_weights) + + # Final tanh activation. Defaults to True for backward compatibility + self.use_tanh_at_final = h.get("use_tanh_at_final", True) + + def forward(self, x): + # Pre-conv + x = self.conv_pre(x) + + for i in range(self.num_upsamples): + # Upsampling + for i_up in range(len(self.ups[i])): + x = self.ups[i][i_up](x) + # AMP blocks + xs = None + for j in range(self.num_kernels): + if xs is None: + xs = self.resblocks[i * self.num_kernels + j](x) + else: + xs += self.resblocks[i * self.num_kernels + j](x) + x = xs / self.num_kernels + + # Post-conv + x = self.activation_post(x) + x = self.conv_post(x) + # Final tanh activation + if self.use_tanh_at_final: + x = torch.tanh(x) + else: + x = torch.clamp(x, min=-1.0, max=1.0) # Bound the output to [-1, 1] + + return x + + def remove_weight_norm(self): + try: + print("Removing weight norm...") + for l in self.ups: + for l_i in l: + remove_weight_norm(l_i) + for l in self.resblocks: + l.remove_weight_norm() + remove_weight_norm(self.conv_pre) + remove_weight_norm(self.conv_post) + except ValueError: + print("[INFO] Model already removed weight norm. Skipping!") + pass + + # Additional methods for huggingface_hub support + def _save_pretrained(self, save_directory: Path) -> None: + """Save weights and config.json from a Pytorch model to a local directory.""" + + model_path = save_directory / "bigvgan_generator.pt" + torch.save({"generator": self.state_dict()}, model_path) + + config_path = save_directory / "config.json" + with open(config_path, "w") as config_file: + json.dump(self.h, config_file, indent=4) + + @classmethod + def _from_pretrained( + cls, + *, + model_id: str, + revision: str, + cache_dir: str, + force_download: bool, + proxies: Optional[Dict], + resume_download: bool, + local_files_only: bool, + token: Union[str, bool, None], + map_location: str = "cpu", # Additional argument + strict: bool = False, # Additional argument + use_cuda_kernel: bool = False, + **model_kwargs, + ): + """Load Pytorch pretrained weights and return the loaded model.""" + + # Download and load hyperparameters (h) used by BigVGAN + if os.path.isdir(model_id): + print("Loading config.json from local directory") + config_file = os.path.join(model_id, "config.json") + else: + config_file = hf_hub_download( + repo_id=model_id, + filename="config.json", + revision=revision, + cache_dir=cache_dir, + force_download=force_download, + proxies=proxies, + resume_download=resume_download, + token=token, + local_files_only=local_files_only, + ) + h = load_hparams_from_json(config_file) + + # instantiate BigVGAN using h + if use_cuda_kernel: + print( + f"[WARNING] You have specified use_cuda_kernel=True during BigVGAN.from_pretrained(). Only inference is supported (training is not implemented)!" + ) + print( + f"[WARNING] You need nvcc and ninja installed in your system that matches your PyTorch build is using to build the kernel. If not, the model will fail to initialize or generate incorrect waveform!" + ) + print( + f"[WARNING] For detail, see the official GitHub repository: https://github.com/NVIDIA/BigVGAN?tab=readme-ov-file#using-custom-cuda-kernel-for-synthesis" + ) + model = cls(h, use_cuda_kernel=use_cuda_kernel) + + # Download and load pretrained generator weight + if os.path.isdir(model_id): + print("Loading weights from local directory") + model_file = os.path.join(model_id, "bigvgan_generator.pt") + else: + print(f"Loading weights from {model_id}") + model_file = hf_hub_download( + repo_id=model_id, + filename="bigvgan_generator.pt", + revision=revision, + cache_dir=cache_dir, + force_download=force_download, + proxies=proxies, + resume_download=resume_download, + token=token, + local_files_only=local_files_only, + ) + + checkpoint_dict = torch.load(model_file, map_location=map_location) + + try: + model.load_state_dict(checkpoint_dict["generator"]) + except RuntimeError: + print( + f"[INFO] the pretrained checkpoint does not contain weight norm. Loading the checkpoint after removing weight norm!" + ) + model.remove_weight_norm() + model.load_state_dict(checkpoint_dict["generator"]) + + return model diff --git a/BigVGAN/configs/bigvgan_22khz_80band.json b/BigVGAN/configs/bigvgan_22khz_80band.json new file mode 100644 index 0000000000000000000000000000000000000000..64bca7846edb4e86d7ee22d9ca7a1554cf7f1042 --- /dev/null +++ b/BigVGAN/configs/bigvgan_22khz_80band.json @@ -0,0 +1,45 @@ +{ + "resblock": "1", + "num_gpus": 0, + "batch_size": 32, + "learning_rate": 0.0001, + "adam_b1": 0.8, + "adam_b2": 0.99, + "lr_decay": 0.9999996, + "seed": 1234, + + "upsample_rates": [4,4,2,2,2,2], + "upsample_kernel_sizes": [8,8,4,4,4,4], + "upsample_initial_channel": 1536, + "resblock_kernel_sizes": [3,7,11], + "resblock_dilation_sizes": [[1,3,5], [1,3,5], [1,3,5]], + + "activation": "snakebeta", + "snake_logscale": true, + + "resolutions": [[1024, 120, 600], [2048, 240, 1200], [512, 50, 240]], + "mpd_reshapes": [2, 3, 5, 7, 11], + "use_spectral_norm": false, + "discriminator_channel_mult": 1, + + "segment_size": 8192, + "num_mels": 80, + "num_freq": 1025, + "n_fft": 1024, + "hop_size": 256, + "win_size": 1024, + + "sampling_rate": 22050, + + "fmin": 0, + "fmax": 8000, + "fmax_for_loss": null, + + "num_workers": 4, + + "dist_config": { + "dist_backend": "nccl", + "dist_url": "tcp://localhost:54321", + "world_size": 1 + } +} diff --git a/BigVGAN/configs/bigvgan_24khz_100band.json b/BigVGAN/configs/bigvgan_24khz_100band.json new file mode 100644 index 0000000000000000000000000000000000000000..e7f7ff08f6697a4640d8e28c0b3fe7e62d0c3fc7 --- /dev/null +++ b/BigVGAN/configs/bigvgan_24khz_100band.json @@ -0,0 +1,45 @@ +{ + "resblock": "1", + "num_gpus": 0, + "batch_size": 32, + "learning_rate": 0.0001, + "adam_b1": 0.8, + "adam_b2": 0.99, + "lr_decay": 0.9999996, + "seed": 1234, + + "upsample_rates": [4,4,2,2,2,2], + "upsample_kernel_sizes": [8,8,4,4,4,4], + "upsample_initial_channel": 1536, + "resblock_kernel_sizes": [3,7,11], + "resblock_dilation_sizes": [[1,3,5], [1,3,5], [1,3,5]], + + "activation": "snakebeta", + "snake_logscale": true, + + "resolutions": [[1024, 120, 600], [2048, 240, 1200], [512, 50, 240]], + "mpd_reshapes": [2, 3, 5, 7, 11], + "use_spectral_norm": false, + "discriminator_channel_mult": 1, + + "segment_size": 8192, + "num_mels": 100, + "num_freq": 1025, + "n_fft": 1024, + "hop_size": 256, + "win_size": 1024, + + "sampling_rate": 24000, + + "fmin": 0, + "fmax": 12000, + "fmax_for_loss": null, + + "num_workers": 4, + + "dist_config": { + "dist_backend": "nccl", + "dist_url": "tcp://localhost:54321", + "world_size": 1 + } +} diff --git a/BigVGAN/configs/bigvgan_base_22khz_80band.json b/BigVGAN/configs/bigvgan_base_22khz_80band.json new file mode 100644 index 0000000000000000000000000000000000000000..fd244848308917f4df7ce49bf6b76530fd04cbc2 --- /dev/null +++ b/BigVGAN/configs/bigvgan_base_22khz_80band.json @@ -0,0 +1,45 @@ +{ + "resblock": "1", + "num_gpus": 0, + "batch_size": 32, + "learning_rate": 0.0001, + "adam_b1": 0.8, + "adam_b2": 0.99, + "lr_decay": 0.9999996, + "seed": 1234, + + "upsample_rates": [8,8,2,2], + "upsample_kernel_sizes": [16,16,4,4], + "upsample_initial_channel": 512, + "resblock_kernel_sizes": [3,7,11], + "resblock_dilation_sizes": [[1,3,5], [1,3,5], [1,3,5]], + + "activation": "snakebeta", + "snake_logscale": true, + + "resolutions": [[1024, 120, 600], [2048, 240, 1200], [512, 50, 240]], + "mpd_reshapes": [2, 3, 5, 7, 11], + "use_spectral_norm": false, + "discriminator_channel_mult": 1, + + "segment_size": 8192, + "num_mels": 80, + "num_freq": 1025, + "n_fft": 1024, + "hop_size": 256, + "win_size": 1024, + + "sampling_rate": 22050, + + "fmin": 0, + "fmax": 8000, + "fmax_for_loss": null, + + "num_workers": 4, + + "dist_config": { + "dist_backend": "nccl", + "dist_url": "tcp://localhost:54321", + "world_size": 1 + } +} diff --git a/BigVGAN/configs/bigvgan_base_24khz_100band.json b/BigVGAN/configs/bigvgan_base_24khz_100band.json new file mode 100644 index 0000000000000000000000000000000000000000..0911508cac4a9346ada8c196bfcc228998da6f42 --- /dev/null +++ b/BigVGAN/configs/bigvgan_base_24khz_100band.json @@ -0,0 +1,45 @@ +{ + "resblock": "1", + "num_gpus": 0, + "batch_size": 32, + "learning_rate": 0.0001, + "adam_b1": 0.8, + "adam_b2": 0.99, + "lr_decay": 0.9999996, + "seed": 1234, + + "upsample_rates": [8,8,2,2], + "upsample_kernel_sizes": [16,16,4,4], + "upsample_initial_channel": 512, + "resblock_kernel_sizes": [3,7,11], + "resblock_dilation_sizes": [[1,3,5], [1,3,5], [1,3,5]], + + "activation": "snakebeta", + "snake_logscale": true, + + "resolutions": [[1024, 120, 600], [2048, 240, 1200], [512, 50, 240]], + "mpd_reshapes": [2, 3, 5, 7, 11], + "use_spectral_norm": false, + "discriminator_channel_mult": 1, + + "segment_size": 8192, + "num_mels": 100, + "num_freq": 1025, + "n_fft": 1024, + "hop_size": 256, + "win_size": 1024, + + "sampling_rate": 24000, + + "fmin": 0, + "fmax": 12000, + "fmax_for_loss": null, + + "num_workers": 4, + + "dist_config": { + "dist_backend": "nccl", + "dist_url": "tcp://localhost:54321", + "world_size": 1 + } +} diff --git a/BigVGAN/configs/bigvgan_v2_22khz_80band_256x.json b/BigVGAN/configs/bigvgan_v2_22khz_80band_256x.json new file mode 100644 index 0000000000000000000000000000000000000000..e96bd5fdd5b99767adba7f13bfcd1f777d5c599a --- /dev/null +++ b/BigVGAN/configs/bigvgan_v2_22khz_80band_256x.json @@ -0,0 +1,61 @@ +{ + "resblock": "1", + "num_gpus": 0, + "batch_size": 4, + "learning_rate": 0.0001, + "adam_b1": 0.8, + "adam_b2": 0.99, + "lr_decay": 0.9999996, + "seed": 1234, + + "upsample_rates": [4,4,2,2,2,2], + "upsample_kernel_sizes": [8,8,4,4,4,4], + "upsample_initial_channel": 1536, + "resblock_kernel_sizes": [3,7,11], + "resblock_dilation_sizes": [[1,3,5], [1,3,5], [1,3,5]], + + "use_tanh_at_final": false, + "use_bias_at_final": false, + + "activation": "snakebeta", + "snake_logscale": true, + + "use_cqtd_instead_of_mrd": true, + "cqtd_filters": 128, + "cqtd_max_filters": 1024, + "cqtd_filters_scale": 1, + "cqtd_dilations": [1, 2, 4], + "cqtd_hop_lengths": [512, 256, 256], + "cqtd_n_octaves": [9, 9, 9], + "cqtd_bins_per_octaves": [24, 36, 48], + + "mpd_reshapes": [2, 3, 5, 7, 11], + "use_spectral_norm": false, + "discriminator_channel_mult": 1, + + "use_multiscale_melloss": true, + "lambda_melloss": 15, + + "clip_grad_norm": 500, + + "segment_size": 65536, + "num_mels": 80, + "num_freq": 1025, + "n_fft": 1024, + "hop_size": 256, + "win_size": 1024, + + "sampling_rate": 22050, + + "fmin": 0, + "fmax": null, + "fmax_for_loss": null, + + "num_workers": 4, + + "dist_config": { + "dist_backend": "nccl", + "dist_url": "tcp://localhost:54321", + "world_size": 1 + } +} diff --git a/BigVGAN/configs/bigvgan_v2_22khz_80band_fmax8k_256x.json b/BigVGAN/configs/bigvgan_v2_22khz_80band_fmax8k_256x.json new file mode 100644 index 0000000000000000000000000000000000000000..a3c9699fbe11948f4fd7e3434d2e623a00c802dd --- /dev/null +++ b/BigVGAN/configs/bigvgan_v2_22khz_80band_fmax8k_256x.json @@ -0,0 +1,61 @@ +{ + "resblock": "1", + "num_gpus": 0, + "batch_size": 4, + "learning_rate": 0.0001, + "adam_b1": 0.8, + "adam_b2": 0.99, + "lr_decay": 0.9999996, + "seed": 1234, + + "upsample_rates": [4,4,2,2,2,2], + "upsample_kernel_sizes": [8,8,4,4,4,4], + "upsample_initial_channel": 1536, + "resblock_kernel_sizes": [3,7,11], + "resblock_dilation_sizes": [[1,3,5], [1,3,5], [1,3,5]], + + "use_tanh_at_final": false, + "use_bias_at_final": false, + + "activation": "snakebeta", + "snake_logscale": true, + + "use_cqtd_instead_of_mrd": true, + "cqtd_filters": 128, + "cqtd_max_filters": 1024, + "cqtd_filters_scale": 1, + "cqtd_dilations": [1, 2, 4], + "cqtd_hop_lengths": [512, 256, 256], + "cqtd_n_octaves": [9, 9, 9], + "cqtd_bins_per_octaves": [24, 36, 48], + + "mpd_reshapes": [2, 3, 5, 7, 11], + "use_spectral_norm": false, + "discriminator_channel_mult": 1, + + "use_multiscale_melloss": true, + "lambda_melloss": 15, + + "clip_grad_norm": 500, + + "segment_size": 65536, + "num_mels": 80, + "num_freq": 1025, + "n_fft": 1024, + "hop_size": 256, + "win_size": 1024, + + "sampling_rate": 22050, + + "fmin": 0, + "fmax": 8000, + "fmax_for_loss": null, + + "num_workers": 4, + + "dist_config": { + "dist_backend": "nccl", + "dist_url": "tcp://localhost:54321", + "world_size": 1 + } +} diff --git a/BigVGAN/configs/bigvgan_v2_24khz_100band_256x.json b/BigVGAN/configs/bigvgan_v2_24khz_100band_256x.json new file mode 100644 index 0000000000000000000000000000000000000000..8057ee267c8ed80615362a41892b923a3ccd27e5 --- /dev/null +++ b/BigVGAN/configs/bigvgan_v2_24khz_100band_256x.json @@ -0,0 +1,61 @@ +{ + "resblock": "1", + "num_gpus": 0, + "batch_size": 4, + "learning_rate": 0.0001, + "adam_b1": 0.8, + "adam_b2": 0.99, + "lr_decay": 0.9999996, + "seed": 1234, + + "upsample_rates": [4,4,2,2,2,2], + "upsample_kernel_sizes": [8,8,4,4,4,4], + "upsample_initial_channel": 1536, + "resblock_kernel_sizes": [3,7,11], + "resblock_dilation_sizes": [[1,3,5], [1,3,5], [1,3,5]], + + "use_tanh_at_final": false, + "use_bias_at_final": false, + + "activation": "snakebeta", + "snake_logscale": true, + + "use_cqtd_instead_of_mrd": true, + "cqtd_filters": 128, + "cqtd_max_filters": 1024, + "cqtd_filters_scale": 1, + "cqtd_dilations": [1, 2, 4], + "cqtd_hop_lengths": [512, 256, 256], + "cqtd_n_octaves": [9, 9, 9], + "cqtd_bins_per_octaves": [24, 36, 48], + + "mpd_reshapes": [2, 3, 5, 7, 11], + "use_spectral_norm": false, + "discriminator_channel_mult": 1, + + "use_multiscale_melloss": true, + "lambda_melloss": 15, + + "clip_grad_norm": 500, + + "segment_size": 65536, + "num_mels": 100, + "num_freq": 1025, + "n_fft": 1024, + "hop_size": 256, + "win_size": 1024, + + "sampling_rate": 24000, + + "fmin": 0, + "fmax": null, + "fmax_for_loss": null, + + "num_workers": 4, + + "dist_config": { + "dist_backend": "nccl", + "dist_url": "tcp://localhost:54321", + "world_size": 1 + } +} diff --git a/BigVGAN/configs/bigvgan_v2_44khz_128band_256x.json b/BigVGAN/configs/bigvgan_v2_44khz_128band_256x.json new file mode 100644 index 0000000000000000000000000000000000000000..b6999d3028e5d741ec99b16b34f153e763d0cfec --- /dev/null +++ b/BigVGAN/configs/bigvgan_v2_44khz_128band_256x.json @@ -0,0 +1,61 @@ +{ + "resblock": "1", + "num_gpus": 0, + "batch_size": 4, + "learning_rate": 0.0001, + "adam_b1": 0.8, + "adam_b2": 0.99, + "lr_decay": 0.9999996, + "seed": 1234, + + "upsample_rates": [4,4,2,2,2,2], + "upsample_kernel_sizes": [8,8,4,4,4,4], + "upsample_initial_channel": 1536, + "resblock_kernel_sizes": [3,7,11], + "resblock_dilation_sizes": [[1,3,5], [1,3,5], [1,3,5]], + + "use_tanh_at_final": false, + "use_bias_at_final": false, + + "activation": "snakebeta", + "snake_logscale": true, + + "use_cqtd_instead_of_mrd": true, + "cqtd_filters": 128, + "cqtd_max_filters": 1024, + "cqtd_filters_scale": 1, + "cqtd_dilations": [1, 2, 4], + "cqtd_hop_lengths": [512, 256, 256], + "cqtd_n_octaves": [9, 9, 9], + "cqtd_bins_per_octaves": [24, 36, 48], + + "mpd_reshapes": [2, 3, 5, 7, 11], + "use_spectral_norm": false, + "discriminator_channel_mult": 1, + + "use_multiscale_melloss": true, + "lambda_melloss": 15, + + "clip_grad_norm": 500, + + "segment_size": 65536, + "num_mels": 128, + "num_freq": 1025, + "n_fft": 1024, + "hop_size": 256, + "win_size": 1024, + + "sampling_rate": 44100, + + "fmin": 0, + "fmax": null, + "fmax_for_loss": null, + + "num_workers": 4, + + "dist_config": { + "dist_backend": "nccl", + "dist_url": "tcp://localhost:54321", + "world_size": 1 + } +} diff --git a/BigVGAN/configs/bigvgan_v2_44khz_128band_512x.json b/BigVGAN/configs/bigvgan_v2_44khz_128band_512x.json new file mode 100644 index 0000000000000000000000000000000000000000..2d7176c910ae0969f208f6d28e3f14abca2dbc7f --- /dev/null +++ b/BigVGAN/configs/bigvgan_v2_44khz_128band_512x.json @@ -0,0 +1,61 @@ +{ + "resblock": "1", + "num_gpus": 0, + "batch_size": 4, + "learning_rate": 0.0001, + "adam_b1": 0.8, + "adam_b2": 0.99, + "lr_decay": 0.9999996, + "seed": 1234, + + "upsample_rates": [8,4,2,2,2,2], + "upsample_kernel_sizes": [16,8,4,4,4,4], + "upsample_initial_channel": 1536, + "resblock_kernel_sizes": [3,7,11], + "resblock_dilation_sizes": [[1,3,5], [1,3,5], [1,3,5]], + + "use_tanh_at_final": false, + "use_bias_at_final": false, + + "activation": "snakebeta", + "snake_logscale": true, + + "use_cqtd_instead_of_mrd": true, + "cqtd_filters": 128, + "cqtd_max_filters": 1024, + "cqtd_filters_scale": 1, + "cqtd_dilations": [1, 2, 4], + "cqtd_hop_lengths": [512, 256, 256], + "cqtd_n_octaves": [9, 9, 9], + "cqtd_bins_per_octaves": [24, 36, 48], + + "mpd_reshapes": [2, 3, 5, 7, 11], + "use_spectral_norm": false, + "discriminator_channel_mult": 1, + + "use_multiscale_melloss": true, + "lambda_melloss": 15, + + "clip_grad_norm": 500, + + "segment_size": 65536, + "num_mels": 128, + "num_freq": 2049, + "n_fft": 2048, + "hop_size": 512, + "win_size": 2048, + + "sampling_rate": 44100, + + "fmin": 0, + "fmax": null, + "fmax_for_loss": null, + + "num_workers": 4, + + "dist_config": { + "dist_backend": "nccl", + "dist_url": "tcp://localhost:54321", + "world_size": 1 + } +} diff --git a/BigVGAN/demo/__init__.py b/BigVGAN/demo/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/BigVGAN/demo/app.py b/BigVGAN/demo/app.py new file mode 100644 index 0000000000000000000000000000000000000000..ace2e4f3a1a08020ac059da14eed764967453d10 --- /dev/null +++ b/BigVGAN/demo/app.py @@ -0,0 +1,441 @@ +# Copyright (c) 2024 NVIDIA CORPORATION. +# Licensed under the MIT license. + +import spaces +import gradio as gr +import pandas as pd +import torch +import os +import sys + +# to import modules from parent_dir +parent_dir = os.path.abspath(os.path.join(os.path.dirname(__file__), "..")) +sys.path.append(parent_dir) + +from meldataset import get_mel_spectrogram, MAX_WAV_VALUE +from bigvgan import BigVGAN +import librosa +import numpy as np +from utils import plot_spectrogram +import PIL + +if torch.cuda.is_available(): + device = torch.device("cuda") + torch.backends.cudnn.benchmark = False + print(f"using GPU") +else: + device = torch.device("cpu") + print(f"using CPU") + + +def inference_gradio(input, model_choice): # Input is audio waveform in [T, channel] + sr, audio = input # Unpack input to sampling rate and audio itself + audio = np.transpose(audio) # Transpose to [channel, T] for librosa + audio = audio / MAX_WAV_VALUE # Convert int16 to float range used by BigVGAN + + model = dict_model[model_choice] + + if sr != model.h.sampling_rate: # Convert audio to model's sampling rate + audio = librosa.resample(audio, orig_sr=sr, target_sr=model.h.sampling_rate) + if len(audio.shape) == 2: # Stereo + audio = librosa.to_mono(audio) # Convert to mono if stereo + audio = librosa.util.normalize(audio) * 0.95 + + output, spec_gen = inference_model( + audio, model + ) # Output is generated audio in ndarray, int16 + + spec_plot_gen = plot_spectrogram(spec_gen) + + output_audio = (model.h.sampling_rate, output) # Tuple for gr.Audio output + + buffer = spec_plot_gen.canvas.buffer_rgba() + output_image = PIL.Image.frombuffer( + "RGBA", spec_plot_gen.canvas.get_width_height(), buffer, "raw", "RGBA", 0, 1 + ) + + return output_audio, output_image + + +@spaces.GPU(duration=120) +def inference_model(audio_input, model): + # Load model to device + model.to(device) + + with torch.inference_mode(): + wav = torch.FloatTensor(audio_input) + # Compute mel spectrogram from the ground truth audio + spec_gt = get_mel_spectrogram(wav.unsqueeze(0), model.h).to(device) + + y_g_hat = model(spec_gt) + + audio_gen = y_g_hat.squeeze().cpu() + spec_gen = get_mel_spectrogram(audio_gen.unsqueeze(0), model.h) + audio_gen = audio_gen.numpy() # [T], float [-1, 1] + audio_gen = (audio_gen * MAX_WAV_VALUE).astype("int16") # [T], int16 + spec_gen = spec_gen.squeeze().numpy() # [C, T_frame] + + # Unload to CPU + model.to("cpu") + # Delete GPU tensor + del spec_gt, y_g_hat + + return audio_gen, spec_gen + + +css = """ + a { + color: inherit; + text-decoration: underline; + } + .gradio-container { + font-family: 'IBM Plex Sans', sans-serif; + } + .gr-button { + color: white; + border-color: #000000; + background: #000000; + } + input[type='range'] { + accent-color: #000000; + } + .dark input[type='range'] { + accent-color: #dfdfdf; + } + .container { + max-width: 730px; + margin: auto; + padding-top: 1.5rem; + } + #gallery { + min-height: 22rem; + margin-bottom: 15px; + margin-left: auto; + margin-right: auto; + border-bottom-right-radius: .5rem !important; + border-bottom-left-radius: .5rem !important; + } + #gallery>div>.h-full { + min-height: 20rem; + } + .details:hover { + text-decoration: underline; + } + .gr-button { + white-space: nowrap; + } + .gr-button:focus { + border-color: rgb(147 197 253 / var(--tw-border-opacity)); + outline: none; + box-shadow: var(--tw-ring-offset-shadow), var(--tw-ring-shadow), var(--tw-shadow, 0 0 #0000); + --tw-border-opacity: 1; + --tw-ring-offset-shadow: var(--tw-ring-inset) 0 0 0 var(--tw-ring-offset-width) var(--tw-ring-offset-color); + --tw-ring-shadow: var(--tw-ring-inset) 0 0 0 calc(3px var(--tw-ring-offset-width)) var(--tw-ring-color); + --tw-ring-color: rgb(191 219 254 / var(--tw-ring-opacity)); + --tw-ring-opacity: .5; + } + #advanced-btn { + font-size: .7rem !important; + line-height: 19px; + margin-top: 12px; + margin-bottom: 12px; + padding: 2px 8px; + border-radius: 14px !important; + } + #advanced-options { + margin-bottom: 20px; + } + .footer { + margin-bottom: 45px; + margin-top: 35px; + text-align: center; + border-bottom: 1px solid #e5e5e5; + } + .footer>p { + font-size: .8rem; + display: inline-block; + padding: 0 10px; + transform: translateY(10px); + background: white; + } + .dark .footer { + border-color: #303030; + } + .dark .footer>p { + background: #0b0f19; + } + .acknowledgments h4{ + margin: 1.25em 0 .25em 0; + font-weight: bold; + font-size: 115%; + } + #container-advanced-btns{ + display: flex; + flex-wrap: wrap; + justify-content: space-between; + align-items: center; + } + .animate-spin { + animation: spin 1s linear infinite; + } + @keyframes spin { + from { + transform: rotate(0deg); + } + to { + transform: rotate(360deg); + } + } + #share-btn-container { + display: flex; padding-left: 0.5rem !important; padding-right: 0.5rem !important; background-color: #000000; justify-content: center; align-items: center; border-radius: 9999px !important; width: 13rem; + margin-top: 10px; + margin-left: auto; + } + #share-btn { + all: initial; color: #ffffff;font-weight: 600; cursor:pointer; font-family: 'IBM Plex Sans', sans-serif; margin-left: 0.5rem !important; padding-top: 0.25rem !important; padding-bottom: 0.25rem !important;right:0; + } + #share-btn * { + all: unset; + } + #share-btn-container div:nth-child(-n+2){ + width: auto !important; + min-height: 0px !important; + } + #share-btn-container .wrap { + display: none !important; + } + .gr-form{ + flex: 1 1 50%; border-top-right-radius: 0; border-bottom-right-radius: 0; + } + #prompt-container{ + gap: 0; + } + #generated_id{ + min-height: 700px + } + #setting_id{ + margin-bottom: 12px; + text-align: center; + font-weight: 900; + } +""" + +# Script for loading the models + +LIST_MODEL_ID = [ + "bigvgan_24khz_100band", + "bigvgan_base_24khz_100band", + "bigvgan_22khz_80band", + "bigvgan_base_22khz_80band", + "bigvgan_v2_22khz_80band_256x", + "bigvgan_v2_22khz_80band_fmax8k_256x", + "bigvgan_v2_24khz_100band_256x", + "bigvgan_v2_44khz_128band_256x", + "bigvgan_v2_44khz_128band_512x", +] + +dict_model = {} +dict_config = {} + +for model_name in LIST_MODEL_ID: + + generator = BigVGAN.from_pretrained("nvidia/" + model_name) + generator.remove_weight_norm() + generator.eval() + + dict_model[model_name] = generator + dict_config[model_name] = generator.h + +# Script for Gradio UI + +iface = gr.Blocks(css=css, title="BigVGAN - Demo") + +with iface: + gr.HTML( + """ +
+
+

+ BigVGAN: A Universal Neural Vocoder with Large-Scale Training +

+
+

+ [Paper] [Code] [Demo] [Project page] +

+
+ """ + ) + gr.HTML( + """ +
+

News

+

[Jul 2024] We release BigVGAN-v2 along with pretrained checkpoints. Below are the highlights:

+
    +
  • Custom CUDA kernel for inference: we provide a fused upsampling + activation kernel written in CUDA for accelerated inference speed. Our test shows 1.5 - 3x faster speed on a single A100 GPU.
    • +
  • Improved discriminator and loss: BigVGAN-v2 is trained using a multi-scale sub-band CQT discriminator and a multi-scale mel spectrogram loss.
    • +
  • Larger training data: BigVGAN-v2 is trained using datasets containing diverse audio types, including speech in multiple languages, environmental sounds, and instruments.
    • +
  • We provide pretrained checkpoints of BigVGAN-v2 using diverse audio configurations, supporting up to 44 kHz sampling rate and 512x upsampling ratio. See the table below for the link.
    • +
+
+ """ + ) + gr.HTML( + """ +
+

Model Overview

+ BigVGAN is a universal neural vocoder model that generates audio waveforms using mel spectrogram as inputs. +
+
+ """ + ) + with gr.Accordion("Input"): + + model_choice = gr.Dropdown( + label="Select the model to use", + info="The default model is bigvgan_v2_24khz_100band_256x", + value="bigvgan_v2_24khz_100band_256x", + choices=[m for m in LIST_MODEL_ID], + interactive=True, + ) + + audio_input = gr.Audio( + label="Input Audio", elem_id="input-audio", interactive=True + ) + + button = gr.Button("Submit") + + with gr.Accordion("Output"): + with gr.Column(): + output_audio = gr.Audio(label="Output Audio", elem_id="output-audio") + output_image = gr.Image( + label="Output Mel Spectrogram", elem_id="output-image-gen" + ) + + button.click( + inference_gradio, + inputs=[audio_input, model_choice], + outputs=[output_audio, output_image], + concurrency_limit=10, + ) + + gr.Examples( + [ + [ + os.path.join(os.path.dirname(__file__), "examples/jensen_24k.wav"), + "bigvgan_v2_24khz_100band_256x", + ], + [ + os.path.join(os.path.dirname(__file__), "examples/libritts_24k.wav"), + "bigvgan_v2_24khz_100band_256x", + ], + [ + os.path.join(os.path.dirname(__file__), "examples/queen_24k.wav"), + "bigvgan_v2_24khz_100band_256x", + ], + [ + os.path.join(os.path.dirname(__file__), "examples/dance_24k.wav"), + "bigvgan_v2_24khz_100band_256x", + ], + [ + os.path.join(os.path.dirname(__file__), "examples/megalovania_24k.wav"), + "bigvgan_v2_24khz_100band_256x", + ], + [ + os.path.join(os.path.dirname(__file__), "examples/hifitts_44k.wav"), + "bigvgan_v2_44khz_128band_256x", + ], + [ + os.path.join(os.path.dirname(__file__), "examples/musdbhq_44k.wav"), + "bigvgan_v2_44khz_128band_256x", + ], + [ + os.path.join(os.path.dirname(__file__), "examples/musiccaps1_44k.wav"), + "bigvgan_v2_44khz_128band_256x", + ], + [ + os.path.join(os.path.dirname(__file__), "examples/musiccaps2_44k.wav"), + "bigvgan_v2_44khz_128band_256x", + ], + ], + fn=inference_gradio, + inputs=[audio_input, model_choice], + outputs=[output_audio, output_image], + ) + + # Define the data for the table + data = { + "Model Name": [ + "bigvgan_v2_44khz_128band_512x", + "bigvgan_v2_44khz_128band_256x", + "bigvgan_v2_24khz_100band_256x", + "bigvgan_v2_22khz_80band_256x", + "bigvgan_v2_22khz_80band_fmax8k_256x", + "bigvgan_24khz_100band", + "bigvgan_base_24khz_100band", + "bigvgan_22khz_80band", + "bigvgan_base_22khz_80band", + ], + "Sampling Rate": [ + "44 kHz", + "44 kHz", + "24 kHz", + "22 kHz", + "22 kHz", + "24 kHz", + "24 kHz", + "22 kHz", + "22 kHz", + ], + "Mel band": [128, 128, 100, 80, 80, 100, 100, 80, 80], + "fmax": [22050, 22050, 12000, 11025, 8000, 12000, 12000, 8000, 8000], + "Upsampling Ratio": [512, 256, 256, 256, 256, 256, 256, 256, 256], + "Parameters": [ + "122M", + "112M", + "112M", + "112M", + "112M", + "112M", + "14M", + "112M", + "14M", + ], + "Dataset": [ + "Large-scale Compilation", + "Large-scale Compilation", + "Large-scale Compilation", + "Large-scale Compilation", + "Large-scale Compilation", + "LibriTTS", + "LibriTTS", + "LibriTTS + VCTK + LJSpeech", + "LibriTTS + VCTK + LJSpeech", + ], + "Fine-Tuned": ["No", "No", "No", "No", "No", "No", "No", "No", "No"], + } + + base_url = "https://huggingface.co/nvidia/" + + df = pd.DataFrame(data) + df["Model Name"] = df["Model Name"].apply( + lambda x: f'{x}' + ) + + html_table = gr.HTML( + f""" +
+ {df.to_html(index=False, escape=False, classes='border="1" cellspacing="0" cellpadding="5" style="margin-left: auto; margin-right: auto;')} +

NOTE: The v1 models are trained using speech audio datasets ONLY! (24kHz models: LibriTTS, 22kHz models: LibriTTS + VCTK + LJSpeech).

+
+ """ + ) + +iface.queue() +iface.launch() diff --git a/BigVGAN/demo/examples/dance_24k.wav b/BigVGAN/demo/examples/dance_24k.wav new file mode 100644 index 0000000000000000000000000000000000000000..6d996759b067c5f2b00ef8925b2f8c3fb4bedc71 Binary files /dev/null and b/BigVGAN/demo/examples/dance_24k.wav differ diff --git a/BigVGAN/demo/examples/hifitts_44k.wav b/BigVGAN/demo/examples/hifitts_44k.wav new file mode 100644 index 0000000000000000000000000000000000000000..b5823928e5a884c5f86e706fc94629ae65d51fa8 Binary files /dev/null and b/BigVGAN/demo/examples/hifitts_44k.wav differ diff --git a/BigVGAN/demo/examples/jensen_24k.wav b/BigVGAN/demo/examples/jensen_24k.wav new file mode 100644 index 0000000000000000000000000000000000000000..0096987ad94072e0108b1a420bb8892fb6a2d7ac Binary files /dev/null and b/BigVGAN/demo/examples/jensen_24k.wav differ diff --git a/BigVGAN/demo/examples/libritts_24k.wav b/BigVGAN/demo/examples/libritts_24k.wav new file mode 100644 index 0000000000000000000000000000000000000000..54fb14a947b241137ec000a97be85569eb4c9bc8 Binary files /dev/null and b/BigVGAN/demo/examples/libritts_24k.wav differ diff --git a/BigVGAN/demo/examples/megalovania_24k.wav b/BigVGAN/demo/examples/megalovania_24k.wav new file mode 100644 index 0000000000000000000000000000000000000000..593127d88b4d829a4fd91ede0972db20eac86766 --- /dev/null +++ b/BigVGAN/demo/examples/megalovania_24k.wav @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7970ac637e680876d48ad84e9185db1b21da01929fe46d855e8794bd83d14c20 +size 1548328 diff --git a/BigVGAN/demo/examples/musdbhq_44k.wav b/BigVGAN/demo/examples/musdbhq_44k.wav new file mode 100644 index 0000000000000000000000000000000000000000..77176040e051a5063e9175b88582fc1fbea614a5 Binary files /dev/null and b/BigVGAN/demo/examples/musdbhq_44k.wav differ diff --git a/BigVGAN/demo/examples/musiccaps1_44k.wav b/BigVGAN/demo/examples/musiccaps1_44k.wav new file mode 100644 index 0000000000000000000000000000000000000000..50a321456021a712ebd43155577785b76038d393 Binary files /dev/null and b/BigVGAN/demo/examples/musiccaps1_44k.wav differ diff --git a/BigVGAN/demo/examples/musiccaps2_44k.wav b/BigVGAN/demo/examples/musiccaps2_44k.wav new file mode 100644 index 0000000000000000000000000000000000000000..debb5cabe73dc408d247e588e957085ad10b653e Binary files /dev/null and b/BigVGAN/demo/examples/musiccaps2_44k.wav differ diff --git a/BigVGAN/demo/examples/queen_24k.wav b/BigVGAN/demo/examples/queen_24k.wav new file mode 100644 index 0000000000000000000000000000000000000000..e40af250f95d4ed6dbf37a2aa5018e1c4ead044b Binary files /dev/null and b/BigVGAN/demo/examples/queen_24k.wav differ diff --git a/BigVGAN/demo/requirements.txt b/BigVGAN/demo/requirements.txt new file mode 100644 index 0000000000000000000000000000000000000000..0c87f2852395a0e6a9a769c52bc27b396429e193 --- /dev/null +++ b/BigVGAN/demo/requirements.txt @@ -0,0 +1,15 @@ +torch +numpy +librosa>=0.8.1 +scipy +tensorboard +soundfile +matplotlib +pesq +auraloss +tqdm +nnAudio +ninja +huggingface_hub>=0.23.4 +gradio>=4.38.1 +spaces>=0.28.3 \ No newline at end of file diff --git a/BigVGAN/discriminators.py b/BigVGAN/discriminators.py new file mode 100644 index 0000000000000000000000000000000000000000..ffdf327bdbd51cbaff4c9a8ea4e3a3b7668e575c --- /dev/null +++ b/BigVGAN/discriminators.py @@ -0,0 +1,651 @@ +# Copyright (c) 2024 NVIDIA CORPORATION. +# Licensed under the MIT license. + +# Adapted from https://github.com/jik876/hifi-gan under the MIT license. +# LICENSE is in incl_licenses directory. + + +import torch +import torch.nn.functional as F +import torch.nn as nn +from torch.nn import Conv2d +from torch.nn.utils import weight_norm, spectral_norm +from torchaudio.transforms import Spectrogram, Resample + +from env import AttrDict +from utils import get_padding +import typing +from typing import Optional, List, Union, Dict, Tuple + + +class DiscriminatorP(torch.nn.Module): + def __init__( + self, + h: AttrDict, + period: List[int], + kernel_size: int = 5, + stride: int = 3, + use_spectral_norm: bool = False, + ): + super().__init__() + self.period = period + self.d_mult = h.discriminator_channel_mult + norm_f = weight_norm if not use_spectral_norm else spectral_norm + + self.convs = nn.ModuleList( + [ + norm_f( + Conv2d( + 1, + int(32 * self.d_mult), + (kernel_size, 1), + (stride, 1), + padding=(get_padding(5, 1), 0), + ) + ), + norm_f( + Conv2d( + int(32 * self.d_mult), + int(128 * self.d_mult), + (kernel_size, 1), + (stride, 1), + padding=(get_padding(5, 1), 0), + ) + ), + norm_f( + Conv2d( + int(128 * self.d_mult), + int(512 * self.d_mult), + (kernel_size, 1), + (stride, 1), + padding=(get_padding(5, 1), 0), + ) + ), + norm_f( + Conv2d( + int(512 * self.d_mult), + int(1024 * self.d_mult), + (kernel_size, 1), + (stride, 1), + padding=(get_padding(5, 1), 0), + ) + ), + norm_f( + Conv2d( + int(1024 * self.d_mult), + int(1024 * self.d_mult), + (kernel_size, 1), + 1, + padding=(2, 0), + ) + ), + ] + ) + self.conv_post = norm_f( + Conv2d(int(1024 * self.d_mult), 1, (3, 1), 1, padding=(1, 0)) + ) + + def forward(self, x: torch.Tensor) -> Tuple[torch.Tensor, List[torch.Tensor]]: + fmap = [] + + # 1d to 2d + b, c, t = x.shape + if t % self.period != 0: # pad first + n_pad = self.period - (t % self.period) + x = F.pad(x, (0, n_pad), "reflect") + t = t + n_pad + x = x.view(b, c, t // self.period, self.period) + + for l in self.convs: + x = l(x) + x = F.leaky_relu(x, 0.1) + fmap.append(x) + x = self.conv_post(x) + fmap.append(x) + x = torch.flatten(x, 1, -1) + + return x, fmap + + +class MultiPeriodDiscriminator(torch.nn.Module): + def __init__(self, h: AttrDict): + super().__init__() + self.mpd_reshapes = h.mpd_reshapes + print(f"mpd_reshapes: {self.mpd_reshapes}") + self.discriminators = nn.ModuleList( + [ + DiscriminatorP(h, rs, use_spectral_norm=h.use_spectral_norm) + for rs in self.mpd_reshapes + ] + ) + + def forward(self, y: torch.Tensor, y_hat: torch.Tensor) -> Tuple[ + List[torch.Tensor], + List[torch.Tensor], + List[List[torch.Tensor]], + List[List[torch.Tensor]], + ]: + y_d_rs = [] + y_d_gs = [] + fmap_rs = [] + fmap_gs = [] + for i, d in enumerate(self.discriminators): + y_d_r, fmap_r = d(y) + y_d_g, fmap_g = d(y_hat) + y_d_rs.append(y_d_r) + fmap_rs.append(fmap_r) + y_d_gs.append(y_d_g) + fmap_gs.append(fmap_g) + + return y_d_rs, y_d_gs, fmap_rs, fmap_gs + + +class DiscriminatorR(nn.Module): + def __init__(self, cfg: AttrDict, resolution: List[List[int]]): + super().__init__() + + self.resolution = resolution + assert ( + len(self.resolution) == 3 + ), f"MRD layer requires list with len=3, got {self.resolution}" + self.lrelu_slope = 0.1 + + norm_f = weight_norm if cfg.use_spectral_norm == False else spectral_norm + if hasattr(cfg, "mrd_use_spectral_norm"): + print( + f"[INFO] overriding MRD use_spectral_norm as {cfg.mrd_use_spectral_norm}" + ) + norm_f = ( + weight_norm if cfg.mrd_use_spectral_norm == False else spectral_norm + ) + self.d_mult = cfg.discriminator_channel_mult + if hasattr(cfg, "mrd_channel_mult"): + print(f"[INFO] overriding mrd channel multiplier as {cfg.mrd_channel_mult}") + self.d_mult = cfg.mrd_channel_mult + + self.convs = nn.ModuleList( + [ + norm_f(nn.Conv2d(1, int(32 * self.d_mult), (3, 9), padding=(1, 4))), + norm_f( + nn.Conv2d( + int(32 * self.d_mult), + int(32 * self.d_mult), + (3, 9), + stride=(1, 2), + padding=(1, 4), + ) + ), + norm_f( + nn.Conv2d( + int(32 * self.d_mult), + int(32 * self.d_mult), + (3, 9), + stride=(1, 2), + padding=(1, 4), + ) + ), + norm_f( + nn.Conv2d( + int(32 * self.d_mult), + int(32 * self.d_mult), + (3, 9), + stride=(1, 2), + padding=(1, 4), + ) + ), + norm_f( + nn.Conv2d( + int(32 * self.d_mult), + int(32 * self.d_mult), + (3, 3), + padding=(1, 1), + ) + ), + ] + ) + self.conv_post = norm_f( + nn.Conv2d(int(32 * self.d_mult), 1, (3, 3), padding=(1, 1)) + ) + + def forward(self, x: torch.Tensor) -> Tuple[torch.Tensor, List[torch.Tensor]]: + fmap = [] + + x = self.spectrogram(x) + x = x.unsqueeze(1) + for l in self.convs: + x = l(x) + x = F.leaky_relu(x, self.lrelu_slope) + fmap.append(x) + x = self.conv_post(x) + fmap.append(x) + x = torch.flatten(x, 1, -1) + + return x, fmap + + def spectrogram(self, x: torch.Tensor) -> torch.Tensor: + n_fft, hop_length, win_length = self.resolution + x = F.pad( + x, + (int((n_fft - hop_length) / 2), int((n_fft - hop_length) / 2)), + mode="reflect", + ) + x = x.squeeze(1) + x = torch.stft( + x, + n_fft=n_fft, + hop_length=hop_length, + win_length=win_length, + center=False, + return_complex=True, + ) + x = torch.view_as_real(x) # [B, F, TT, 2] + mag = torch.norm(x, p=2, dim=-1) # [B, F, TT] + + return mag + + +class MultiResolutionDiscriminator(nn.Module): + def __init__(self, cfg, debug=False): + super().__init__() + self.resolutions = cfg.resolutions + assert ( + len(self.resolutions) == 3 + ), f"MRD requires list of list with len=3, each element having a list with len=3. Got {self.resolutions}" + self.discriminators = nn.ModuleList( + [DiscriminatorR(cfg, resolution) for resolution in self.resolutions] + ) + + def forward(self, y: torch.Tensor, y_hat: torch.Tensor) -> Tuple[ + List[torch.Tensor], + List[torch.Tensor], + List[List[torch.Tensor]], + List[List[torch.Tensor]], + ]: + y_d_rs = [] + y_d_gs = [] + fmap_rs = [] + fmap_gs = [] + + for i, d in enumerate(self.discriminators): + y_d_r, fmap_r = d(x=y) + y_d_g, fmap_g = d(x=y_hat) + y_d_rs.append(y_d_r) + fmap_rs.append(fmap_r) + y_d_gs.append(y_d_g) + fmap_gs.append(fmap_g) + + return y_d_rs, y_d_gs, fmap_rs, fmap_gs + + +# Method based on descript-audio-codec: https://github.com/descriptinc/descript-audio-codec +# Modified code adapted from https://github.com/gemelo-ai/vocos under the MIT license. +# LICENSE is in incl_licenses directory. +class DiscriminatorB(nn.Module): + def __init__( + self, + window_length: int, + channels: int = 32, + hop_factor: float = 0.25, + bands: Tuple[Tuple[float, float], ...] = ( + (0.0, 0.1), + (0.1, 0.25), + (0.25, 0.5), + (0.5, 0.75), + (0.75, 1.0), + ), + ): + super().__init__() + self.window_length = window_length + self.hop_factor = hop_factor + self.spec_fn = Spectrogram( + n_fft=window_length, + hop_length=int(window_length * hop_factor), + win_length=window_length, + power=None, + ) + n_fft = window_length // 2 + 1 + bands = [(int(b[0] * n_fft), int(b[1] * n_fft)) for b in bands] + self.bands = bands + convs = lambda: nn.ModuleList( + [ + weight_norm(nn.Conv2d(2, channels, (3, 9), (1, 1), padding=(1, 4))), + weight_norm( + nn.Conv2d(channels, channels, (3, 9), (1, 2), padding=(1, 4)) + ), + weight_norm( + nn.Conv2d(channels, channels, (3, 9), (1, 2), padding=(1, 4)) + ), + weight_norm( + nn.Conv2d(channels, channels, (3, 9), (1, 2), padding=(1, 4)) + ), + weight_norm( + nn.Conv2d(channels, channels, (3, 3), (1, 1), padding=(1, 1)) + ), + ] + ) + self.band_convs = nn.ModuleList([convs() for _ in range(len(self.bands))]) + + self.conv_post = weight_norm( + nn.Conv2d(channels, 1, (3, 3), (1, 1), padding=(1, 1)) + ) + + def spectrogram(self, x: torch.Tensor) -> List[torch.Tensor]: + # Remove DC offset + x = x - x.mean(dim=-1, keepdims=True) + # Peak normalize the volume of input audio + x = 0.8 * x / (x.abs().max(dim=-1, keepdim=True)[0] + 1e-9) + x = self.spec_fn(x) + x = torch.view_as_real(x) + x = x.permute(0, 3, 2, 1) # [B, F, T, C] -> [B, C, T, F] + # Split into bands + x_bands = [x[..., b[0] : b[1]] for b in self.bands] + return x_bands + + def forward(self, x: torch.Tensor) -> Tuple[torch.Tensor, List[torch.Tensor]]: + x_bands = self.spectrogram(x.squeeze(1)) + fmap = [] + x = [] + + for band, stack in zip(x_bands, self.band_convs): + for i, layer in enumerate(stack): + band = layer(band) + band = torch.nn.functional.leaky_relu(band, 0.1) + if i > 0: + fmap.append(band) + x.append(band) + + x = torch.cat(x, dim=-1) + x = self.conv_post(x) + fmap.append(x) + + return x, fmap + + +# Method based on descript-audio-codec: https://github.com/descriptinc/descript-audio-codec +# Modified code adapted from https://github.com/gemelo-ai/vocos under the MIT license. +# LICENSE is in incl_licenses directory. +class MultiBandDiscriminator(nn.Module): + def __init__( + self, + h, + ): + """ + Multi-band multi-scale STFT discriminator, with the architecture based on https://github.com/descriptinc/descript-audio-codec. + and the modified code adapted from https://github.com/gemelo-ai/vocos. + """ + super().__init__() + # fft_sizes (list[int]): Tuple of window lengths for FFT. Defaults to [2048, 1024, 512] if not set in h. + self.fft_sizes = h.get("mbd_fft_sizes", [2048, 1024, 512]) + self.discriminators = nn.ModuleList( + [DiscriminatorB(window_length=w) for w in self.fft_sizes] + ) + + def forward(self, y: torch.Tensor, y_hat: torch.Tensor) -> Tuple[ + List[torch.Tensor], + List[torch.Tensor], + List[List[torch.Tensor]], + List[List[torch.Tensor]], + ]: + + y_d_rs = [] + y_d_gs = [] + fmap_rs = [] + fmap_gs = [] + + for d in self.discriminators: + y_d_r, fmap_r = d(x=y) + y_d_g, fmap_g = d(x=y_hat) + y_d_rs.append(y_d_r) + fmap_rs.append(fmap_r) + y_d_gs.append(y_d_g) + fmap_gs.append(fmap_g) + + return y_d_rs, y_d_gs, fmap_rs, fmap_gs + + +# Adapted from https://github.com/open-mmlab/Amphion/blob/main/models/vocoders/gan/discriminator/mssbcqtd.py under the MIT license. +# LICENSE is in incl_licenses directory. +class DiscriminatorCQT(nn.Module): + def __init__(self, cfg: AttrDict, hop_length: int, n_octaves:int, bins_per_octave: int): + super().__init__() + self.cfg = cfg + + self.filters = cfg["cqtd_filters"] + self.max_filters = cfg["cqtd_max_filters"] + self.filters_scale = cfg["cqtd_filters_scale"] + self.kernel_size = (3, 9) + self.dilations = cfg["cqtd_dilations"] + self.stride = (1, 2) + + self.in_channels = cfg["cqtd_in_channels"] + self.out_channels = cfg["cqtd_out_channels"] + self.fs = cfg["sampling_rate"] + self.hop_length = hop_length + self.n_octaves = n_octaves + self.bins_per_octave = bins_per_octave + + # Lazy-load + from nnAudio import features + + self.cqt_transform = features.cqt.CQT2010v2( + sr=self.fs * 2, + hop_length=self.hop_length, + n_bins=self.bins_per_octave * self.n_octaves, + bins_per_octave=self.bins_per_octave, + output_format="Complex", + pad_mode="constant", + ) + + self.conv_pres = nn.ModuleList() + for _ in range(self.n_octaves): + self.conv_pres.append( + nn.Conv2d( + self.in_channels * 2, + self.in_channels * 2, + kernel_size=self.kernel_size, + padding=self.get_2d_padding(self.kernel_size), + ) + ) + + self.convs = nn.ModuleList() + + self.convs.append( + nn.Conv2d( + self.in_channels * 2, + self.filters, + kernel_size=self.kernel_size, + padding=self.get_2d_padding(self.kernel_size), + ) + ) + + in_chs = min(self.filters_scale * self.filters, self.max_filters) + for i, dilation in enumerate(self.dilations): + out_chs = min( + (self.filters_scale ** (i + 1)) * self.filters, self.max_filters + ) + self.convs.append( + weight_norm( + nn.Conv2d( + in_chs, + out_chs, + kernel_size=self.kernel_size, + stride=self.stride, + dilation=(dilation, 1), + padding=self.get_2d_padding(self.kernel_size, (dilation, 1)), + ) + ) + ) + in_chs = out_chs + out_chs = min( + (self.filters_scale ** (len(self.dilations) + 1)) * self.filters, + self.max_filters, + ) + self.convs.append( + weight_norm( + nn.Conv2d( + in_chs, + out_chs, + kernel_size=(self.kernel_size[0], self.kernel_size[0]), + padding=self.get_2d_padding( + (self.kernel_size[0], self.kernel_size[0]) + ), + ) + ) + ) + + self.conv_post = weight_norm( + nn.Conv2d( + out_chs, + self.out_channels, + kernel_size=(self.kernel_size[0], self.kernel_size[0]), + padding=self.get_2d_padding((self.kernel_size[0], self.kernel_size[0])), + ) + ) + + self.activation = torch.nn.LeakyReLU(negative_slope=0.1) + self.resample = Resample(orig_freq=self.fs, new_freq=self.fs * 2) + + self.cqtd_normalize_volume = self.cfg.get("cqtd_normalize_volume", False) + if self.cqtd_normalize_volume: + print( + f"[INFO] cqtd_normalize_volume set to True. Will apply DC offset removal & peak volume normalization in CQTD!" + ) + + def get_2d_padding( + self, + kernel_size: typing.Tuple[int, int], + dilation: typing.Tuple[int, int] = (1, 1), + ): + return ( + ((kernel_size[0] - 1) * dilation[0]) // 2, + ((kernel_size[1] - 1) * dilation[1]) // 2, + ) + + def forward(self, x: torch.tensor) -> Tuple[torch.Tensor, List[torch.Tensor]]: + fmap = [] + + if self.cqtd_normalize_volume: + # Remove DC offset + x = x - x.mean(dim=-1, keepdims=True) + # Peak normalize the volume of input audio + x = 0.8 * x / (x.abs().max(dim=-1, keepdim=True)[0] + 1e-9) + + x = self.resample(x) + + z = self.cqt_transform(x) + + z_amplitude = z[:, :, :, 0].unsqueeze(1) + z_phase = z[:, :, :, 1].unsqueeze(1) + + z = torch.cat([z_amplitude, z_phase], dim=1) + z = torch.permute(z, (0, 1, 3, 2)) # [B, C, W, T] -> [B, C, T, W] + + latent_z = [] + for i in range(self.n_octaves): + latent_z.append( + self.conv_pres[i]( + z[ + :, + :, + :, + i * self.bins_per_octave : (i + 1) * self.bins_per_octave, + ] + ) + ) + latent_z = torch.cat(latent_z, dim=-1) + + for i, l in enumerate(self.convs): + latent_z = l(latent_z) + + latent_z = self.activation(latent_z) + fmap.append(latent_z) + + latent_z = self.conv_post(latent_z) + + return latent_z, fmap + + +class MultiScaleSubbandCQTDiscriminator(nn.Module): + def __init__(self, cfg: AttrDict): + super().__init__() + + self.cfg = cfg + # Using get with defaults + self.cfg["cqtd_filters"] = self.cfg.get("cqtd_filters", 32) + self.cfg["cqtd_max_filters"] = self.cfg.get("cqtd_max_filters", 1024) + self.cfg["cqtd_filters_scale"] = self.cfg.get("cqtd_filters_scale", 1) + self.cfg["cqtd_dilations"] = self.cfg.get("cqtd_dilations", [1, 2, 4]) + self.cfg["cqtd_in_channels"] = self.cfg.get("cqtd_in_channels", 1) + self.cfg["cqtd_out_channels"] = self.cfg.get("cqtd_out_channels", 1) + # Multi-scale params to loop over + self.cfg["cqtd_hop_lengths"] = self.cfg.get("cqtd_hop_lengths", [512, 256, 256]) + self.cfg["cqtd_n_octaves"] = self.cfg.get("cqtd_n_octaves", [9, 9, 9]) + self.cfg["cqtd_bins_per_octaves"] = self.cfg.get( + "cqtd_bins_per_octaves", [24, 36, 48] + ) + + self.discriminators = nn.ModuleList( + [ + DiscriminatorCQT( + self.cfg, + hop_length=self.cfg["cqtd_hop_lengths"][i], + n_octaves=self.cfg["cqtd_n_octaves"][i], + bins_per_octave=self.cfg["cqtd_bins_per_octaves"][i], + ) + for i in range(len(self.cfg["cqtd_hop_lengths"])) + ] + ) + + def forward(self, y: torch.Tensor, y_hat: torch.Tensor) -> Tuple[ + List[torch.Tensor], + List[torch.Tensor], + List[List[torch.Tensor]], + List[List[torch.Tensor]], + ]: + + y_d_rs = [] + y_d_gs = [] + fmap_rs = [] + fmap_gs = [] + + for disc in self.discriminators: + y_d_r, fmap_r = disc(y) + y_d_g, fmap_g = disc(y_hat) + y_d_rs.append(y_d_r) + fmap_rs.append(fmap_r) + y_d_gs.append(y_d_g) + fmap_gs.append(fmap_g) + + return y_d_rs, y_d_gs, fmap_rs, fmap_gs + + +class CombinedDiscriminator(nn.Module): + """ + Wrapper of chaining multiple discrimiantor architectures. + Example: combine mbd and cqtd as a single class + """ + + def __init__(self, list_discriminator: List[nn.Module]): + super().__init__() + self.discrimiantor = nn.ModuleList(list_discriminator) + + def forward(self, y: torch.Tensor, y_hat: torch.Tensor) -> Tuple[ + List[torch.Tensor], + List[torch.Tensor], + List[List[torch.Tensor]], + List[List[torch.Tensor]], + ]: + + y_d_rs = [] + y_d_gs = [] + fmap_rs = [] + fmap_gs = [] + + for disc in self.discrimiantor: + y_d_r, y_d_g, fmap_r, fmap_g = disc(y, y_hat) + y_d_rs.extend(y_d_r) + fmap_rs.extend(fmap_r) + y_d_gs.extend(y_d_g) + fmap_gs.extend(fmap_g) + + return y_d_rs, y_d_gs, fmap_rs, fmap_gs diff --git a/BigVGAN/env.py b/BigVGAN/env.py new file mode 100644 index 0000000000000000000000000000000000000000..cf8ac6cea644c78d115dd3902b902993f366ee61 --- /dev/null +++ b/BigVGAN/env.py @@ -0,0 +1,18 @@ +# Adapted from https://github.com/jik876/hifi-gan under the MIT license. +# LICENSE is in incl_licenses directory. + +import os +import shutil + + +class AttrDict(dict): + def __init__(self, *args, **kwargs): + super(AttrDict, self).__init__(*args, **kwargs) + self.__dict__ = self + + +def build_env(config, config_name, path): + t_path = os.path.join(path, config_name) + if config != t_path: + os.makedirs(path, exist_ok=True) + shutil.copyfile(config, os.path.join(path, config_name)) diff --git a/BigVGAN/filelists/LibriTTS/dev-clean.txt b/BigVGAN/filelists/LibriTTS/dev-clean.txt new file mode 100644 index 0000000000000000000000000000000000000000..563b86e601c12604b511548fe50a21d57524e438 --- /dev/null +++ b/BigVGAN/filelists/LibriTTS/dev-clean.txt @@ -0,0 +1,115 @@ +dev-clean/1272/128104/1272_128104_000001_000000|A 'JOLLY' ART CRITIC +dev-clean/1272/141231/1272_141231_000007_000003|And when he attacked, it was always there to beat him aside. +dev-clean/1272/141231/1272_141231_000033_000002|If anything, he was pressing the attack. +dev-clean/1462/170138/1462_170138_000012_000002|Dear me, Mac, the girl couldn't possibly be better, you know." +dev-clean/1462/170142/1462_170142_000002_000005|Alexander did not sit down. +dev-clean/1462/170142/1462_170142_000029_000001|"I meant to, but somehow I couldn't. +dev-clean/1462/170142/1462_170142_000046_000004|The sight of you, Bartley, to see you living and happy and successful-can I never make you understand what that means to me?" She pressed his shoulders gently. +dev-clean/1462/170145/1462_170145_000012_000003|There is a letter for you there, in my desk drawer. +dev-clean/1462/170145/1462_170145_000033_000000|She felt the strength leap in the arms that held her so lightly. +dev-clean/1673/143397/1673_143397_000031_000007|He attempted to remove or intimidate the leaders by a common sentence, of acquittal or condemnation; he invested his representatives at Ephesus with ample power and military force; he summoned from either party eight chosen deputies to a free and candid conference in the neighborhood of the capital, far from the contagion of popular frenzy. +dev-clean/174/168635/174_168635_000040_000000|To teach Cosette to read, and to let her play, this constituted nearly the whole of Jean Valjean's existence. +dev-clean/174/50561/174_50561_000058_000001|They have the end of the game to themselves.) +dev-clean/174/84280/174_84280_000015_000000|And perhaps in this story I have said enough for you to understand why Mary has identified herself with something world-wide, has added to herself a symbolical value, and why it is I find in the whole crowded spectacle of mankind, a quality that is also hers, a sense of fine things entangled and stifled and unable to free themselves from the ancient limiting jealousies which law and custom embody. +dev-clean/1919/142785/1919_142785_000063_000000|[Illustration: SHALOT.] +dev-clean/1919/142785/1919_142785_000131_000001|Cut the bread into thin slices, place them in a cool oven overnight, and when thoroughly dry and crisp, roll them down into fine crumbs. +dev-clean/1988/147956/1988_147956_000016_000009|He was neatly dressed. +dev-clean/1988/148538/1988_148538_000015_000007|These persons then displayed towards each other precisely the same puerile jealousies which animate the men of democracies, the same eagerness to snatch the smallest advantages which their equals contested, and the same desire to parade ostentatiously those of which they were in possession. +dev-clean/1988/24833/1988_24833_000028_000003|He's taking the kid for a walk when a thunderstorm blows up. +dev-clean/1988/24833/1988_24833_000059_000000|"Doesn't pay enough?" Pop asks. +dev-clean/1993/147149/1993_147149_000051_000002|So leaving kind messages to George and Jane Wilson, and hesitating whether she might dare to send a few kind words to Jem, and deciding that she had better not, she stepped out into the bright morning light, so fresh a contrast to the darkened room where death had been. +dev-clean/1993/147965/1993_147965_000003_000004|I suppose, in the crowded clutter of their cave, the old man had come to believe that peace and order had vanished from the earth, or existed only in the old world he had left so far behind. +dev-clean/1993/147966/1993_147966_000020_000003|We found the chickens asleep; perhaps they thought night had come to stay. +dev-clean/2035/147960/2035_147960_000019_000001|He is all over Jimmy's boots. I scream for him to run, but he just hit and hit that snake like he was crazy." +dev-clean/2035/147961/2035_147961_000011_000002|He grew more and more excited, and kept pointing all around his bed, as if there were things there and he wanted mr Shimerda to see them. +dev-clean/2035/147961/2035_147961_000025_000002|Beside a frozen pond something happened to the other sledge; peter saw it plainly. +dev-clean/2035/152373/2035_152373_000010_000007|saint Aidan, the Apostle of Northumbria, had refused the late Egfrid's father absolution, on one occasion, until he solemnly promised to restore their freedom to certain captives of this description. +dev-clean/2086/149214/2086_149214_000005_000002|It is a legend prolonging itself, from an epoch now gray in the distance, down into our own broad daylight, and bringing along with it some of its legendary mist, which the reader, according to his pleasure, may either disregard, or allow it to float almost imperceptibly about the characters and events for the sake of a picturesque effect. +dev-clean/2086/149220/2086_149220_000016_000003|In short, I make pictures out of sunshine; and, not to be too much dazzled with my own trade, I have prevailed with Miss Hepzibah to let me lodge in one of these dusky gables. +dev-clean/2086/149220/2086_149220_000028_000000|Phoebe was on the point of retreating, but turned back, with some hesitation; for she did not exactly comprehend his manner, although, on better observation, its feature seemed rather to be lack of ceremony than any approach to offensive rudeness. +dev-clean/2277/149874/2277_149874_000007_000001|Her husband asked a few questions and sat down to read the evening paper. +dev-clean/2277/149896/2277_149896_000007_000006|He saw only her pretty face and neat figure and wondered why life was not arranged so that such joy as he found with her could be steadily maintained. +dev-clean/2277/149896/2277_149896_000025_000008|He jangled it fiercely several times in succession, but without avail. +dev-clean/2277/149897/2277_149897_000023_000000|"Well?" said Hurstwood. +dev-clean/2277/149897/2277_149897_000046_000002|He troubled over many little details and talked perfunctorily to everybody. +dev-clean/2412/153954/2412_153954_000004_000005|Even in middle age they were still comely, and the old grey haired women at their cottage doors had a dignity, not to say majesty, of their own. +dev-clean/2428/83699/2428_83699_000009_000000|Now it is a remarkable thing that I have always had an extraordinary predilection for the name Madge. +dev-clean/2428/83699/2428_83699_000024_000004|I had long been wishing that an old-fashioned Christmas had been completely extinct before I had thought of adventuring in quest of one. +dev-clean/2428/83699/2428_83699_000047_000000|"Perhaps you had better come inside." +dev-clean/2428/83705/2428_83705_000015_000004|I did not want any unpleasantness; and I am quite sure there would have been unpleasantness had I demurred. +dev-clean/2428/83705/2428_83705_000034_000002|"And what," inquired mrs Macpherson, "has Mary Ann given you?" +dev-clean/251/118436/251_118436_000017_000001|This man was clad in a brown camel hair robe and sandals, and a green turban was on his head. His expression was tranquil, his gaze impersonal. +dev-clean/251/136532/251_136532_000000_000003|Fitzgerald was still trying to find out how the germ had been transmitted. +dev-clean/251/136532/251_136532_000020_000004|Without question, he had become, overnight, the most widely known archaeologist in history. +dev-clean/251/137823/251_137823_000025_000001|Or grazed, at least," Tom added thankfully. +dev-clean/251/137823/251_137823_000054_000002|The two girls were as much upset as Tom's mother. +dev-clean/2803/154320/2803_154320_000017_000004|Think of Lady Glenarvan; think of Mary Grant!" +dev-clean/2803/154328/2803_154328_000028_000000|Wilson and Olbinett joined their companions, and all united to dig through the wall-john with his dagger, the others with stones taken from the ground, or with their nails, while Mulrady, stretched along the ground, watched the native guard through a crevice of the matting. +dev-clean/2803/154328/2803_154328_000080_000003|Where chance led them, but at any rate they were free. +dev-clean/2803/161169/2803_161169_000011_000019|What do you think of that from the coal tar. +dev-clean/2902/9008/2902_9008_000009_000001|He was a Greek, also, but of a more common, and, perhaps, lower type; dark and fiery, thin and graceful; his delicate figure and cheeks, wasted by meditation, harmonised well with the staid and simple philosophic cloak which he wore as a sign of his profession. +dev-clean/2902/9008/2902_9008_000048_000003|For aught I know or care, the plot may be an exactly opposite one, and the Christians intend to murder all the Jews. +dev-clean/3000/15664/3000_15664_000013_000004|These volcanic caves are not wanting in interest, and it is well to light a pitch pine torch and take a walk in these dark ways of the underworld whenever opportunity offers, if for no other reason to see with new appreciation on returning to the sunshine the beauties that lie so thick about us. +dev-clean/3000/15664/3000_15664_000029_000002|Thus the Shasta River issues from a large lake like spring in Shasta Valley, and about two thirds of the volume of the McCloud gushes forth in a grand spring on the east side of the mountain, a few miles back from its immediate base. +dev-clean/3170/137482/3170_137482_000010_000004|The nobility, the merchants, even workmen in good circumstances, are never seen in the 'magazzino', for cleanliness is not exactly worshipped in such places. +dev-clean/3170/137482/3170_137482_000037_000001|He was celebrated in Venice not only for his eloquence and his great talents as a statesman, but also for the gallantries of his youth. +dev-clean/3536/23268/3536_23268_000028_000000|"It is not the first time, I believe, you have acted contrary to that, Miss Milner," replied mrs Horton, and affected a tenderness of voice, to soften the harshness of her words. +dev-clean/3576/138058/3576_138058_000019_000003|He wondered to see the lance leaning against the tree, the shield on the ground, and Don Quixote in armour and dejected, with the saddest and most melancholy face that sadness itself could produce; and going up to him he said, "Be not so cast down, good man, for you have not fallen into the hands of any inhuman Busiris, but into Roque Guinart's, which are more merciful than cruel." +dev-clean/3752/4943/3752_4943_000026_000002|Lie quiet!" +dev-clean/3752/4943/3752_4943_000056_000002|His flogging wouldn't have killed a flea." +dev-clean/3752/4944/3752_4944_000031_000000|"Well now!" said Meekin, with asperity, "I don't agree with you. Everybody seems to be against that poor fellow-Captain Frere tried to make me think that his letters contained a hidden meaning, but I don't believe they did. +dev-clean/3752/4944/3752_4944_000063_000003|He'd rather kill himself." +dev-clean/3752/4944/3752_4944_000094_000000|"The Government may go to----, and you, too!" roared Burgess. +dev-clean/3853/163249/3853_163249_000058_000000|"I've done it, mother: tell me you're not sorry." +dev-clean/3853/163249/3853_163249_000125_000004|Help me to be brave and strong, David: don't let me complain or regret, but show me what lies beyond, and teach me to believe that simply doing the right is reward and happiness enough." +dev-clean/5338/24615/5338_24615_000004_000003|It had been built at a period when castles were no longer necessary, and when the Scottish architects had not yet acquired the art of designing a domestic residence. +dev-clean/5338/284437/5338_284437_000031_000001|A powerful ruler ought to be rich and to live in a splendid palace. +dev-clean/5536/43358/5536_43358_000012_000001|Being a natural man, the Indian was intensely poetical. +dev-clean/5536/43359/5536_43359_000015_000000|The family was not only the social unit, but also the unit of government. +dev-clean/5694/64025/5694_64025_000004_000006|Our regiment was the advance guard on Saturday evening, and did a little skirmishing; but General Gladden's brigade passed us and assumed a position in our immediate front. +dev-clean/5694/64029/5694_64029_000006_000005|I read it, and looked up to hand it back to him, when I discovered that he had a pistol cocked and leveled in my face, and says he, "Drop that gun; you are my prisoner." I saw there was no use in fooling about it. +dev-clean/5694/64029/5694_64029_000024_000002|The ground was literally covered with blue coats dead; and, if I remember correctly, there were eighty dead horses. +dev-clean/5694/64038/5694_64038_000015_000002|I could not imagine what had become of him. +dev-clean/5895/34615/5895_34615_000013_000003|Man can do nothing to create beauty, but everything to produce ugliness. +dev-clean/5895/34615/5895_34615_000025_000000|With this exception, Gwynplaine's laugh was everlasting. +dev-clean/5895/34622/5895_34622_000029_000002|In the opposite corner was the kitchen. +dev-clean/5895/34629/5895_34629_000021_000005|The sea is a wall; and if Voltaire-a thing which he very much regretted when it was too late-had not thrown a bridge over to Shakespeare, Shakespeare might still be in England, on the other side of the wall, a captive in insular glory. +dev-clean/6241/61943/6241_61943_000020_000000|My uncle came out of his cabin pale, haggard, thin, but full of enthusiasm, his eyes dilated with pleasure and satisfaction. +dev-clean/6241/61946/6241_61946_000014_000000|The rugged summits of the rocky hills were dimly visible on the edge of the horizon, through the misty fogs; every now and then some heavy flakes of snow showed conspicuous in the morning light, while certain lofty and pointed rocks were first lost in the grey low clouds, their summits clearly visible above, like jagged reefs rising from a troublous sea. +dev-clean/6241/61946/6241_61946_000051_000001|Then my uncle, myself, and guide, two boatmen and the four horses got into a very awkward flat bottom boat. +dev-clean/6295/64301/6295_64301_000010_000002|The music was broken, and Joseph left alone with the dumb instruments. +dev-clean/6313/66125/6313_66125_000020_000002|"Are you hurt?" +dev-clean/6313/66125/6313_66125_000053_000000|"Are you ready?" +dev-clean/6313/66129/6313_66129_000011_000001|"Cold water is the most nourishing thing we've touched since last night." +dev-clean/6313/66129/6313_66129_000045_000004|Of course, dogs can't follow the trail of an animal as well, now, as they could with snow on the ground. +dev-clean/6313/66129/6313_66129_000081_000000|Stacy dismounted and removed the hat carefully to one side. +dev-clean/6313/76958/6313_76958_000029_000000|Instantly there was a chorus of yells and snarls from the disturbed cowpunchers, accompanied by dire threats as to what they would do to the gopher did he ever disturb their rest in that way again. +dev-clean/6313/76958/6313_76958_000073_000001|"Those fellows have to go out. +dev-clean/6319/275224/6319_275224_000014_000001|And what is the matter with the beautiful straggling branches, that they are to be cut off as fast as they appear? +dev-clean/6319/57405/6319_57405_000019_000000|"It is rather a silly thing to do," said Deucalion; "and yet there can be no harm in it, and we shall see what will happen." +dev-clean/6319/64726/6319_64726_000017_000002|Then the prince took the princess by the hand; she was dressed in great splendour, but he did not hint that she looked as he had seen pictures of his great grandmother look; he thought her all the more charming for that. +dev-clean/6345/93302/6345_93302_000000_000001|All LibriVox recordings are in the public domain. +dev-clean/6345/93302/6345_93302_000049_000000|The fine tact of a noble woman seemed to have deserted her. +dev-clean/6345/93302/6345_93302_000073_000000|So she said- +dev-clean/6345/93306/6345_93306_000024_000002|What is it? +dev-clean/652/130737/652_130737_000031_000001|Good aroma. +dev-clean/7850/111771/7850_111771_000009_000001|After various flanking movements and costly assaults, the problem of taking Lee narrowed itself down to a siege of Petersburg. +dev-clean/7850/281318/7850_281318_000012_000000|She began to show them how to weave the bits of things together into nests, as they should be made. +dev-clean/7850/286674/7850_286674_000006_000001|You would think that, with six legs apiece and three joints in each leg, they might walk quite fast, yet they never did. +dev-clean/7850/73752/7850_73752_000006_000003|What a Neapolitan ball was his career then! +dev-clean/7976/105575/7976_105575_000009_000000|The burying party the next morning found nineteen dead Rebels lying together at one place. +dev-clean/7976/105575/7976_105575_000017_000000|Our regiment now pursued the flying Rebels with great vigor. +dev-clean/7976/110124/7976_110124_000021_000001|"We two are older and wiser than you are. It is for us to determine what shall be done. +dev-clean/7976/110124/7976_110124_000053_000002|The doors were strong and held securely. +dev-clean/7976/110523/7976_110523_000027_000000|"We will go in here," said Hansel, "and have a glorious feast. +dev-clean/8297/275154/8297_275154_000008_000000|Was this man-haggard, pallid, shabby, looking at him piteously with bloodshot eyes-the handsome, pleasant, prosperous brother whom he remembered? +dev-clean/8297/275154/8297_275154_000024_000011|Tell me where my wife is living now?" +dev-clean/8297/275155/8297_275155_000013_000006|What a perfect gentleman!" +dev-clean/8297/275155/8297_275155_000037_000000|"Say thoroughly worthy of the course forced upon me and my daughter by your brother's infamous conduct-and you will be nearer the mark!" +dev-clean/8297/275156/8297_275156_000013_000005|No more of it now. +dev-clean/84/121123/84_121123_000009_000000|But in less than five minutes the staircase groaned beneath an extraordinary weight. +dev-clean/84/121123/84_121123_000054_000000|It was something terrible to witness the silent agony, the mute despair of Noirtier, whose tears silently rolled down his cheeks. +dev-clean/84/121550/84_121550_000064_000000|And lo! a sudden lustre ran across On every side athwart the spacious forest, Such that it made me doubt if it were lightning. +dev-clean/84/121550/84_121550_000156_000000|Nor prayer for inspiration me availed, By means of which in dreams and otherwise I called him back, so little did he heed them. +dev-clean/84/121550/84_121550_000247_000000|Thus Beatrice; and I, who at the feet Of her commandments all devoted was, My mind and eyes directed where she willed. +dev-clean/8842/302203/8842_302203_000001_000001|And I remember that on the ninth day, being overcome with intolerable pain, a thought came into my mind concerning my lady: but when it had a little nourished this thought, my mind returned to its brooding over mine enfeebled body. diff --git a/BigVGAN/filelists/LibriTTS/dev-other.txt b/BigVGAN/filelists/LibriTTS/dev-other.txt new file mode 100644 index 0000000000000000000000000000000000000000..9952353bccd0d62f22da2c409ba6764df09bc005 --- /dev/null +++ b/BigVGAN/filelists/LibriTTS/dev-other.txt @@ -0,0 +1,93 @@ +dev-other/116/288045/116_288045_000003_000000|PART one +dev-other/116/288045/116_288045_000034_000001|He was only an idol. +dev-other/116/288047/116_288047_000002_000002|Observing the sun, the moon, and the stars overhead, the primitive man wished to account for them. +dev-other/116/288048/116_288048_000001_000000|Let me now give an idea of the method I propose to follow in the study of this subject. +dev-other/116/288048/116_288048_000020_000003|Leaving out Judas, and counting Matthias, who was elected in his place, we have thirteen apostles. +dev-other/1255/138279/1255_138279_000012_000000|"One. +dev-other/1255/138279/1255_138279_000049_000001|Will it be by banns or license?" +dev-other/1255/74899/1255_74899_000020_000000|"Pardon me. +dev-other/1255/90407/1255_90407_000006_000001|But, as the rain gave not the least sign of cessation, he observed: 'I think we shall have to go back.' +dev-other/1255/90407/1255_90407_000039_000002|Into it they plodded without pause, crossing the harbour bridge about midnight, wet to the skin. +dev-other/1255/90413/1255_90413_000023_000001|'Now what the devil this means I cannot tell,' he said to himself, reflecting stock still for a moment on the stairs. +dev-other/1585/131718/1585_131718_000025_000009|Edison marginalized documents extensively. +dev-other/1630/141772/1630_141772_000000_000002|Suddenly he again felt that he was alive and suffering from a burning, lacerating pain in his head. +dev-other/1630/141772/1630_141772_000039_000000|The quiet home life and peaceful happiness of Bald Hills presented itself to him. +dev-other/1630/73710/1630_73710_000019_000003|I almost wish papa would return, though I dread to see him. +dev-other/1630/96099/1630_96099_000033_000001|Why did you not follow him? +dev-other/1650/157641/1650_157641_000037_000001|mr w m +dev-other/1650/173551/1650_173551_000025_000000|Pierre went into that gloomy study which he had entered with such trepidation in his benefactor's lifetime. +dev-other/1651/136854/1651_136854_000046_000005|I have, however, this of gratitude, that I think of you with regard, when I do not, perhaps, give the proofs which I ought, of being, Sir, +dev-other/1686/142278/1686_142278_000015_000000|'No! not doubts as to religion; not the slightest injury to that.' He paused. +dev-other/1686/142278/1686_142278_000042_000001|Margaret was nearly upset again into a burst of crying. +dev-other/1701/141759/1701_141759_000001_000001|Not till midwinter was the count at last handed a letter addressed in his son's handwriting. +dev-other/1701/141759/1701_141759_000048_000000|"Why should you be ashamed?" +dev-other/1701/141760/1701_141760_000013_000003|"I only sent you the note yesterday by Bolkonski-an adjutant of Kutuzov's, who's a friend of mine. +dev-other/1701/141760/1701_141760_000056_000000|In spite of Prince Andrew's disagreeable, ironical tone, in spite of the contempt with which Rostov, from his fighting army point of view, regarded all these little adjutants on the staff of whom the newcomer was evidently one, Rostov felt confused, blushed, and became silent. +dev-other/2506/13150/2506_13150_000022_000000|--Nay-if you don't believe me, you may read the chapter for your pains. +dev-other/3660/172182/3660_172182_000012_000007|And a year, and a second, and a third, he proceeded thus, until his fame had flown over the face of the kingdom. +dev-other/3660/172183/3660_172183_000011_000000|So the maiden went forward, keeping in advance of Geraint, as he had desired her; and it grieved him as much as his wrath would permit, to see a maiden so illustrious as she having so much trouble with the care of the horses. +dev-other/3660/172183/3660_172183_000019_000040|Come with me to the court of a son in law of my sister, which is near here, and thou shalt have the best medical assistance in the kingdom." +dev-other/3660/6517/3660_6517_000036_000002|Bright sunshine. +dev-other/3660/6517/3660_6517_000059_000005|Not a single one has lost his good spirits. +dev-other/3663/172005/3663_172005_000022_000000|She must cross the Slide Brook valley, if possible, and gain the mountain opposite. +dev-other/3663/172528/3663_172528_000016_000008|He had been brought by my very dear friend Luca Martini, who passed the larger portion of the day with me. +dev-other/3915/57461/3915_57461_000018_000001|In a fit of madness I was tempted to kill and rob you. +dev-other/3915/98647/3915_98647_000018_000006|Thus the old custom is passing away. +dev-other/4323/13259/4323_13259_000009_000011|What would Jesus do? +dev-other/4323/13259/4323_13259_000020_000003|It seems she had been recently converted during the evangelist's meetings, and was killed while returning from one of the meetings in company with other converts and some of her friends. +dev-other/4323/18416/4323_18416_000019_000001|So she was asked to sing at musicales and receptions without end, until Alexia exclaimed at last, "They are all raving, stark mad over her, and it's all Polly's own fault, the whole of it." +dev-other/4323/18416/4323_18416_000050_000000|"I know, child; you think your old Grandpapa does just about right," said mr King soothingly, and highly gratified. +dev-other/4323/18416/4323_18416_000079_000002|"And I can't tolerate any thoughts I cannot speak." +dev-other/4323/55228/4323_55228_000028_000000|"Pete told you that I didn't care for any girl, only to paint?" demanded Bertram, angry and mystified. +dev-other/4323/55228/4323_55228_000071_000000|There was another silence. +dev-other/4570/102353/4570_102353_000001_000000|CHAPTER four. +dev-other/4570/14911/4570_14911_000009_000002|EYES-Brown, dark hazel or hazel, not deep set nor bulgy, and with a mild expression. +dev-other/4570/56594/4570_56594_000012_000000|"'No,' says the gentleman. +dev-other/4831/18525/4831_18525_000028_000000|"Oh! isn't it 'Oats, Peas, Beans, and Barley grow'?" cried Polly, as they watched them intently. +dev-other/4831/18525/4831_18525_000078_000001|"I want to write, too, I do," she cried, very much excited. +dev-other/4831/18525/4831_18525_000122_000000|"O dear me!" exclaimed Polly, softly, for she couldn't even yet get over that dreadful beginning. +dev-other/4831/25894/4831_25894_000022_000003|The other days were very much like this; sometimes they made more, sometimes less, but Tommo always 'went halves;' and Tessa kept on, in spite of cold and weariness, for her plans grew as her earnings increased, and now she hoped to get useful things, instead of candy and toys alone. +dev-other/4831/29134/4831_29134_000001_000000|The session was drawing toward its close. +dev-other/4831/29134/4831_29134_000018_000000|"So this poor little boy grew up to be a man, and had to go out in the world, far from home and friends to earn his living. +dev-other/5543/27761/5543_27761_000019_000000|Her mother went to hide. +dev-other/5543/27761/5543_27761_000065_000000|"Agathya says so, madam," answered Fedosya; "it's she that knows." +dev-other/5543/27761/5543_27761_000107_000000|"Sima, my dear, don't agitate yourself," said Sergey Modestovich in a whisper. +dev-other/5849/50873/5849_50873_000026_000000|"He has promised to do so." +dev-other/5849/50873/5849_50873_000074_000000|"The boy did it! +dev-other/5849/50962/5849_50962_000010_000000|"It's a schooner," said mr Bingham to mr Minturn, "and she has a very heavy cargo." +dev-other/5849/50963/5849_50963_000009_000003|Well, it was a long, slow job to drag those heavy logs around that point, and just when we were making headway, along comes a storm that drove the schooner and canoes out of business." +dev-other/5849/50964/5849_50964_000018_000001|There were the shells to be looked after, the fish nets, besides Downy, the duck, and Snoop, the cat. +dev-other/6123/59150/6123_59150_000016_000001|He kicked him two or three times with his heel in the face. +dev-other/6123/59186/6123_59186_000008_000000|"Catering care" is an appalling phrase. +dev-other/6267/53049/6267_53049_000007_000001|"I'd better be putting my grey matter into that algebra instead of wasting it plotting for a party dress that I certainly can't get. +dev-other/6267/53049/6267_53049_000045_000001|I am named after her." +dev-other/6267/65525/6267_65525_000018_000000|Dear mr Lincoln: +dev-other/6267/65525/6267_65525_000045_000006|You can't mistake it." +dev-other/6455/66379/6455_66379_000020_000002|(Deal, sir, if you please; better luck next time.)" +dev-other/6455/67803/6455_67803_000038_000000|"Yes," he answered. +dev-other/6467/56885/6467_56885_000012_000001|As you are so generously taking her on trust, may she never cause you a moment's regret. +dev-other/6467/97061/6467_97061_000010_000000|A terrible battle ensued, in which both kings performed prodigies of valour. +dev-other/6841/88291/6841_88291_000006_000006|One stood waiting for them to finish, a sheaf of long j h stamping irons in his hand. +dev-other/6841/88291/6841_88291_000019_000006|Cries arose in a confusion: "Marker" "Hot iron!" "Tally one!" Dust eddied and dissipated. +dev-other/6841/88294/6841_88294_000010_000003|Usually I didn't bother with his talk, for it didn't mean anything, but something in his voice made me turn. +dev-other/6841/88294/6841_88294_000048_000000|He stood there looking straight at me without winking or offering to move. +dev-other/700/122866/700_122866_000006_000003|You've been thirteen for a month, so I suppose it doesn't seem such a novelty to you as it does to me. +dev-other/700/122866/700_122866_000023_000006|Ruby Gillis is rather sentimental. +dev-other/700/122867/700_122867_000012_000004|My career is closed. +dev-other/700/122867/700_122867_000033_000003|At the end of the week Marilla said decidedly: +dev-other/700/122868/700_122868_000015_000003|mrs Lynde says that all play acting is abominably wicked." +dev-other/700/122868/700_122868_000038_000001|And Ruby is in hysterics-oh, Anne, how did you escape?" +dev-other/7601/101622/7601_101622_000018_000002|The very girls themselves set them on: +dev-other/7601/175351/7601_175351_000031_000008|Still, during the nights which followed the fifteenth of August, darkness was never profound; although the sun set, he still gave sufficient light by refraction. +dev-other/7641/96252/7641_96252_000003_000006|For these are careful only for themselves, for their own egoism, just like the bandit, from whom they are only distinguished by the absurdity of their means. +dev-other/7641/96670/7641_96670_000013_000001|The mist lifted suddenly and she saw three strangers in the palace courtyard. +dev-other/7641/96684/7641_96684_000009_000000|"What years of happiness have been mine, O Apollo, through your friendship for me," said Admetus. +dev-other/7641/96684/7641_96684_000031_000002|How noble it was of Admetus to bring him into his house and give entertainment to him while such sorrow was upon him. +dev-other/7697/105815/7697_105815_000048_000002|And they brought out the jaw bone of an ass with which Samson did such great feats, and the sling and stone with which David slew Goliath of Gath. +dev-other/8173/294714/8173_294714_000006_000001|"Don't spoil my pleasure in seeing you again by speaking of what can never be! Have you still to be told how it is that you find me here alone with my child?" +dev-other/8173/294714/8173_294714_000027_000001|What was there to prevent her from insuring her life, if she pleased, and from so disposing of the insurance as to give Van Brandt a direct interest in her death? +dev-other/8254/115543/8254_115543_000034_000000|"Yes, and how he orders every one about him. +dev-other/8254/84205/8254_84205_000029_000000|"I'm not afraid of them hitting me, my lad," said Griggs confidently. "Being shot at by fellows with bows and arrows sounds bad enough, but there's not much risk here." +dev-other/8254/84205/8254_84205_000073_000000|"Right; I do, neighbour, and it's very handsome of you to offer me the chance to back out. +dev-other/8288/274162/8288_274162_000023_000000|"Exactly. +dev-other/8288/274162/8288_274162_000078_000000|"So much the worse. diff --git a/BigVGAN/filelists/LibriTTS/parse_libritts.py b/BigVGAN/filelists/LibriTTS/parse_libritts.py new file mode 100644 index 0000000000000000000000000000000000000000..dc1d4d3bd7383b38903055b1a99ecf4d784a39f5 --- /dev/null +++ b/BigVGAN/filelists/LibriTTS/parse_libritts.py @@ -0,0 +1,77 @@ +# Copyright (c) 2024 NVIDIA CORPORATION. +# Licensed under the MIT license. + +import os, glob + + +def get_wav_and_text_filelist(data_root, data_type, subsample=1): + wav_list = sorted( + [ + path.replace(data_root, "")[1:] + for path in glob.glob(os.path.join(data_root, data_type, "**/**/*.wav")) + ] + ) + wav_list = wav_list[::subsample] + txt_filelist = [path.replace(".wav", ".normalized.txt") for path in wav_list] + + txt_list = [] + for txt_file in txt_filelist: + with open(os.path.join(data_root, txt_file), "r") as f_txt: + text = f_txt.readline().strip("\n") + txt_list.append(text) + wav_list = [path.replace(".wav", "") for path in wav_list] + + return wav_list, txt_list + + +def write_filelist(output_path, wav_list, txt_list): + with open(output_path, "w") as f: + for i in range(len(wav_list)): + filename = wav_list[i] + "|" + txt_list[i] + f.write(filename + "\n") + + +if __name__ == "__main__": + + data_root = "filelists/LibriTTS" + + # Dev and test sets. subsample each sets to get ~100 utterances + data_type_list = ["dev-clean", "dev-other", "test-clean", "test-other"] + subsample_list = [50, 50, 50, 50] + for data_type, subsample in zip(data_type_list, subsample_list): + print(f"processing {data_type}") + data_path = os.path.join(data_root, data_type) + assert os.path.exists(data_path), ( + f"path {data_path} not found. make sure the path is accessible by creating the symbolic link using the following command: " + f"ln -s /path/to/your/{data_path} {data_path}" + ) + wav_list, txt_list = get_wav_and_text_filelist(data_root, data_type, subsample) + write_filelist(os.path.join(data_root, data_type + ".txt"), wav_list, txt_list) + + # Training and seen speaker validation datasets (libritts-full): train-clean-100 + train-clean-360 + train-other-500 + wav_list_train, txt_list_train = [], [] + for data_type in ["train-clean-100", "train-clean-360", "train-other-500"]: + print(f"processing {data_type}") + data_path = os.path.join(data_root, data_type) + assert os.path.exists(data_path), ( + f"path {data_path} not found. make sure the path is accessible by creating the symbolic link using the following command: " + f"ln -s /path/to/your/{data_path} {data_path}" + ) + wav_list, txt_list = get_wav_and_text_filelist(data_root, data_type) + wav_list_train.extend(wav_list) + txt_list_train.extend(txt_list) + + # Split the training set so that the seen speaker validation set contains ~100 utterances + subsample_val = 3000 + wav_list_val, txt_list_val = ( + wav_list_train[::subsample_val], + txt_list_train[::subsample_val], + ) + del wav_list_train[::subsample_val] + del txt_list_train[::subsample_val] + write_filelist( + os.path.join(data_root, "train-full.txt"), wav_list_train, txt_list_train + ) + write_filelist(os.path.join(data_root, "val-full.txt"), wav_list_val, txt_list_val) + + print("done") diff --git a/BigVGAN/filelists/LibriTTS/test-clean.txt b/BigVGAN/filelists/LibriTTS/test-clean.txt new file mode 100644 index 0000000000000000000000000000000000000000..5bbfab4da31088b3f42dcf2e15238e046d1781dc --- /dev/null +++ b/BigVGAN/filelists/LibriTTS/test-clean.txt @@ -0,0 +1,97 @@ +test-clean/1089/134686/1089_134686_000001_000001|He hoped there would be stew for dinner, turnips and carrots and bruised potatoes and fat mutton pieces to be ladled out in thick peppered flour fattened sauce. Stuff it into you, his belly counselled him. +test-clean/1089/134686/1089_134686_000020_000001|We can scut the whole hour. +test-clean/1089/134691/1089_134691_000004_000001|Yet her mistrust pricked him more keenly than his father's pride and he thought coldly how he had watched the faith which was fading down in his soul ageing and strengthening in her eyes. +test-clean/1089/134691/1089_134691_000027_000004|Now, at the name of the fabulous artificer, he seemed to hear the noise of dim waves and to see a winged form flying above the waves and slowly climbing the air. +test-clean/1188/133604/1188_133604_000018_000002|There are just four touches-fine as the finest penmanship-to do that beak; and yet you will find that in the peculiar paroquettish mumbling and nibbling action of it, and all the character in which this nibbling beak differs from the tearing beak of the eagle, it is impossible to go farther or be more precise. +test-clean/121/121726/121_121726_000046_000003|Tied to a woman. +test-clean/121/127105/121_127105_000024_000000|He laughed for the first time. +test-clean/1284/1180/1284_1180_000001_000000|The Crooked Magician +test-clean/1284/1181/1284_1181_000005_000000|The head of the Patchwork Girl was the most curious part of her. +test-clean/1320/122612/1320_122612_000019_000005|It is true that the horses are here, but the Hurons are gone; let us, then, hunt for the path by which they parted." +test-clean/1320/122612/1320_122612_000056_000002|Then he reappeared, creeping along the earth, from which his dress was hardly distinguishable, directly in the rear of his intended captive. +test-clean/1580/141083/1580_141083_000012_000000|"The first page on the floor, the second in the window, the third where you left it," said he. +test-clean/1580/141083/1580_141083_000041_000003|Above were three students, one on each story. +test-clean/1580/141083/1580_141083_000063_000001|Holmes held it out on his open palm in the glare of the electric light. +test-clean/1580/141083/1580_141083_000110_000001|Where were you when you began to feel bad?" +test-clean/1580/141084/1580_141084_000024_000002|Pencils, too, and knives-all was satisfactory. +test-clean/1580/141084/1580_141084_000060_000001|"I frankly admit that I am unable to prove it. +test-clean/1580/141084/1580_141084_000085_000000|"Good heavens! have you nothing to add?" cried Soames. +test-clean/1995/1826/1995_1826_000022_000001|Miss Taylor did not know much about cotton, but at least one more remark seemed called for. +test-clean/1995/1836/1995_1836_000016_000001|No, of course there was no immediate danger; but when people were suddenly thrust beyond their natural station, filled with wild ideas and impossible ambitions, it meant terrible danger to Southern white women. +test-clean/1995/1837/1995_1837_000024_000000|He heard that she was down stairs and ran to meet her with beating heart. +test-clean/2300/131720/2300_131720_000016_000005|Having travelled around the world, I had cultivated an indifference to any special difficulties of that kind. +test-clean/2300/131720/2300_131720_000030_000005|I telephoned again, and felt something would happen, but fortunately it did not. +test-clean/237/126133/237_126133_000002_000004|It got to be noticed finally; and one and all redoubled their exertions to make everything twice as pleasant as ever! +test-clean/237/126133/237_126133_000049_000000|But the chubby face didn't look up brightly, as usual: and the next moment, without a bit of warning, Phronsie sprang past them all, even Polly, and flung herself into mr King's arms, in a perfect torrent of sobs. +test-clean/237/134493/237_134493_000008_000003|Alexandra lets you sleep late. +test-clean/237/134500/237_134500_000001_000001|Frank sat up until a late hour reading the Sunday newspapers. +test-clean/237/134500/237_134500_000014_000000|"I don't know all of them, but I know lindens are. +test-clean/237/134500/237_134500_000034_000000|She sighed despondently. +test-clean/260/123286/260_123286_000019_000002|Therefore don't talk to me about views and prospects." +test-clean/260/123286/260_123286_000049_000005|He shakes his head negatively. +test-clean/260/123288/260_123288_000016_000002|It rushes on from the farthest recesses of the vast cavern. +test-clean/260/123288/260_123288_000043_000001|I could just see my uncle at full length on the raft, and Hans still at his helm and spitting fire under the action of the electricity which has saturated him. +test-clean/2830/3979/2830_3979_000007_000000|PREFACE +test-clean/2830/3980/2830_3980_000018_000001|Humble man that he was, he will not now take a back seat. +test-clean/2961/961/2961_961_000004_000037|Then your city did bravely, and won renown over the whole earth. +test-clean/2961/961/2961_961_000023_000003|But violent as were the internal and alimentary fluids, the tide became still more violent when the body came into contact with flaming fire, or the solid earth, or gliding waters, or the stormy wind; the motions produced by these impulses pass through the body to the soul and have the name of sensations. +test-clean/3570/5694/3570_5694_000009_000003|The canon of reputability is at hand and seizes upon such innovations as are, according to its standard, fit to survive. +test-clean/3570/5695/3570_5695_000001_000003|But the middle class wife still carries on the business of vicarious leisure, for the good name of the household and its master. +test-clean/3570/5695/3570_5695_000009_000005|Considered by itself simply-taken in the first degree-this added provocation to which the artisan and the urban laboring classes are exposed may not very seriously decrease the amount of savings; but in its cumulative action, through raising the standard of decent expenditure, its deterrent effect on the tendency to save cannot but be very great. +test-clean/3570/5696/3570_5696_000011_000006|For this is the basis of award of the instinct of workmanship, and that instinct is the court of final appeal in any question of economic truth or adequacy. +test-clean/3729/6852/3729_6852_000004_000003|In order to please her, I spoke to her of the Abbe Conti, and I had occasion to quote two lines of that profound writer. +test-clean/4077/13754/4077_13754_000002_000000|The troops, once in Utah, had to be provisioned; and everything the settlers could spare was eagerly bought at an unusual price. The gold changed hands. +test-clean/4446/2271/4446_2271_000003_000004|There's everything in seeing Hilda while she's fresh in a part. +test-clean/4446/2271/4446_2271_000020_000001|Lady Westmere is very fond of Hilda." +test-clean/4446/2273/4446_2273_000008_000002|I've no need for fine clothes in Mac's play this time, so I can afford a few duddies for myself. +test-clean/4446/2273/4446_2273_000027_000004|She did my blouses beautifully the last time I was there, and was so delighted to see me again. +test-clean/4446/2273/4446_2273_000046_000001|"Aren't you afraid to let the wind low like that on your neck? +test-clean/4446/2275/4446_2275_000013_000000|Hilda was pale by this time, and her eyes were wide with fright. +test-clean/4446/2275/4446_2275_000038_000006|"You want to tell me that you can only see me like this, as old friends do, or out in the world among people? +test-clean/4507/16021/4507_16021_000011_000000|It engenders a whole world, la pegre, for which read theft, and a hell, la pegrenne, for which read hunger. +test-clean/4507/16021/4507_16021_000030_000001|Facts form one of these, and ideas the other. +test-clean/4970/29093/4970_29093_000010_000000|Delightful illusion of paint and tinsel and silk attire, of cheap sentiment and high and mighty dialogue! +test-clean/4970/29093/4970_29093_000047_000000|"Never mind the map. +test-clean/4970/29095/4970_29095_000021_000000|"I will practice it." +test-clean/4970/29095/4970_29095_000055_000002|He took it with him from the Southern Hotel, when he went to walk, and read it over and again in an unfrequented street as he stumbled along. +test-clean/4992/41797/4992_41797_000014_000002|He keeps the thou shalt not commandments first rate, Hen Lord does! +test-clean/4992/41806/4992_41806_000020_000001|Thou who settest the solitary in families, bless the life that is sheltered here. +test-clean/5105/28241/5105_28241_000004_000004|The late astounding events, however, had rendered Procope manifestly uneasy, and not the less so from his consciousness that the count secretly partook of his own anxiety. +test-clean/5142/33396/5142_33396_000004_000004|At the prow I carved the head with open mouth and forked tongue thrust out. +test-clean/5142/33396/5142_33396_000039_000000|"The thralls were bringing in a great pot of meat. +test-clean/5142/36377/5142_36377_000013_000003|I liked Naomi Colebrook at first sight; liked her pleasant smile; liked her hearty shake of the hand when we were presented to each other. +test-clean/5639/40744/5639_40744_000003_000006|Mother! dear father! do you hear me? +test-clean/5639/40744/5639_40744_000022_000000|Just then Leocadia came to herself, and embracing the cross seemed changed into a sea of tears, and the gentleman remained in utter bewilderment, until his wife had repeated to him, from beginning to end, Leocadia's whole story; and he believed it, through the blessed dispensation of Heaven, which had confirmed it by so many convincing testimonies. +test-clean/5683/32865/5683_32865_000018_000000|Well, it was pretty-French, I dare say-a little set of tablets-a toy-the cover of enamel, studded in small jewels, with a slender border of symbolic flowers, and with a heart in the centre, a mosaic of little carbuncles, rubies, and other red and crimson stones, placed with a view to light and shade. +test-clean/5683/32866/5683_32866_000005_000000|'Did you see that?' said Wylder in my ear, with a chuckle; and, wagging his head, he added, rather loftily for him, 'Miss Brandon, I reckon, has taken your measure, Master Stanley, as well as i I wonder what the deuce the old dowager sees in him. +test-clean/5683/32866/5683_32866_000047_000002|I was not a bit afraid of being found out. +test-clean/5683/32879/5683_32879_000036_000002|Be he near, or be he far, I regard his very name with horror.' +test-clean/6829/68769/6829_68769_000011_000000|So as soon as breakfast was over the next morning Beth and Kenneth took one of the automobiles, the boy consenting unwillingly to this sort of locomotion because it would save much time. +test-clean/6829/68769/6829_68769_000051_000001|One morning she tried to light the fire with kerosene, and lost her sight. +test-clean/6829/68769/6829_68769_000089_000001|Why should you do all this?" +test-clean/6829/68771/6829_68771_000018_000003|A speakers' stand, profusely decorated, had been erected on the lawn, and hundreds of folding chairs provided for seats. +test-clean/6930/75918/6930_75918_000000_000001|Night. +test-clean/6930/81414/6930_81414_000041_000001|Here is his scarf, which has evidently been strained, and on it are spots of blood, while all around are marks indicating a struggle. +test-clean/7021/79740/7021_79740_000010_000006|I observe that, when you both wish for the same thing, you don't quarrel for it and try to pull it away from one another; but one waits like a lady until the other has done with it. +test-clean/7021/85628/7021_85628_000017_000000|"I am going to the court ball," answered Anders. +test-clean/7127/75946/7127_75946_000022_000002|It is necessary, therefore, that he should comply." +test-clean/7127/75946/7127_75946_000061_000001|Disdainful of a success of which Madame showed no acknowledgement, he thought of nothing but boldly regaining the marked preference of the princess. +test-clean/7127/75947/7127_75947_000035_000000|"Quite true, and I believe you are right. +test-clean/7176/88083/7176_88083_000002_000003|He was too imposing in appearance, too gorgeous in apparel, too bold and vigilant in demeanor to be so misunderstood. +test-clean/7176/88083/7176_88083_000017_000000|Immediately over his outstretched gleaming head flew the hawk. +test-clean/7176/92135/7176_92135_000011_000000|And, so on in the same vein for some thirty lines. +test-clean/7176/92135/7176_92135_000074_000001|Tea, please, Matthews. +test-clean/7729/102255/7729_102255_000011_000003|The Free State Hotel served as barracks. +test-clean/7729/102255/7729_102255_000028_000009|They were squads of Kansas militia, companies of "peaceful emigrants," or gangs of irresponsible outlaws, to suit the chance, the whim, or the need of the moment. +test-clean/8230/279154/8230_279154_000003_000002|In the present lecture I shall attempt the analysis of memory knowledge, both as an introduction to the problem of knowledge in general, and because memory, in some form, is presupposed in almost all other knowledge. +test-clean/8230/279154/8230_279154_000013_000003|One of these is context. +test-clean/8230/279154/8230_279154_000027_000000|A further stage is RECOGNITION. +test-clean/8455/210777/8455_210777_000022_000003|And immediately on his sitting down, there got up a gentleman to whom I had not been introduced before this day, and gave the health of Mrs Neverbend and the ladies of Britannula. +test-clean/8455/210777/8455_210777_000064_000001|Government that he shall be treated with all respect, and that those honours shall be paid to him which are due to the President of a friendly republic. +test-clean/8463/287645/8463_287645_000023_000001|For instance, Jacob Taylor was noticed on the record book as being twenty three years of age, and the name of his master was entered as "William Pollit;" but as Jacob had never been allowed to learn to read, he might have failed in giving a correct pronunciation of the name. +test-clean/8463/294825/8463_294825_000048_000000|- CENTIMETER Roughly two fifths of an inch +test-clean/8463/294828/8463_294828_000046_000001|Conseil did them in a flash, and I was sure the lad hadn't missed a thing, because he classified shirts and suits as expertly as birds and mammals. +test-clean/8555/284447/8555_284447_000018_000002|The poor Queen, by the way, was seldom seen, as she passed all her time playing solitaire with a deck that was one card short, hoping that before she had lived her entire six hundred years she would win the game. +test-clean/8555/284447/8555_284447_000049_000000|Now, indeed, the Boolooroo was as angry as he was amazed. +test-clean/8555/284449/8555_284449_000039_000000|When the courtiers and the people assembled saw the goat they gave a great cheer, for the beast had helped to dethrone their wicked Ruler. +test-clean/8555/292519/8555_292519_000041_000000|She was alone that night. He had broken into her courtyard. Above the gurgling gutters he heard- surely- a door unchained? diff --git a/BigVGAN/filelists/LibriTTS/test-other.txt b/BigVGAN/filelists/LibriTTS/test-other.txt new file mode 100644 index 0000000000000000000000000000000000000000..930b528eb7a3e92597cd1d5ad3d2519143ad2051 --- /dev/null +++ b/BigVGAN/filelists/LibriTTS/test-other.txt @@ -0,0 +1,103 @@ +test-other/1688/142285/1688_142285_000000_000000|'Margaret!' said mr Hale, as he returned from showing his guest downstairs; 'I could not help watching your face with some anxiety, when mr Thornton made his confession of having been a shop boy. +test-other/1688/142285/1688_142285_000046_000000|'No, mamma; that Anne Buckley would never have done.' +test-other/1998/15444/1998_15444_000012_000000|Simple filtration will sometimes suffice to separate the required substance; in other cases dialysis will be necessary, in order that crystalloid substances may be separated from colloid bodies. +test-other/1998/29454/1998_29454_000021_000001|Fried eggs and bacon-he had one egg and the man had three-bread and butter-and if the bread was thick, so was the butter-and as many cups of tea as you liked to say thank you for. +test-other/1998/29454/1998_29454_000053_000001|It almost looked, Dickie thought, as though he had brought them out for some special purpose. +test-other/1998/29455/1998_29455_000022_000000|It was a wonderful day. +test-other/1998/29455/1998_29455_000082_000003|But 'e's never let it out." +test-other/2414/128292/2414_128292_000003_000000|"What!" said he, "have not the most ludicrous things always happened to us old anchorites and saints? +test-other/2609/156975/2609_156975_000036_000004|The cruel fate of his people and the painful experience in Egypt that had driven him into the wilderness prepared his mind to receive this training. +test-other/3005/163389/3005_163389_000017_000001|And they laughed all the time, and that made the duke mad; and everybody left, anyway, before the show was over, but one boy which was asleep. +test-other/3005/163390/3005_163390_000023_000021|S'pose people left money laying around where he was what did he do? +test-other/3005/163391/3005_163391_000021_000000|"It's a pretty long journey. +test-other/3005/163399/3005_163399_000013_000002|When we got there she set me down in a split bottomed chair, and set herself down on a little low stool in front of me, holding both of my hands, and says: +test-other/3005/163399/3005_163399_000045_000000|He sprung to the window at the head of the bed, and that give mrs Phelps the chance she wanted. +test-other/3080/5040/3080_5040_000000_000010|You have no such ladies in Ireland? +test-other/3331/159605/3331_159605_000006_000002|I could do so much for all at home how I should enjoy that!" And Polly let her thoughts revel in the luxurious future her fancy painted. +test-other/3331/159605/3331_159605_000082_000000|"Who got up that nice idea, I should like to know?" demanded Polly, as Fanny stopped for breath. +test-other/3528/168656/3528_168656_000003_000003|She told wonders of the Abbey of Fontevrault,--that it was like a city, and that there were streets in the monastery. +test-other/3528/168669/3528_168669_000030_000000|A silence ensued. +test-other/3528/168669/3528_168669_000075_000000|"Like yourself, reverend Mother." +test-other/3528/168669/3528_168669_000123_000000|"But the commissary of police-" +test-other/3528/168669/3528_168669_000137_000000|"That is well." +test-other/3528/168669/3528_168669_000164_000008|I shall have my lever. +test-other/3538/142836/3538_142836_000021_000003|However, as late as the reigns of our two last Georges, fabulous sums were often expended upon fanciful desserts. +test-other/3538/163619/3538_163619_000054_000000|'Now he says that you are to make haste and throw yourself overboard,' answered the step mother. +test-other/3538/163622/3538_163622_000069_000000|So they travelled onwards again, for many and many a mile, over hill and dale. +test-other/3538/163624/3538_163624_000038_000000|Then Sigurd went down into that deep place, and dug many pits in it, and in one of the pits he lay hidden with his sword drawn. +test-other/367/130732/367_130732_000002_000001|Probably nowhere in San Francisco could one get lobster better served than in the Old Delmonico restaurant of the days before the fire. +test-other/3764/168670/3764_168670_000003_000000|"But you, Father Madeleine?" +test-other/3764/168670/3764_168670_000043_000000|"Yes." +test-other/3764/168670/3764_168670_000083_000005|He grumbled:-- +test-other/3764/168671/3764_168671_000012_000003|He did what he liked with him. +test-other/3764/168671/3764_168671_000046_000000|"Comrade!" cried Fauchelevent. +test-other/3997/180294/3997_180294_000023_000000|Then, when God allows love to a courtesan, that love, which at first seems like a pardon, becomes for her almost without penitence. +test-other/3997/180294/3997_180294_000065_000001|The count will be coming back, and there is nothing to be gained by his finding you here." +test-other/3997/180297/3997_180297_000034_000004|For these people we have to be merry when they are merry, well when they want to sup, sceptics like themselves. +test-other/3997/182399/3997_182399_000014_000003|Oh, my, no! +test-other/4198/61336/4198_61336_000000_000003|It is significant to note in this connection that the new king was an unswerving adherent of the cult of Ashur, by the adherents of which he was probably strongly supported. +test-other/4198/61336/4198_61336_000033_000001|Nabonassar had died and was succeeded by his son Nabu nadin zeri, who, after reigning for two years, was slain in a rebellion. +test-other/4294/14317/4294_14317_000022_000011|I do not condescend to smite you. He looked at me submissively and said nothing. +test-other/4294/35475/4294_35475_000018_000001|At last they reached a wide chasm that bounded the Ogre's domain. +test-other/4294/35475/4294_35475_000050_000002|They said, "We are only waiting to lay some wily plan to capture the Ogre." +test-other/4294/9934/4294_9934_000025_000000|"Gold; here it is." +test-other/4350/10919/4350_10919_000006_000000|"Immediately, princess. +test-other/4350/9170/4350_9170_000005_000001|Authority, in the sense in which the word is ordinarily understood, is a means of forcing a man to act in opposition to his desires. +test-other/4350/9170/4350_9170_000056_000000|But the fatal significance of universal military service, as the manifestation of the contradiction inherent in the social conception of life, is not only apparent in that. +test-other/4852/28311/4852_28311_000031_000001|After a step or two, not finding his friend beside him, he turned. +test-other/4852/28319/4852_28319_000013_000002|mr Wicker waited patiently beside him for a few moments for Chris to get up his courage. +test-other/533/1066/533_1066_000008_000000|"I mean," he persisted, "do you feel as though you could go through with something rather unusual?" +test-other/533/131562/533_131562_000018_000000|mr Huntingdon then went up stairs. +test-other/5442/41168/5442_41168_000002_000001|Sergey Ivanovitch, waiting till the malignant gentleman had finished speaking, said that he thought the best solution would be to refer to the act itself, and asked the secretary to find the act. +test-other/5442/41169/5442_41169_000003_000000|"He's such a blackguard! +test-other/5442/41169/5442_41169_000030_000000|"And with what he made he'd increase his stock, or buy some land for a trifle, and let it out in lots to the peasants," Levin added, smiling. He had evidently more than once come across those commercial calculations. +test-other/5484/24317/5484_24317_000040_000006|Let us hope that you will make this three leaved clover the luck promising four leaved one. +test-other/5484/24318/5484_24318_000015_000002|The blood of these innocent men would be on his head if he did not listen to her representations. +test-other/5484/24318/5484_24318_000068_000001|He was appearing before his companions only to give truth its just due. +test-other/5764/299665/5764_299665_000041_000004|He saw the seeds that man had planted wither and perish, but he sent no rain. +test-other/5764/299665/5764_299665_000070_000000|Think of the egotism of a man who believes that an infinite being wants his praise! +test-other/5764/299665/5764_299665_000102_000000|The first stone is that matter-substance-cannot be destroyed, cannot be annihilated. +test-other/5764/299665/5764_299665_000134_000000|You cannot reform these people with tracts and talk. +test-other/6070/63485/6070_63485_000025_000003|Hand me the cash, and I will hand you the pocketbook." +test-other/6070/86744/6070_86744_000027_000000|"Have you bachelor's apartments there? +test-other/6070/86745/6070_86745_000001_000002|Two windows only of the pavilion faced the street; three other windows looked into the court, and two at the back into the garden. +test-other/6128/63240/6128_63240_000012_000002|Neither five nor fifteen, and yet not ten exactly, but either nine or eleven. +test-other/6128/63240/6128_63240_000042_000002|mrs Luna explained to her sister that her freedom of speech was caused by his being a relation-though, indeed, he didn't seem to know much about them. +test-other/6128/63244/6128_63244_000002_000000|"I can't talk to those people, I can't!" said Olive Chancellor, with a face which seemed to plead for a remission of responsibility. +test-other/6432/63722/6432_63722_000026_000000|"Not the least in the world-not as much as you do," was the cool answer. +test-other/6432/63722/6432_63722_000050_000004|Queen Elizabeth was very fond of watches and clocks, and her friends, knowing that, used to present her with beautiful specimens. Some of the watches of her day were made in the form of crosses, purses, little books, and even skulls." +test-other/6432/63722/6432_63722_000080_000003|When it does it will create a sensation." +test-other/6432/63723/6432_63723_000026_000000|"No; but he will, or I'll sue him and get judgment. +test-other/6432/63723/6432_63723_000057_000000|"Then for the love of-" +test-other/6432/63723/6432_63723_000080_000000|"Hello, Harry! +test-other/6938/70848/6938_70848_000046_000003|Show me the source!" +test-other/6938/70848/6938_70848_000104_000000|With biting sarcasm he went on to speak of the Allied diplomats, till then contemptuous of Russia's invitation to an armistice, which had been accepted by the Central Powers. +test-other/7105/2330/7105_2330_000021_000000|"He won't go unless he has a brass band. +test-other/7105/2340/7105_2340_000015_000001|We feel that we must live on cream for the rest of our lives. +test-other/7902/96591/7902_96591_000008_000001|I did not come to frighten you; you frightened me." +test-other/7902/96591/7902_96591_000048_000000|"No," he thought to himself, "I don't believe they would kill me, but they would knock me about." +test-other/7902/96592/7902_96592_000024_000001|Once out of that room he could ran, and by daylight the smugglers dare not hunt him down. +test-other/7902/96592/7902_96592_000063_000000|"What for?" cried Ram. +test-other/7902/96594/7902_96594_000014_000001|These fellows are very cunning, but we shall be too many for them one of these days." +test-other/7902/96594/7902_96594_000062_000001|Keep a sharp look out on the cliff to see if Mr Raystoke is making signals for a boat. +test-other/7902/96595/7902_96595_000039_000000|The man shook his head, and stared as if he didn't half understand the drift of what was said. +test-other/7975/280057/7975_280057_000009_000000|Naturally we were Southerners in sympathy and in fact. +test-other/7975/280057/7975_280057_000025_000004|On reaching the camp the first person I saw whom I knew was Cole Younger. +test-other/7975/280076/7975_280076_000013_000001|I will give you this outline and sketch of my whereabouts and actions at the time of certain robberies with which I am charged. +test-other/7975/280084/7975_280084_000007_000000|But between the time we broke camp and the time they reached the bridge the three who went ahead drank a quart of whisky, and there was the initial blunder at Northfield. +test-other/7975/280085/7975_280085_000005_000002|Some of the boys wanted to kill him, on the theory that "dead men tell no tales," while others urged binding him and leaving him in the woods. +test-other/8131/117016/8131_117016_000005_000000|The Stonewall gang numbered perhaps five hundred. +test-other/8131/117016/8131_117016_000025_000001|"And don't let them get away!" +test-other/8131/117016/8131_117016_000047_000006|I can always go back to Earth, and I'll try to take you along. +test-other/8131/117017/8131_117017_000005_000000|Gordon hit the signal switch, and the Marspeaker let out a shrill whistle. +test-other/8131/117017/8131_117017_000020_000003|There's no graft out here." +test-other/8131/117029/8131_117029_000007_000002|Wrecks were being broken up, with salvageable material used for newer homes. Gordon came to a row of temporary bubbles, individual dwellings built like the dome, but opaque for privacy. +test-other/8131/117029/8131_117029_000023_000004|But there'll be pushers as long as weak men turn to drugs, and graft as long as voters allow the thing to get out of their hands. +test-other/8188/269288/8188_269288_000018_000000|A few moments later there came a tap at the door. +test-other/8188/269288/8188_269288_000053_000001|"Do you want to kill me? +test-other/8188/269290/8188_269290_000035_000001|"But now, Leslie, what is the trouble? +test-other/8188/269290/8188_269290_000065_000000|"I don't think she is quite well," replied Leslie. +test-other/8280/266249/8280_266249_000030_000000|The ladies were weary, and retired to their state rooms shortly after tea, but the gentlemen sought the open air again and paced the deck for some time. +test-other/8280/266249/8280_266249_000113_000000|It was the last game of cards for that trip. +test-other/8461/278226/8461_278226_000026_000000|Laura thanked the French artist and then took her husband's arm and walked away with him. +test-other/8461/281231/8461_281231_000029_000002|Before long the towering flames had surmounted every obstruction, and rose to the evening skies one huge and burning beacon, seen far and wide through the adjacent country; tower after tower crashed down, with blazing roof and rafter. diff --git a/BigVGAN/filelists/LibriTTS/train-full.txt b/BigVGAN/filelists/LibriTTS/train-full.txt new file mode 100644 index 0000000000000000000000000000000000000000..934d7fa43dbc343a40951f265c5b61e591f1adc8 --- /dev/null +++ b/BigVGAN/filelists/LibriTTS/train-full.txt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:40fa291ec58520a7832baefbe934d40cd0a4d8aa516894357a4d6f5ac3dcfb23 +size 51789567 diff --git a/BigVGAN/filelists/LibriTTS/val-full.txt b/BigVGAN/filelists/LibriTTS/val-full.txt new file mode 100644 index 0000000000000000000000000000000000000000..e06cd55f99d1b7835af616420d8f026eab221134 --- /dev/null +++ b/BigVGAN/filelists/LibriTTS/val-full.txt @@ -0,0 +1,119 @@ +train-clean-100/103/1241/103_1241_000000_000001|matthew Cuthbert is surprised +train-clean-100/1594/135914/1594_135914_000033_000001|He told them, that having taken refuge in a small village, he there fell sick; that some charitable peasants had taken care of him, but finding he did not recover, a camel driver had undertaken to carry him to the hospital at Bagdad. +train-clean-100/233/155990/233_155990_000018_000002|I did, however, receive aid from the Emperor of Germany. +train-clean-100/3240/131231/3240_131231_000041_000003|Some persons, thinking them to be sea fishes, placed them in salt water, according to mr Roberts. +train-clean-100/40/222/40_222_000026_000000|"No, read it yourself," cried Catherine, whose second thoughts were clearer. +train-clean-100/4406/16882/4406_16882_000014_000002|Then they set me upon a horse with my wounded child in my lap, and there being no furniture upon the horse's back, as we were going down a steep hill we both fell over the horse's head, at which they, like inhumane creatures, laughed, and rejoiced to see it, though I thought we should there have ended our days, as overcome with so many difficulties. +train-clean-100/5393/19218/5393_19218_000115_000000|"Where is it going then?" +train-clean-100/6147/34606/6147_34606_000013_000008|One was "a dancing master;" that is to say he made the rustics frisk about by pricking the calves of their legs with the point of his sword. +train-clean-100/6848/76049/6848_76049_000003_000007|But suppose she was not all ordinary female person.... +train-clean-100/7505/258964/7505_258964_000026_000007|During the Boer War horses and mules rose in price in the United States on account of British purchases. +train-clean-100/831/130739/831_130739_000015_000000|But enough of these revelations. +train-clean-100/887/123291/887_123291_000028_000000|Here the Professor laid hold of the fossil skeleton, and handled it with the skill of a dexterous showman. +train-clean-360/112/123216/112_123216_000035_000009|The wonderful day had come and Roy's violets had no place in it. +train-clean-360/1323/149236/1323_149236_000007_000004|It was vain to hope that mere words would quiet a nation which had not, in any age, been very amenable to control, and which was now agitated by hopes and resentments, such as great revolutions, following great oppressions, naturally engender. +train-clean-360/1463/134465/1463_134465_000058_000000|Both Sandy and I began to laugh. +train-clean-360/1748/1562/1748_1562_000067_000000|"Oh, Pocket, Pocket," said I; but by this time the party which had gone towards the house, rushed out again, shouting and screaming with laughter. +train-clean-360/1914/133440/1914_133440_000014_000001|With the last twenty or thirty feet of it a deadly nausea came upon me. +train-clean-360/207/143321/207_143321_000070_000002|The canoes were not on the river bank. +train-clean-360/2272/152267/2272_152267_000003_000001|After supper the knight shared his own bed with the leper. +train-clean-360/2517/135227/2517_135227_000006_000005|As I was anxious to witness some of their purely religious ceremonies, I wished to go. +train-clean-360/2709/158074/2709_158074_000054_000000|Meanwhile the women continued to protest. +train-clean-360/2929/86777/2929_86777_000009_000000|A long silence followed; the peach, like the grapes, fell to the ground. +train-clean-360/318/124224/318_124224_000022_000010|In spite of his prejudice against Edward, he could put himself into Mr Waller's place, and see the thing from his point of view. +train-clean-360/3368/170952/3368_170952_000006_000000|And can he be fearless of death, or will he choose death in battle rather than defeat and slavery, who believes the world below to be real and terrible? +train-clean-360/3549/9203/3549_9203_000005_000004|We must hope so. There are examples. +train-clean-360/3835/178028/3835_178028_000007_000001|That day Prince Vasili no longer boasted of his protege Kutuzov, but remained silent when the commander in chief was mentioned. +train-clean-360/3994/149798/3994_149798_000005_000002|Afterward we can visit the mountain and punish the cruel magician of the Flatheads." +train-clean-360/4257/6397/4257_6397_000009_000000|At that time Nostromo had been already long enough in the country to raise to the highest pitch Captain Mitchell's opinion of the extraordinary value of his discovery. +train-clean-360/454/134728/454_134728_000133_000000|After a week of physical anguish, Unrest and pain, and feverish heat, Toward the ending day a calm and lull comes on, Three hours of peace and soothing rest of brain. +train-clean-360/4848/28247/4848_28247_000026_000002|Had he gained this arduous height only to behold the rocks carpeted with ice and snow, and reaching interminably to the far off horizon? +train-clean-360/5039/1189/5039_1189_000091_000000|The Shaggy Man sat down again and seemed well pleased. +train-clean-360/5261/19373/5261_19373_000011_000001|Some cause was evidently at work on this distant planet, causing it to disagree with its motion as calculated according to the law of gravitation. +train-clean-360/5538/70919/5538_70919_000032_000001|Only one person in the world could have laid those discoloured pearls at his door in the dead of night. The black figure in the garden, with the chiffon fluttering about its head, was Evelina Grey-or what was left of her. +train-clean-360/5712/48848/5712_48848_000060_000003|Lily for the time had been raised to a pinnacle,--a pinnacle which might be dangerous, but which was, at any rate, lofty. +train-clean-360/5935/43322/5935_43322_000050_000002|I think too-yes, I think that on the whole the ritual is impressive. +train-clean-360/6115/58433/6115_58433_000007_000002|We must run the risk." +train-clean-360/6341/64956/6341_64956_000040_000000|"Why, papa, I thought we were going to have such a nice time, and she just spoiled it all." +train-clean-360/6509/67147/6509_67147_000028_000003|It "was n't done" in England. +train-clean-360/6694/70837/6694_70837_000027_000002|There an enormous smiling sailor stopped me, and when I showed my pass, just said, "If you were Saint Michael himself, comrade, you couldn't pass here!" Through the glass of the door I made out the distorted face and gesticulating arms of a French correspondent, locked in.... +train-clean-360/6956/76046/6956_76046_000055_000001|Twelve hundred, fifteen hundred millions perhaps." +train-clean-360/7145/87280/7145_87280_000004_000003|This modern Ulysses made a masterful effort, but alas! had no ships to carry him away, and no wax with which to fill his ears. +train-clean-360/7314/77782/7314_77782_000011_000000|"Well, then, what in thunder is the matter with you?" cried the Lawyer, irritated. +train-clean-360/7525/92915/7525_92915_000034_000001|It was desperate, too, and lasted nearly all day-and it was one of the important battles of the world, although the numbers engaged in it were not large. +train-clean-360/7754/108640/7754_108640_000001_000004|Was I aware-was I fully aware of the discrepancy between us? +train-clean-360/7909/106369/7909_106369_000006_000002|And Colchian Aea lies at the edge of Pontus and of the world." +train-clean-360/8011/280922/8011_280922_000009_000000|He stretched out his hand, and all at once stroked my cheek. +train-clean-360/8176/115046/8176_115046_000027_000001|"Bless my soul, I never can understand it!" +train-clean-360/8459/292347/8459_292347_000015_000000|A woman near Gort, in Galway, says: 'There is a boy, now, of the Cloran's; but I wouldn't for the world let them think I spoke of him; it's two years since he came from America, and since that time he never went to Mass, or to church, or to fairs, or to market, or to stand on the cross roads, or to hurling, or to nothing. +train-clean-360/8699/291107/8699_291107_000003_000005|He leaned closer over it, regardless of the thin choking haze that spread about his face. In his attitude there was a rigidity of controlled excitement out of keeping with the seeming harmlessness of the experiment. +train-clean-360/8855/283242/8855_283242_000061_000000|"That couldn't be helped, grannie. +train-other-500/102/129232/102_129232_000050_000000|Is it otherwise in the newest romance? +train-other-500/1124/134775/1124_134775_000087_000001|Some of them are enclosed only by hedges, which lends a cheerful aspect to the street. +train-other-500/1239/138254/1239_138254_000010_000001|It was past twelve when all preparations were finished. +train-other-500/1373/132103/1373_132103_000056_000000|So they moved on. +train-other-500/1566/153036/1566_153036_000087_000003|You enter the river close by the trees, and then keep straight for the pile of stones, which is some fifty yards higher up, for the ford crosses the river at an angle." +train-other-500/1653/142352/1653_142352_000005_000002|If he should not come! +train-other-500/1710/133294/1710_133294_000023_000000|When the Indians were the sole inhabitants of the wilds from whence they have since been expelled, their wants were few. +train-other-500/1773/139602/1773_139602_000032_000001|When the rabbit saw that the badger was getting well, he thought of another plan by which he could compass the creature's death. +train-other-500/1920/1793/1920_1793_000037_000001|She has a little Blenheim lapdog, that she loves a thousand times more than she ever will me!" +train-other-500/2067/143535/2067_143535_000009_000002|Indeed, there, to the left, was a stone shelf with a little ledge to it three inches or so high, and on the shelf lay what I took to be a corpse; at any rate, it looked like one, with something white thrown over it. +train-other-500/2208/11020/2208_11020_000037_000001|It's at my place over there.' +train-other-500/2312/157868/2312_157868_000019_000002|I am the manager of the theatre, and I'm thundering glad that your first play has been produced at the 'New York,' sir. +train-other-500/2485/151992/2485_151992_000028_000005|At last he looked up at his wife and said, in a gentle tone: +train-other-500/2587/54186/2587_54186_000015_000000|Concerning the work as a whole he wrote to Clara while in the throes of composition: "This music now in me, and always such beautiful melodies! +train-other-500/2740/288813/2740_288813_000018_000003|But Philip had kept him apart, had banked him off, and yet drained him to the dregs. +train-other-500/2943/171001/2943_171001_000122_000000|The sound of his voice pronouncing her name aroused her. +train-other-500/3063/138651/3063_138651_000028_000000|But, as may be imagined, the unfortunate john was as much surprised by this rencounter as the other two. +train-other-500/3172/166439/3172_166439_000050_000000|And now at last was clear a thing that had puzzled greatly-the mechanism of that opening process by which sphere became oval disk, pyramid a four pointed star and-as I had glimpsed in the play of the Little Things about Norhala, could see now so plainly in the Keeper-the blocks took this inverted cruciform shape. +train-other-500/331/132019/331_132019_000038_000000|"I say, this is folly! +train-other-500/3467/166570/3467_166570_000054_000001|Does he never mention Orlando?" +train-other-500/3587/140711/3587_140711_000015_000001|O fie, mrs Jervis, said I, how could you serve me so? Besides, it looks too free both in me, and to him. +train-other-500/3675/187020/3675_187020_000026_000001|"I wonder what would be suitable? +train-other-500/3819/134146/3819_134146_000019_000001|Also the figure half hidden by the cupboard door-was a female figure, massive, and in flowing robes. +train-other-500/3912/77626/3912_77626_000003_000004|You may almost distinguish the figures on the clock that has just told the hour. +train-other-500/4015/63729/4015_63729_000058_000000|"It does." +train-other-500/413/22436/413_22436_000035_000003|I conjecture, the French squadron is bound for Malta and Alexandria, and the Spanish fleet for the attack of Minorca." +train-other-500/4218/41159/4218_41159_000028_000002|Yes? That worries Alexey. +train-other-500/4352/10940/4352_10940_000037_000002|He doesn't exist." +train-other-500/4463/26871/4463_26871_000023_000000|"I did not notice him following me," she said timidly. +train-other-500/4591/14356/4591_14356_000019_000000|"Within three days," cried the enchanter, loudly, "bring Rinaldo and Ricciardetto into the pass of Ronces Valles. +train-other-500/4738/291957/4738_291957_000000_000001|ODE ON THE SPRING. +train-other-500/4824/36029/4824_36029_000045_000003|And indeed Janet herself had taken no part in the politics, content merely to advise the combatants upon their demeanour. +train-other-500/4936/65528/4936_65528_000014_000007|I immediately responded, "Yes, they are most terrible struck on each other," and I said it in a tone that indicated I thought it a most beautiful and lovely thing that they should be so. +train-other-500/5019/38670/5019_38670_000017_000000|"Let me make you a present of the gloves," she said, with her irresistible smile. +train-other-500/5132/33409/5132_33409_000016_000001|They waited on the table in Valhalla. +train-other-500/52/121057/52_121057_000019_000000|"I," cried the steward with a strange expression. +train-other-500/5321/53046/5321_53046_000025_000003|I gather from what mrs joel said that she's rather touched in her mind too, and has an awful hankering to get home here-to this very house. +train-other-500/5429/210770/5429_210770_000029_000006|But this was not all. +train-other-500/557/129797/557_129797_000072_000001|The guns were manned, the gunners already kindling fuses, when the buccaneer fleet, whilst still heading for Palomas, was observed to bear away to the west. +train-other-500/572/128861/572_128861_000016_000002|My home was desolate. +train-other-500/5826/53497/5826_53497_000044_000001|If it be as you say, he will have shown himself noble, and his nobility will have consisted in this, that he has been willing to take that which he does not want, in order that he may succour one whom he loves. +train-other-500/5906/52158/5906_52158_000055_000000|The impression that he gets this knowledge or suspicion from the outside is due, the scientists say, to the fact that his thinking has proceeded at such lightning like speed that he was unable to watch the wheels go round. +train-other-500/6009/57639/6009_57639_000038_000000|This, friendly reader, is my only motive. +train-other-500/6106/58196/6106_58196_000007_000001|I tell you that you must make the dress. +train-other-500/6178/86034/6178_86034_000079_000004|Then she will grow calmer, and will know you again. +train-other-500/6284/63091/6284_63091_000133_000001|I don't want to go anywhere where anybody'll see me." +train-other-500/6436/104980/6436_104980_000009_000002|I guess you never heard about this house." +train-other-500/6540/232291/6540_232291_000017_000003|The girl was not wholly a savage. +train-other-500/6627/67844/6627_67844_000046_000002|The other girls had stopped talking, and now looked at Sylvia as if wondering what she would say. +train-other-500/6707/77351/6707_77351_000002_000006|But our first words I may give you, because though they conveyed nothing to me at the time, afterwards they meant much. +train-other-500/6777/76694/6777_76694_000013_000011|When they are forcibly put out of Garraway's on Saturday night-which they must be, for they never would go out of their own accord-where do they vanish until Monday morning? +train-other-500/690/133452/690_133452_000011_000000|Campany lifted his quill pen and pointed to a case of big leather bound volumes in a far corner of the room. +train-other-500/7008/34667/7008_34667_000032_000002|What had happened? +train-other-500/7131/92815/7131_92815_000039_000001|The cabman tried to pass to the left, but a heavy express wagon cut him off. +train-other-500/7220/77911/7220_77911_000005_000000|"Do? +train-other-500/7326/245693/7326_245693_000008_000000|Whether the Appetite Is a Special Power of the Soul? +train-other-500/7392/105672/7392_105672_000013_000005|Whoever, being required, refused to answer upon oath to any article of this act of settlement, was declared to be guilty of treason; and by this clause a species of political inquisition was established in the kingdom, as well as the accusations of treason multiplied to an unreasonable degree. +train-other-500/7512/98636/7512_98636_000017_000002|A man thus rarely makes provision for the future, and looks with scorn on foreign customs which seem to betoken a fear lest, in old age, ungrateful children may neglect their parents and cast them aside. +train-other-500/7654/258963/7654_258963_000007_000007|Egypt, for a time reduced to a semi desert condition, has only in the past century been restored to a certain extent by the use of new methods and a return to the old ones. +train-other-500/7769/99396/7769_99396_000020_000002|I had to go out once a day in search of food. +train-other-500/791/127519/791_127519_000086_000000|This was how it came about. +train-other-500/8042/113769/8042_113769_000021_000000|House the second. +train-other-500/8180/274725/8180_274725_000010_000000|"What fools men are in love matters," quoth Patty to herself-"at least most men!" with a thought backward to Mark's sensible choosing. +train-other-500/8291/282929/8291_282929_000031_000006|He's in a devil of a-Well, he needs the money, and I've got to get it for him. You know my word's good, Cooper." +train-other-500/8389/120181/8389_120181_000022_000000|"No," I answered. +train-other-500/8476/269293/8476_269293_000078_000001|Annie, in some wonder, went downstairs alone. +train-other-500/8675/295195/8675_295195_000004_000004|Everything had gone on prosperously with them, and they had reared many successive families of young Nutcrackers, who went forth to assume their places in the forest of life, and to reflect credit on their bringing up,--so that naturally enough they began to have a very easy way of considering themselves models of wisdom. +train-other-500/9000/282381/9000_282381_000016_000008|Bank facings seemed to indicate that the richest pay dirt lay at bed rock. +train-other-500/978/132494/978_132494_000017_000001|And what made you come at that very minute? diff --git a/BigVGAN/incl_licenses/LICENSE_1 b/BigVGAN/incl_licenses/LICENSE_1 new file mode 100644 index 0000000000000000000000000000000000000000..5afae394d6b37da0e12ba6b290d2512687f421ac --- /dev/null +++ b/BigVGAN/incl_licenses/LICENSE_1 @@ -0,0 +1,21 @@ +MIT License + +Copyright (c) 2020 Jungil Kong + +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. \ No newline at end of file diff --git a/BigVGAN/incl_licenses/LICENSE_2 b/BigVGAN/incl_licenses/LICENSE_2 new file mode 100644 index 0000000000000000000000000000000000000000..322b758863c4219be68291ae3826218baa93cb4c --- /dev/null +++ b/BigVGAN/incl_licenses/LICENSE_2 @@ -0,0 +1,21 @@ +MIT License + +Copyright (c) 2020 Edward Dixon + +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. \ No newline at end of file diff --git a/BigVGAN/incl_licenses/LICENSE_3 b/BigVGAN/incl_licenses/LICENSE_3 new file mode 100644 index 0000000000000000000000000000000000000000..56ee3c8c4cc2b4b32e0975d17258f9ba515fdbcc --- /dev/null +++ b/BigVGAN/incl_licenses/LICENSE_3 @@ -0,0 +1,201 @@ + Apache License + Version 2.0, January 2004 + http://www.apache.org/licenses/ + + TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION + + 1. Definitions. + + "License" shall mean the terms and conditions for use, reproduction, + and distribution as defined by Sections 1 through 9 of this document. + + "Licensor" shall mean the copyright owner or entity authorized by + the copyright owner that is granting the License. + + "Legal Entity" shall mean the union of the acting entity and all + other entities that control, are controlled by, or are under common + control with that entity. For the purposes of this definition, + "control" means (i) the power, direct or indirect, to cause the + direction or management of such entity, whether by contract or + otherwise, or (ii) ownership of fifty percent (50%) or more of the + outstanding shares, or (iii) beneficial ownership of such entity. + + "You" (or "Your") shall mean an individual or Legal Entity + exercising permissions granted by this License. + + "Source" form shall mean the preferred form for making modifications, + including but not limited to software source code, documentation + source, and configuration files. + + "Object" form shall mean any form resulting from mechanical + transformation or translation of a Source form, including but + not limited to compiled object code, generated documentation, + and conversions to other media types. + + "Work" shall mean the work of authorship, whether in Source or + Object form, made available under the License, as indicated by a + copyright notice that is included in or attached to the work + (an example is provided in the Appendix below). + + "Derivative Works" shall mean any work, whether in Source or Object + form, that is based on (or derived from) the Work and for which the + editorial revisions, annotations, elaborations, or other modifications + represent, as a whole, an original work of authorship. For the purposes + of this License, Derivative Works shall not include works that remain + separable from, or merely link (or bind by name) to the interfaces of, + the Work and Derivative Works thereof. + + "Contribution" shall mean any work of authorship, including + the original version of the Work and any modifications or additions + to that Work or Derivative Works thereof, that is intentionally + submitted to Licensor for inclusion in the Work 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 Work, but + excluding communication that is conspicuously marked or otherwise + designated in writing by the copyright owner as "Not a Contribution." + + "Contributor" shall mean Licensor and any individual or Legal Entity + on behalf of whom a Contribution has been received by Licensor and + subsequently incorporated within the Work. + + 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 Derivative Works of, + publicly display, publicly perform, sublicense, and distribute the + Work and such Derivative Works in Source or Object form. + + 3. Grant of Patent 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 + (except as stated in this section) patent license to make, have made, + use, offer to sell, sell, import, and otherwise transfer the Work, + 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 Work 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 Work + or a Contribution incorporated within the Work constitutes direct + or contributory patent infringement, then any patent licenses + granted to You under this License for that Work shall terminate + as of the date such litigation is filed. + + 4. Redistribution. You may reproduce and distribute copies of the + Work or Derivative Works thereof in any medium, with or without + modifications, and in Source or Object form, provided that You + meet the following conditions: + + (a) You must give any other recipients of the Work or + Derivative Works a copy of this License; and + + (b) You must cause any modified files to carry prominent notices + stating that You changed the files; and + + (c) You must retain, in the Source form of any Derivative Works + that You distribute, all copyright, patent, trademark, and + attribution notices from the Source form of the Work, + excluding those notices that do not pertain to any part of + the Derivative Works; and + + (d) If the Work includes a "NOTICE" text file as part of its + distribution, then any Derivative Works that You distribute must + include a readable copy of the attribution notices contained + within such NOTICE file, excluding those notices that do not + pertain to any part of the Derivative Works, in at least one + of the following places: within a NOTICE text file distributed + as part of the Derivative Works; within the Source form or + documentation, if provided along with the Derivative Works; or, + within a display generated by the Derivative Works, if and + wherever such third-party notices normally appear. The contents + of the NOTICE file are for informational purposes only and + do not modify the License. You may add Your own attribution + notices within Derivative Works that You distribute, alongside + or as an addendum to the NOTICE text from the Work, provided + that such additional attribution notices cannot be construed + as modifying the License. + + You may add Your own copyright statement to Your modifications and + may provide additional or different license terms and conditions + for use, reproduction, or distribution of Your modifications, or + for any such Derivative Works as a whole, provided Your use, + reproduction, and distribution of the Work otherwise complies with + the conditions stated in this License. + + 5. Submission of Contributions. Unless You explicitly state otherwise, + any Contribution intentionally submitted for inclusion in the Work + by You to the Licensor shall be under the terms and conditions of + this License, without any additional terms or conditions. + Notwithstanding the above, nothing herein shall supersede or modify + the terms of any separate license agreement you may have executed + with Licensor regarding such Contributions. + + 6. Trademarks. This License does not grant permission to use the trade + names, trademarks, service marks, or product names of the Licensor, + except as required for reasonable and customary use in describing the + origin of the Work and reproducing the content of the NOTICE file. + + 7. Disclaimer of Warranty. Unless required by applicable law or + agreed to in writing, Licensor provides the Work (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 Work and assume any + risks associated with Your exercise of permissions under this License. + + 8. 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 + Work (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. + + 9. Accepting Warranty or Additional Liability. While redistributing + the Work or Derivative Works 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. + + END OF TERMS AND CONDITIONS + + APPENDIX: How to apply the Apache License to your work. + + To apply the Apache License to your work, attach the following + boilerplate notice, with the fields enclosed by brackets "[]" + replaced with your own identifying information. (Don't include + the brackets!) The text should be enclosed in the appropriate + comment syntax for the file format. We also recommend that a + file or class name and description of purpose be included on the + same "printed page" as the copyright notice for easier + identification within third-party archives. + + Copyright [yyyy] [name of copyright owner] + + 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. \ No newline at end of file diff --git a/BigVGAN/incl_licenses/LICENSE_4 b/BigVGAN/incl_licenses/LICENSE_4 new file mode 100644 index 0000000000000000000000000000000000000000..48fd1a1ba8d81a94b6c7d1c2ff1a1f307cc5371d --- /dev/null +++ b/BigVGAN/incl_licenses/LICENSE_4 @@ -0,0 +1,29 @@ +BSD 3-Clause License + +Copyright (c) 2019, Seungwon Park 박승원 +All rights reserved. + +Redistribution and use in source and binary forms, with or without +modification, are permitted provided that the following conditions are met: + +1. Redistributions of source code must retain the above copyright notice, this + list of conditions and the following disclaimer. + +2. Redistributions in binary form must reproduce the above copyright notice, + this list of conditions and the following disclaimer in the documentation + and/or other materials provided with the distribution. + +3. Neither the name of the copyright holder nor the names of its + contributors may be used to endorse or promote products derived from + this software without specific prior written permission. + +THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" +AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE +IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE +DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE +FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL +DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR +SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER +CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, +OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE +OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. \ No newline at end of file diff --git a/BigVGAN/incl_licenses/LICENSE_5 b/BigVGAN/incl_licenses/LICENSE_5 new file mode 100644 index 0000000000000000000000000000000000000000..01ae5538e6b7c787bb4f5d6f2cd9903520d6e465 --- /dev/null +++ b/BigVGAN/incl_licenses/LICENSE_5 @@ -0,0 +1,16 @@ +Copyright 2020 Alexandre Défossez + +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. \ No newline at end of file diff --git a/BigVGAN/incl_licenses/LICENSE_6 b/BigVGAN/incl_licenses/LICENSE_6 new file mode 100644 index 0000000000000000000000000000000000000000..2569ec0b6c85f94f3cd071ba16e9028ccf156be2 --- /dev/null +++ b/BigVGAN/incl_licenses/LICENSE_6 @@ -0,0 +1,21 @@ +MIT License + +Copyright (c) 2023-present, Descript + +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. \ No newline at end of file diff --git a/BigVGAN/incl_licenses/LICENSE_7 b/BigVGAN/incl_licenses/LICENSE_7 new file mode 100644 index 0000000000000000000000000000000000000000..c37bdaf99c6921f5849425d546069e972f52d7fa --- /dev/null +++ b/BigVGAN/incl_licenses/LICENSE_7 @@ -0,0 +1,21 @@ +MIT License + +Copyright (c) 2023 Charactr Inc. + +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. \ No newline at end of file diff --git a/BigVGAN/incl_licenses/LICENSE_8 b/BigVGAN/incl_licenses/LICENSE_8 new file mode 100644 index 0000000000000000000000000000000000000000..ab3d7ffe795779f54e339078e4e752ad9019aae8 --- /dev/null +++ b/BigVGAN/incl_licenses/LICENSE_8 @@ -0,0 +1,21 @@ +MIT License + +Copyright (c) 2023 Amphion + +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. \ No newline at end of file diff --git a/BigVGAN/inference.py b/BigVGAN/inference.py new file mode 100644 index 0000000000000000000000000000000000000000..a213f3185a0f932d459ae397732ddaa69ec131d8 --- /dev/null +++ b/BigVGAN/inference.py @@ -0,0 +1,89 @@ +# Adapted from https://github.com/jik876/hifi-gan under the MIT license. +# LICENSE is in incl_licenses directory. + +from __future__ import absolute_import, division, print_function, unicode_literals + +import os +import argparse +import json +import torch +import librosa +from utils import load_checkpoint +from meldataset import get_mel_spectrogram +from scipy.io.wavfile import write +from env import AttrDict +from meldataset import MAX_WAV_VALUE +from bigvgan import BigVGAN as Generator + +h = None +device = None +torch.backends.cudnn.benchmark = False + + +def inference(a, h): + generator = Generator(h, use_cuda_kernel=a.use_cuda_kernel).to(device) + + state_dict_g = load_checkpoint(a.checkpoint_file, device) + generator.load_state_dict(state_dict_g["generator"]) + + filelist = os.listdir(a.input_wavs_dir) + + os.makedirs(a.output_dir, exist_ok=True) + + generator.eval() + generator.remove_weight_norm() + with torch.no_grad(): + for i, filname in enumerate(filelist): + # Load the ground truth audio and resample if necessary + wav, sr = librosa.load( + os.path.join(a.input_wavs_dir, filname), sr=h.sampling_rate, mono=True + ) + wav = torch.FloatTensor(wav).to(device) + # Compute mel spectrogram from the ground truth audio + x = get_mel_spectrogram(wav.unsqueeze(0), generator.h) + + y_g_hat = generator(x) + + audio = y_g_hat.squeeze() + audio = audio * MAX_WAV_VALUE + audio = audio.cpu().numpy().astype("int16") + + output_file = os.path.join( + a.output_dir, os.path.splitext(filname)[0] + "_generated.wav" + ) + write(output_file, h.sampling_rate, audio) + print(output_file) + + +def main(): + print("Initializing Inference Process..") + + parser = argparse.ArgumentParser() + parser.add_argument("--input_wavs_dir", default="test_files") + parser.add_argument("--output_dir", default="generated_files") + parser.add_argument("--checkpoint_file", required=True) + parser.add_argument("--use_cuda_kernel", action="store_true", default=False) + + a = parser.parse_args() + + config_file = os.path.join(os.path.split(a.checkpoint_file)[0], "config.json") + with open(config_file) as f: + data = f.read() + + global h + json_config = json.loads(data) + h = AttrDict(json_config) + + torch.manual_seed(h.seed) + global device + if torch.cuda.is_available(): + torch.cuda.manual_seed(h.seed) + device = torch.device("cuda") + else: + device = torch.device("cpu") + + inference(a, h) + + +if __name__ == "__main__": + main() diff --git a/BigVGAN/inference_e2e.py b/BigVGAN/inference_e2e.py new file mode 100644 index 0000000000000000000000000000000000000000..a39dc67a671355343ca8f716d34b31a0a8da8d77 --- /dev/null +++ b/BigVGAN/inference_e2e.py @@ -0,0 +1,102 @@ +# Adapted from https://github.com/jik876/hifi-gan under the MIT license. +# LICENSE is in incl_licenses directory. + +from __future__ import absolute_import, division, print_function, unicode_literals + +import glob +import os +import numpy as np +import argparse +import json +import torch +from scipy.io.wavfile import write +from env import AttrDict +from meldataset import MAX_WAV_VALUE +from bigvgan import BigVGAN as Generator + +h = None +device = None +torch.backends.cudnn.benchmark = False + + +def load_checkpoint(filepath, device): + assert os.path.isfile(filepath) + print(f"Loading '{filepath}'") + checkpoint_dict = torch.load(filepath, map_location=device) + print("Complete.") + return checkpoint_dict + + +def scan_checkpoint(cp_dir, prefix): + pattern = os.path.join(cp_dir, prefix + "*") + cp_list = glob.glob(pattern) + if len(cp_list) == 0: + return "" + return sorted(cp_list)[-1] + + +def inference(a, h): + generator = Generator(h, use_cuda_kernel=a.use_cuda_kernel).to(device) + + state_dict_g = load_checkpoint(a.checkpoint_file, device) + generator.load_state_dict(state_dict_g["generator"]) + + filelist = os.listdir(a.input_mels_dir) + + os.makedirs(a.output_dir, exist_ok=True) + + generator.eval() + generator.remove_weight_norm() + with torch.no_grad(): + for i, filname in enumerate(filelist): + # Load the mel spectrogram in .npy format + x = np.load(os.path.join(a.input_mels_dir, filname)) + x = torch.FloatTensor(x).to(device) + if len(x.shape) == 2: + x = x.unsqueeze(0) + + y_g_hat = generator(x) + + audio = y_g_hat.squeeze() + audio = audio * MAX_WAV_VALUE + audio = audio.cpu().numpy().astype("int16") + + output_file = os.path.join( + a.output_dir, os.path.splitext(filname)[0] + "_generated_e2e.wav" + ) + write(output_file, h.sampling_rate, audio) + print(output_file) + + +def main(): + print("Initializing Inference Process..") + + parser = argparse.ArgumentParser() + parser.add_argument("--input_mels_dir", default="test_mel_files") + parser.add_argument("--output_dir", default="generated_files_from_mel") + parser.add_argument("--checkpoint_file", required=True) + parser.add_argument("--use_cuda_kernel", action="store_true", default=False) + + a = parser.parse_args() + + config_file = os.path.join(os.path.split(a.checkpoint_file)[0], "config.json") + with open(config_file) as f: + data = f.read() + + global h + json_config = json.loads(data) + h = AttrDict(json_config) + + torch.manual_seed(h.seed) + global device + if torch.cuda.is_available(): + torch.cuda.manual_seed(h.seed) + device = torch.device("cuda") + else: + device = torch.device("cpu") + + inference(a, h) + + +if __name__ == "__main__": + main() diff --git a/BigVGAN/loss.py b/BigVGAN/loss.py new file mode 100644 index 0000000000000000000000000000000000000000..d581151266f5c067ab1ec2f29dbf6b0bc095aaca --- /dev/null +++ b/BigVGAN/loss.py @@ -0,0 +1,254 @@ +# Copyright (c) 2024 NVIDIA CORPORATION. +# Licensed under the MIT license. + +# Adapted from https://github.com/jik876/hifi-gan under the MIT license. +# LICENSE is in incl_licenses directory. + + +import torch +import torch.nn.functional as F +import torch.nn as nn +from librosa.filters import mel as librosa_mel_fn +from scipy import signal + +import typing +from typing import Optional, List, Union, Dict, Tuple +from collections import namedtuple +import math +import functools + + +# Adapted from https://github.com/descriptinc/descript-audio-codec/blob/main/dac/nn/loss.py under the MIT license. +# LICENSE is in incl_licenses directory. +class MultiScaleMelSpectrogramLoss(nn.Module): + """Compute distance between mel spectrograms. Can be used + in a multi-scale way. + + Parameters + ---------- + n_mels : List[int] + Number of mels per STFT, by default [5, 10, 20, 40, 80, 160, 320], + window_lengths : List[int], optional + Length of each window of each STFT, by default [32, 64, 128, 256, 512, 1024, 2048] + loss_fn : typing.Callable, optional + How to compare each loss, by default nn.L1Loss() + clamp_eps : float, optional + Clamp on the log magnitude, below, by default 1e-5 + mag_weight : float, optional + Weight of raw magnitude portion of loss, by default 0.0 (no ampliciation on mag part) + log_weight : float, optional + Weight of log magnitude portion of loss, by default 1.0 + pow : float, optional + Power to raise magnitude to before taking log, by default 1.0 + weight : float, optional + Weight of this loss, by default 1.0 + match_stride : bool, optional + Whether to match the stride of convolutional layers, by default False + + Implementation copied from: https://github.com/descriptinc/lyrebird-audiotools/blob/961786aa1a9d628cca0c0486e5885a457fe70c1a/audiotools/metrics/spectral.py + Additional code copied and modified from https://github.com/descriptinc/audiotools/blob/master/audiotools/core/audio_signal.py + """ + + def __init__( + self, + sampling_rate: int, + n_mels: List[int] = [5, 10, 20, 40, 80, 160, 320], + window_lengths: List[int] = [32, 64, 128, 256, 512, 1024, 2048], + loss_fn: typing.Callable = nn.L1Loss(), + clamp_eps: float = 1e-5, + mag_weight: float = 0.0, + log_weight: float = 1.0, + pow: float = 1.0, + weight: float = 1.0, + match_stride: bool = False, + mel_fmin: List[float] = [0, 0, 0, 0, 0, 0, 0], + mel_fmax: List[float] = [None, None, None, None, None, None, None], + window_type: str = "hann", + ): + super().__init__() + self.sampling_rate = sampling_rate + + STFTParams = namedtuple( + "STFTParams", + ["window_length", "hop_length", "window_type", "match_stride"], + ) + + self.stft_params = [ + STFTParams( + window_length=w, + hop_length=w // 4, + match_stride=match_stride, + window_type=window_type, + ) + for w in window_lengths + ] + self.n_mels = n_mels + self.loss_fn = loss_fn + self.clamp_eps = clamp_eps + self.log_weight = log_weight + self.mag_weight = mag_weight + self.weight = weight + self.mel_fmin = mel_fmin + self.mel_fmax = mel_fmax + self.pow = pow + + @staticmethod + @functools.lru_cache(None) + def get_window( + window_type, + window_length, + ): + return signal.get_window(window_type, window_length) + + @staticmethod + @functools.lru_cache(None) + def get_mel_filters(sr, n_fft, n_mels, fmin, fmax): + return librosa_mel_fn(sr=sr, n_fft=n_fft, n_mels=n_mels, fmin=fmin, fmax=fmax) + + def mel_spectrogram( + self, + wav, + n_mels, + fmin, + fmax, + window_length, + hop_length, + match_stride, + window_type, + ): + """ + Mirrors AudioSignal.mel_spectrogram used by BigVGAN-v2 training from: + https://github.com/descriptinc/audiotools/blob/master/audiotools/core/audio_signal.py + """ + B, C, T = wav.shape + + if match_stride: + assert ( + hop_length == window_length // 4 + ), "For match_stride, hop must equal n_fft // 4" + right_pad = math.ceil(T / hop_length) * hop_length - T + pad = (window_length - hop_length) // 2 + else: + right_pad = 0 + pad = 0 + + wav = torch.nn.functional.pad(wav, (pad, pad + right_pad), mode="reflect") + + window = self.get_window(window_type, window_length) + window = torch.from_numpy(window).to(wav.device).float() + + stft = torch.stft( + wav.reshape(-1, T), + n_fft=window_length, + hop_length=hop_length, + window=window, + return_complex=True, + center=True, + ) + _, nf, nt = stft.shape + stft = stft.reshape(B, C, nf, nt) + if match_stride: + """ + Drop first two and last two frames, which are added, because of padding. Now num_frames * hop_length = num_samples. + """ + stft = stft[..., 2:-2] + magnitude = torch.abs(stft) + + nf = magnitude.shape[2] + mel_basis = self.get_mel_filters( + self.sampling_rate, 2 * (nf - 1), n_mels, fmin, fmax + ) + mel_basis = torch.from_numpy(mel_basis).to(wav.device) + mel_spectrogram = magnitude.transpose(2, -1) @ mel_basis.T + mel_spectrogram = mel_spectrogram.transpose(-1, 2) + + return mel_spectrogram + + def forward(self, x: torch.Tensor, y: torch.Tensor) -> torch.Tensor: + """Computes mel loss between an estimate and a reference + signal. + + Parameters + ---------- + x : torch.Tensor + Estimate signal + y : torch.Tensor + Reference signal + + Returns + ------- + torch.Tensor + Mel loss. + """ + + loss = 0.0 + for n_mels, fmin, fmax, s in zip( + self.n_mels, self.mel_fmin, self.mel_fmax, self.stft_params + ): + kwargs = { + "n_mels": n_mels, + "fmin": fmin, + "fmax": fmax, + "window_length": s.window_length, + "hop_length": s.hop_length, + "match_stride": s.match_stride, + "window_type": s.window_type, + } + + x_mels = self.mel_spectrogram(x, **kwargs) + y_mels = self.mel_spectrogram(y, **kwargs) + x_logmels = torch.log( + x_mels.clamp(min=self.clamp_eps).pow(self.pow) + ) / torch.log(torch.tensor(10.0)) + y_logmels = torch.log( + y_mels.clamp(min=self.clamp_eps).pow(self.pow) + ) / torch.log(torch.tensor(10.0)) + + loss += self.log_weight * self.loss_fn(x_logmels, y_logmels) + loss += self.mag_weight * self.loss_fn(x_logmels, y_logmels) + + return loss + + +# Loss functions +def feature_loss( + fmap_r: List[List[torch.Tensor]], fmap_g: List[List[torch.Tensor]] +) -> torch.Tensor: + + loss = 0 + for dr, dg in zip(fmap_r, fmap_g): + for rl, gl in zip(dr, dg): + loss += torch.mean(torch.abs(rl - gl)) + + return loss * 2 # This equates to lambda=2.0 for the feature matching loss + + +def discriminator_loss( + disc_real_outputs: List[torch.Tensor], disc_generated_outputs: List[torch.Tensor] +) -> Tuple[torch.Tensor, List[torch.Tensor], List[torch.Tensor]]: + + loss = 0 + r_losses = [] + g_losses = [] + for dr, dg in zip(disc_real_outputs, disc_generated_outputs): + r_loss = torch.mean((1 - dr) ** 2) + g_loss = torch.mean(dg**2) + loss += r_loss + g_loss + r_losses.append(r_loss.item()) + g_losses.append(g_loss.item()) + + return loss, r_losses, g_losses + + +def generator_loss( + disc_outputs: List[torch.Tensor], +) -> Tuple[torch.Tensor, List[torch.Tensor]]: + + loss = 0 + gen_losses = [] + for dg in disc_outputs: + l = torch.mean((1 - dg) ** 2) + gen_losses.append(l) + loss += l + + return loss, gen_losses diff --git a/BigVGAN/meldataset.py b/BigVGAN/meldataset.py new file mode 100644 index 0000000000000000000000000000000000000000..a5859b9e127e639b2f0444343bdb8e2c99cfdf1c --- /dev/null +++ b/BigVGAN/meldataset.py @@ -0,0 +1,396 @@ +# Copyright (c) 2024 NVIDIA CORPORATION. +# Licensed under the MIT license. + +# Adapted from https://github.com/jik876/hifi-gan under the MIT license. +# LICENSE is in incl_licenses directory. + +import math +import os +import random +import torch +import torch.utils.data +import numpy as np +import librosa +from librosa.filters import mel as librosa_mel_fn +import pathlib +from tqdm import tqdm +from typing import List, Tuple, Optional +from .env import AttrDict + +MAX_WAV_VALUE = 32767.0 # NOTE: 32768.0 -1 to prevent int16 overflow (results in popping sound in corner cases) + + +def dynamic_range_compression(x, C=1, clip_val=1e-5): + return np.log(np.clip(x, a_min=clip_val, a_max=None) * C) + + +def dynamic_range_decompression(x, C=1): + return np.exp(x) / C + + +def dynamic_range_compression_torch(x, C=1, clip_val=1e-5): + return torch.log(torch.clamp(x, min=clip_val) * C) + + +def dynamic_range_decompression_torch(x, C=1): + return torch.exp(x) / C + + +def spectral_normalize_torch(magnitudes): + return dynamic_range_compression_torch(magnitudes) + + +def spectral_de_normalize_torch(magnitudes): + return dynamic_range_decompression_torch(magnitudes) + + +mel_basis_cache = {} +hann_window_cache = {} + + +def mel_spectrogram( + y: torch.Tensor, + n_fft: int, + num_mels: int, + sampling_rate: int, + hop_size: int, + win_size: int, + fmin: int, + fmax: int = None, + center: bool = False, +) -> torch.Tensor: + """ + Calculate the mel spectrogram of an input signal. + This function uses slaney norm for the librosa mel filterbank (using librosa.filters.mel) and uses Hann window for STFT (using torch.stft). + + Args: + y (torch.Tensor): Input signal. + n_fft (int): FFT size. + num_mels (int): Number of mel bins. + sampling_rate (int): Sampling rate of the input signal. + hop_size (int): Hop size for STFT. + win_size (int): Window size for STFT. + fmin (int): Minimum frequency for mel filterbank. + fmax (int): Maximum frequency for mel filterbank. If None, defaults to half the sampling rate (fmax = sr / 2.0) inside librosa_mel_fn + center (bool): Whether to pad the input to center the frames. Default is False. + + Returns: + torch.Tensor: Mel spectrogram. + """ + if torch.min(y) < -1.0: + print(f"[WARNING] Min value of input waveform signal is {torch.min(y)}") + if torch.max(y) > 1.0: + print(f"[WARNING] Max value of input waveform signal is {torch.max(y)}") + + device = y.device + key = f"{n_fft}_{num_mels}_{sampling_rate}_{hop_size}_{win_size}_{fmin}_{fmax}_{device}" + + if key not in mel_basis_cache: + mel = librosa_mel_fn( + sr=sampling_rate, n_fft=n_fft, n_mels=num_mels, fmin=fmin, fmax=fmax + ) + mel_basis_cache[key] = torch.from_numpy(mel).float().to(device) + hann_window_cache[key] = torch.hann_window(win_size).to(device) + + mel_basis = mel_basis_cache[key] + hann_window = hann_window_cache[key] + + padding = (n_fft - hop_size) // 2 + y = torch.nn.functional.pad( + y.unsqueeze(1), (padding, padding), mode="reflect" + ).squeeze(1) + + spec = torch.stft( + y, + n_fft, + hop_length=hop_size, + win_length=win_size, + window=hann_window, + center=center, + pad_mode="reflect", + normalized=False, + onesided=True, + return_complex=True, + ) + spec = torch.sqrt(torch.view_as_real(spec).pow(2).sum(-1) + 1e-9) + + mel_spec = torch.matmul(mel_basis, spec) + mel_spec = spectral_normalize_torch(mel_spec) + + return mel_spec + + +def get_mel_spectrogram(wav, h): + """ + Generate mel spectrogram from a waveform using given hyperparameters. + + Args: + wav (torch.Tensor): Input waveform. + h: Hyperparameters object with attributes n_fft, num_mels, sampling_rate, hop_size, win_size, fmin, fmax. + + Returns: + torch.Tensor: Mel spectrogram. + """ + return mel_spectrogram( + wav, + h.n_fft, + h.num_mels, + h.sampling_rate, + h.hop_size, + h.win_size, + h.fmin, + h.fmax, + ) + + +def get_dataset_filelist(a): + training_files = [] + validation_files = [] + list_unseen_validation_files = [] + + with open(a.input_training_file, "r", encoding="utf-8") as fi: + training_files = [ + os.path.join(a.input_wavs_dir, x.split("|")[0] + ".wav") + for x in fi.read().split("\n") + if len(x) > 0 + ] + print(f"first training file: {training_files[0]}") + + with open(a.input_validation_file, "r", encoding="utf-8") as fi: + validation_files = [ + os.path.join(a.input_wavs_dir, x.split("|")[0] + ".wav") + for x in fi.read().split("\n") + if len(x) > 0 + ] + print(f"first validation file: {validation_files[0]}") + + for i in range(len(a.list_input_unseen_validation_file)): + with open(a.list_input_unseen_validation_file[i], "r", encoding="utf-8") as fi: + unseen_validation_files = [ + os.path.join(a.list_input_unseen_wavs_dir[i], x.split("|")[0] + ".wav") + for x in fi.read().split("\n") + if len(x) > 0 + ] + print( + f"first unseen {i}th validation fileset: {unseen_validation_files[0]}" + ) + list_unseen_validation_files.append(unseen_validation_files) + + return training_files, validation_files, list_unseen_validation_files + + +class MelDataset(torch.utils.data.Dataset): + def __init__( + self, + training_files: List[str], + hparams: AttrDict, + segment_size: int, + n_fft: int, + num_mels: int, + hop_size: int, + win_size: int, + sampling_rate: int, + fmin: int, + fmax: Optional[int], + split: bool = True, + shuffle: bool = True, + device: str = None, + fmax_loss: Optional[int] = None, + fine_tuning: bool = False, + base_mels_path: str = None, + is_seen: bool = True, + ): + self.audio_files = training_files + random.seed(1234) + if shuffle: + random.shuffle(self.audio_files) + self.hparams = hparams + self.is_seen = is_seen + if self.is_seen: + self.name = pathlib.Path(self.audio_files[0]).parts[0] + else: + self.name = "-".join(pathlib.Path(self.audio_files[0]).parts[:2]).strip("/") + + self.segment_size = segment_size + self.sampling_rate = sampling_rate + self.split = split + self.n_fft = n_fft + self.num_mels = num_mels + self.hop_size = hop_size + self.win_size = win_size + self.fmin = fmin + self.fmax = fmax + self.fmax_loss = fmax_loss + self.device = device + self.fine_tuning = fine_tuning + self.base_mels_path = base_mels_path + + print("[INFO] checking dataset integrity...") + for i in tqdm(range(len(self.audio_files))): + assert os.path.exists( + self.audio_files[i] + ), f"{self.audio_files[i]} not found" + + def __getitem__( + self, index: int + ) -> Tuple[torch.Tensor, torch.Tensor, str, torch.Tensor]: + try: + filename = self.audio_files[index] + + # Use librosa.load that ensures loading waveform into mono with [-1, 1] float values + # Audio is ndarray with shape [T_time]. Disable auto-resampling here to minimize overhead + # The on-the-fly resampling during training will be done only for the obtained random chunk + audio, source_sampling_rate = librosa.load(filename, sr=None, mono=True) + + # Main logic that uses pair for training BigVGAN + if not self.fine_tuning: + if self.split: # Training step + # Obtain randomized audio chunk + if source_sampling_rate != self.sampling_rate: + # Adjust segment size to crop if the source sr is different + target_segment_size = math.ceil( + self.segment_size + * (source_sampling_rate / self.sampling_rate) + ) + else: + target_segment_size = self.segment_size + + # Compute upper bound index for the random chunk + random_chunk_upper_bound = max( + 0, audio.shape[0] - target_segment_size + ) + + # Crop or pad audio to obtain random chunk with target_segment_size + if audio.shape[0] >= target_segment_size: + audio_start = random.randint(0, random_chunk_upper_bound) + audio = audio[audio_start : audio_start + target_segment_size] + else: + audio = np.pad( + audio, + (0, target_segment_size - audio.shape[0]), + mode="constant", + ) + + # Resample audio chunk to self.sampling rate + if source_sampling_rate != self.sampling_rate: + audio = librosa.resample( + audio, + orig_sr=source_sampling_rate, + target_sr=self.sampling_rate, + ) + if audio.shape[0] > self.segment_size: + # trim last elements to match self.segment_size (e.g., 16385 for 44khz downsampled to 24khz -> 16384) + audio = audio[: self.segment_size] + + else: # Validation step + # Resample full audio clip to target sampling rate + if source_sampling_rate != self.sampling_rate: + audio = librosa.resample( + audio, + orig_sr=source_sampling_rate, + target_sr=self.sampling_rate, + ) + # Trim last elements to match audio length to self.hop_size * n for evaluation + if (audio.shape[0] % self.hop_size) != 0: + audio = audio[: -(audio.shape[0] % self.hop_size)] + + # BigVGAN is trained using volume-normalized waveform + audio = librosa.util.normalize(audio) * 0.95 + + # Cast ndarray to torch tensor + audio = torch.FloatTensor(audio) + audio = audio.unsqueeze(0) # [B(1), self.segment_size] + + # Compute mel spectrogram corresponding to audio + mel = mel_spectrogram( + audio, + self.n_fft, + self.num_mels, + self.sampling_rate, + self.hop_size, + self.win_size, + self.fmin, + self.fmax, + center=False, + ) # [B(1), self.num_mels, self.segment_size // self.hop_size] + + # Fine-tuning logic that uses pre-computed mel. Example: Using TTS model-generated mel as input + else: + # For fine-tuning, assert that the waveform is in the defined sampling_rate + # Fine-tuning won't support on-the-fly resampling to be fool-proof (the dataset should have been prepared properly) + assert ( + source_sampling_rate == self.sampling_rate + ), f"For fine_tuning, waveform must be in the spcified sampling rate {self.sampling_rate}, got {source_sampling_rate}" + + # Cast ndarray to torch tensor + audio = torch.FloatTensor(audio) + audio = audio.unsqueeze(0) # [B(1), T_time] + + # Load pre-computed mel from disk + mel = np.load( + os.path.join( + self.base_mels_path, + os.path.splitext(os.path.split(filename)[-1])[0] + ".npy", + ) + ) + mel = torch.from_numpy(mel) + + if len(mel.shape) < 3: + mel = mel.unsqueeze(0) # ensure [B, C, T] + + if self.split: + frames_per_seg = math.ceil(self.segment_size / self.hop_size) + + if audio.size(1) >= self.segment_size: + mel_start = random.randint(0, mel.size(2) - frames_per_seg - 1) + mel = mel[:, :, mel_start : mel_start + frames_per_seg] + audio = audio[ + :, + mel_start + * self.hop_size : (mel_start + frames_per_seg) + * self.hop_size, + ] + + # Pad pre-computed mel and audio to match length to ensuring fine-tuning without error. + # NOTE: this may introduce a single-frame misalignment of the + # To remove possible misalignment, it is recommended to prepare the pair where the audio length is the integer multiple of self.hop_size + mel = torch.nn.functional.pad( + mel, (0, frames_per_seg - mel.size(2)), "constant" + ) + audio = torch.nn.functional.pad( + audio, (0, self.segment_size - audio.size(1)), "constant" + ) + + # Compute mel_loss used by spectral regression objective. Uses self.fmax_loss instead (usually None) + mel_loss = mel_spectrogram( + audio, + self.n_fft, + self.num_mels, + self.sampling_rate, + self.hop_size, + self.win_size, + self.fmin, + self.fmax_loss, + center=False, + ) # [B(1), self.num_mels, self.segment_size // self.hop_size] + + # Shape sanity checks + assert ( + audio.shape[1] == mel.shape[2] * self.hop_size + and audio.shape[1] == mel_loss.shape[2] * self.hop_size + ), f"Audio length must be mel frame length * hop_size. Got audio shape {audio.shape} mel shape {mel.shape} mel_loss shape {mel_loss.shape}" + + return (mel.squeeze(), audio.squeeze(0), filename, mel_loss.squeeze()) + + # If it encounters error during loading the data, skip this sample and load random other sample to the batch + except Exception as e: + if self.fine_tuning: + raise e # Terminate training if it is fine-tuning. The dataset should have been prepared properly. + else: + print( + f"[WARNING] Failed to load waveform, skipping! filename: {filename} Error: {e}" + ) + return self[random.randrange(len(self))] + + def __len__(self): + return len(self.audio_files) diff --git a/BigVGAN/nv-modelcard++/.gitkeep b/BigVGAN/nv-modelcard++/.gitkeep new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/BigVGAN/nv-modelcard++/bias.md b/BigVGAN/nv-modelcard++/bias.md new file mode 100644 index 0000000000000000000000000000000000000000..4b388c28d09b8ca3aab5096304c52e1a5dac0e16 --- /dev/null +++ b/BigVGAN/nv-modelcard++/bias.md @@ -0,0 +1,4 @@ +| Field | Response | +| :--------------------------------------------------------------------------------------------------------- | :--------------------------------------------------- | +| Participation considerations from adversely impacted groups protected classes in model design and testing: | None | +| Measures taken to mitigate against unwanted bias: | No measures taken to mitigate against unwanted bias. | diff --git a/BigVGAN/nv-modelcard++/explainability.md b/BigVGAN/nv-modelcard++/explainability.md new file mode 100644 index 0000000000000000000000000000000000000000..6f1a16676e438ba95f9d411a19e04a0f13409e54 --- /dev/null +++ b/BigVGAN/nv-modelcard++/explainability.md @@ -0,0 +1,13 @@ +| Field | Response | +| :---------------------------------------------------------------------------------------------------- | :--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | +| Intended Application & Domain: | Generating waveform from mel spectrogram. | +| Model Type: | Convolutional Neural Network (CNN) | +| Intended Users: | This model is intended for developers to synthesize and generate waveforms from the AI-generated mel spectrograms. | +| Output: | Audio Waveform | +| Describe how the model works: | Model generates audio waveform corresponding to the input mel spectrogram. | +| Name the adversely impacted groups this has been tested to deliver comparable outcomes regardless of: | Not Applicable | +| Technical Limitations: | This may not perform well on synthetically-generated mel spectrograms that deviate significantly from the profile of mel spectrograms on which this was trained. | +| Verified to have met prescribed NVIDIA quality standards: | Yes | +| Performance Metrics: | Perceptual Evaluation of Speech Quality (PESQ), Virtual Speech Quality Objective Listener (VISQOL), Multi-resolution STFT (MRSTFT), Mel cepstral distortion (MCD), Periodicity RMSE, Voice/Unvoiced F1 Score (V/UV F1) | +| Potential Known Risks: | This model may generate low-quality or distorted soundwaves. | +| Licensing: | https://github.com/NVIDIA/BigVGAN/blob/main/LICENSE | diff --git a/BigVGAN/nv-modelcard++/overview.md b/BigVGAN/nv-modelcard++/overview.md new file mode 100644 index 0000000000000000000000000000000000000000..a39cba0b49a4a32a37afa90f2baf4630dcd9cadc --- /dev/null +++ b/BigVGAN/nv-modelcard++/overview.md @@ -0,0 +1,126 @@ +# Model Overview + +## Description: + +BigVGAN is a generative AI model specialized in synthesizing audio waveforms using Mel spectrogram as inputs. + +
+ +BigVGAN is a fully convolutional architecture with several upsampling blocks using transposed convolution followed by multiple residual dilated convolution layers. + +BigVGAN consists of a novel module, called anti-aliased multi-periodicity composition (AMP), which is specifically designed for generating waveforms. AMP is specialized in synthesizing high-frequency and periodic soundwaves drawing inspiration from audio signal processing principles. + +It applies a periodic activation function, called Snake, which provides an inductive bias to the architecture in generating periodic soundwaves. It also applies anti-aliasing filters to reduce undesired artifacts in the generated waveforms.
+ +This model is ready for commercial use.
+ +## References(s): + +- [BigVGAN: A Universal Neural Vocoder with Large-Scale Training](https://arxiv.org/abs/2206.04658)
+- [Project Page](https://research.nvidia.com/labs/adlr/projects/bigvgan/)
+- [Audio Demo](https://bigvgan-demo.github.io/)
+ +## Model Architecture: + +**Architecture Type:** Convolution Neural Network (CNN)
+**Network Architecture:** You can see the details of this model on this link: https://github.com/NVIDIA/BigVGAN and the related paper can be found here: https://arxiv.org/abs/2206.04658
+**Model Version:** 2.0
+ +## Input: + +**Input Type:** Audio
+**Input Format:** Mel Spectrogram
+**Input Parameters:** None
+**Other Properties Related to Input:** The input mel spectrogram has shape `[batch, channels, frames]`, where `channels` refers to the number of mel bands defined by the model and `frames` refers to the temporal length. The model supports arbitrary long `frames` that fits into the GPU memory. + +## Output: + +**Input Type:** Audio
+**Output Format:** Audio Waveform
+**Output Parameters:** None
+**Other Properties Related to Output:** The output audio waveform has shape `[batch, 1, time]`, where `1` refers to the mono audio channels and `time` refers to the temporal length. `time` is defined as a fixed integer multiple of input `frames`, which is an upsampling ratio of the model (`time = upsampling ratio * frames`). The output audio waveform consitutes float values with a range of `[-1, 1]`. + +## Software Integration: + +**Runtime Engine(s):** PyTorch + +**Supported Hardware Microarchitecture Compatibility:** NVIDIA Ampere, NVIDIA Hopper, NVIDIA Lovelace, NVIDIA Turing, NVIDIA Volta
+ +## Preferred/Supported Operating System(s): + +Linux + +## Model Version(s): + +v2.0 + +## Training, Testing, and Evaluation Datasets: + +### Training Dataset: + +The dataset contains diverse audio types, including speech in multiple languages, environmental sounds, and instruments. + +**Links:** + +- [AAM: Artificial Audio Multitracks Dataset](https://zenodo.org/records/5794629) +- [AudioCaps](https://audiocaps.github.io/) +- [AudioSet](https://research.google.com/audioset/index.html) +- [common-accent](https://huggingface.co/datasets/DTU54DL/common-accent) +- [Crowd Sourced Emotional Multimodal Actors Dataset (CREMA-D)](https://ieeexplore.ieee.org/document/6849440) +- [DCASE2017 Challenge, Task 4: Large-scale weakly supervised sound event detection for smart cars](https://dcase.community/challenge2017/task-large-scale-sound-event-detection) +- [FSDnoisy18k](https://zenodo.org/records/2529934) +- [Free Universal Sound Separation Dataset](https://zenodo.org/records/3694384) +- [Greatest Hits dataset](https://andrewowens.com/vis/) +- [GTZAN](https://ieeexplore.ieee.org/document/1021072) +- [JL corpus](https://www.kaggle.com/datasets/tli725/jl-corpus) +- [Medley-solos-DB: a cross-collection dataset for musical instrument recognition](https://zenodo.org/records/3464194) +- [MUSAN: A Music, Speech, and Noise Corpus](https://www.openslr.org/17/) +- [MusicBench](https://huggingface.co/datasets/amaai-lab/MusicBench) +- [MusicCaps](https://www.kaggle.com/datasets/googleai/musiccaps) +- [MusicNet](https://www.kaggle.com/datasets/imsparsh/musicnet-dataset) +- [NSynth](https://magenta.tensorflow.org/datasets/nsynth) +- [OnAir-Music-Dataset](https://github.com/sevagh/OnAir-Music-Dataset) +- [Audio Piano Triads Dataset](https://zenodo.org/records/4740877) +- [Pitch Audio Dataset (Surge synthesizer)](https://zenodo.org/records/4677097) +- [SONYC Urban Sound Tagging (SONYC-UST): a multilabel dataset from an urban acoustic sensor network](https://zenodo.org/records/3966543) +- [VocalSound: A Dataset for Improving Human Vocal Sounds Recognition](https://arxiv.org/abs/2205.03433) +- [WavText5K](https://github.com/microsoft/WavText5K) +- [CSS10: A Collection of Single Speaker Speech Datasets for 10 Languages](https://github.com/Kyubyong/css10) +- [Hi-Fi Multi-Speaker English TTS Dataset (Hi-Fi TTS)](https://www.openslr.org/109/) +- [IIIT-H Indic Speech Databases](http://festvox.org/databases/iiit_voices/) +- [Libri-Light: A Benchmark for ASR with Limited or No Supervision](https://arxiv.org/abs/1912.07875) +- [LibriTTS: A Corpus Derived from LibriSpeech for Text-to-Speech](https://www.openslr.org/60) +- [LibriTTS-R: A Restored Multi-Speaker Text-to-Speech Corpus](https://www.openslr.org/141/) +- [The SIWIS French Speech Synthesis Database](https://datashare.ed.ac.uk/handle/10283/2353) +- [Crowdsourced high-quality Colombian Spanish speech data set](https://openslr.org/72/) +- [TTS-Portuguese Corpus](https://github.com/Edresson/TTS-Portuguese-Corpus) +- [CSTR VCTK Corpus: English Multi-speaker Corpus for CSTR Voice Cloning Toolkit](https://datashare.ed.ac.uk/handle/10283/3443) + +\*\* Data Collection Method by dataset
+ +- Human
+ +\*\* Labeling Method by dataset (for those with labels)
+ +- Hybrid: Automated, Human, Unknown
+ +### Evaluating Dataset: + +Properties: The audio generation quality of BigVGAN is evaluated using `dev` splits of the [LibriTTS dataset](https://www.openslr.org/60/) and [Hi-Fi TTS dataset](https://www.openslr.org/109/). The datasets include speech in English language with equal balance of genders. + +\*\* Data Collection Method by dataset
+ +- Human
+ +\*\* Labeling Method by dataset
+ +- Automated
+ +## Inference: + +**Engine:** PyTorch
+**Test Hardware:** NVIDIA A100 GPU
+ +## Ethical Considerations: + +NVIDIA believes Trustworthy AI is a shared responsibility and we have established policies and practices to enable development for a wide array of AI applications. When downloaded or used in accordance with our terms of service, developers should work with their internal model team to ensure this model meets requirements for the relevant industry and use case and addresses unforeseen product misuse. For more detailed information on ethical considerations for this model, please see the Model Card++ Explainability, Bias, Safety & Security, and Privacy Subcards. Please report security vulnerabilities or NVIDIA AI Concerns [here](https://www.nvidia.com/en-us/support/submit-security-vulnerability/). diff --git a/BigVGAN/nv-modelcard++/privacy.md b/BigVGAN/nv-modelcard++/privacy.md new file mode 100644 index 0000000000000000000000000000000000000000..73554a998384ca1b1050239ebd51bda46aec1878 --- /dev/null +++ b/BigVGAN/nv-modelcard++/privacy.md @@ -0,0 +1,14 @@ +| Field | Response | +| :------------------------------------------------------------------------------------------------------------------------------------- | :--------------------------------------------- | +| Generatable or reverse engineerable personal information? | None | +| Protected class data used to create this model? | None | +| Was consent obtained for any personal data used? | Not Applicable (No Personal Data) | +| How often is dataset reviewed? | Before Release | +| Is a mechanism in place to honor data subject right of access or deletion of personal data? | Not Applicable | +| If personal collected for the development of the model, was it collected directly by NVIDIA? | Not Applicable | +| If personal collected for the development of the model by NVIDIA, do you maintain or have access to disclosures made to data subjects? | Not Applicable | +| If personal collected for the development of this AI model, was it minimized to only what was required? | Not Applicable | +| Is data in dataset traceable? | Yes | +| Is there provenance for all datasets used in training? | Yes | +| Does data labeling (annotation, metadata) comply with privacy laws? | Yes | +| Is data compliant with data subject requests for data correction or removal, if such a request was made? | No, not possible with externally-sourced data. | diff --git a/BigVGAN/nv-modelcard++/safety.md b/BigVGAN/nv-modelcard++/safety.md new file mode 100644 index 0000000000000000000000000000000000000000..ed30370dfedbbb49748706034a7153d54f1a668f --- /dev/null +++ b/BigVGAN/nv-modelcard++/safety.md @@ -0,0 +1,6 @@ +| Field | Response | +| :---------------------------------------------- | :---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | +| Model Application(s): | Synethic Audio Generation | +| Describe the life critical impact (if present). | Not Applicable | +| Use Case Restrictions: | None | +| Model and dataset restrictions: | The Principle of least privilege (PoLP) is applied limiting access for dataset generation and model development. Restrictions enforce dataset access during training, and dataset license constraints adhered to. | diff --git a/BigVGAN/requirements.txt b/BigVGAN/requirements.txt new file mode 100644 index 0000000000000000000000000000000000000000..6e61d3203966612e6ad193bbabdef10b1d3fed84 --- /dev/null +++ b/BigVGAN/requirements.txt @@ -0,0 +1,13 @@ +torch +numpy +librosa>=0.8.1 +scipy +tensorboard +soundfile +matplotlib +pesq +auraloss +tqdm +nnAudio +ninja +huggingface_hub>=0.23.4 \ No newline at end of file diff --git a/BigVGAN/tests/test_activation.py b/BigVGAN/tests/test_activation.py new file mode 100644 index 0000000000000000000000000000000000000000..146600e5b52689851d055c949938f100bd56ed4f --- /dev/null +++ b/BigVGAN/tests/test_activation.py @@ -0,0 +1,65 @@ +# Copyright (c) 2024 NVIDIA CORPORATION. +# Licensed under the MIT license. + +import os +import sys +# to import modules from parent_dir +parent_dir = os.path.abspath(os.path.join(os.path.dirname(__file__), "..")) +sys.path.append(parent_dir) + +import torch +from alias_free_activation.cuda import activation1d +from activations import Snake + + +def test_load_fused_kernels(): + try: + print("[Success] load_fused_kernels") + except ImportError as e: + print("[Fail] load_fused_kernels") + raise e + + +def test_anti_alias_activation(): + data = torch.rand((10, 10, 200), device="cuda") + + # Check activations.Snake cuda vs. torch + fused_anti_alias_activation = activation1d.Activation1d( + activation=Snake(10), fused=True + ).cuda() + fused_activation_output = fused_anti_alias_activation(data) + + torch_anti_alias_activation = activation1d.Activation1d( + activation=Snake(10), fused=False + ).cuda() + torch_activation_output = torch_anti_alias_activation(data) + + test_result = (fused_activation_output - torch_activation_output).abs() + + while test_result.dim() != 1: + test_result = test_result.mean(dim=-1) + + diff = test_result.mean(dim=-1) + + if diff <= 1e-3: + print( + f"\n[Success] test_fused_anti_alias_activation" + f"\n > mean_difference={diff}" + f"\n > fused_values={fused_activation_output[-1][-1][:].tolist()}" + f"\n > torch_values={torch_activation_output[-1][-1][:].tolist()}" + ) + else: + print( + f"\n[Fail] test_fused_anti_alias_activation" + f"\n > mean_difference={diff}, " + f"\n > fused_values={fused_activation_output[-1][-1][:].tolist()}, " + f"\n > torch_values={torch_activation_output[-1][-1][:].tolist()}" + ) + + +if __name__ == "__main__": + from alias_free_activation.cuda import load + + load.load() + test_load_fused_kernels() + test_anti_alias_activation() diff --git a/BigVGAN/tests/test_activation_snake_beta.py b/BigVGAN/tests/test_activation_snake_beta.py new file mode 100644 index 0000000000000000000000000000000000000000..3e653859e931501ec204ee57c8c03910bb2a695b --- /dev/null +++ b/BigVGAN/tests/test_activation_snake_beta.py @@ -0,0 +1,66 @@ +# Copyright (c) 2024 NVIDIA CORPORATION. +# Licensed under the MIT license. + +import os +import sys +# to import modules from parent_dir +parent_dir = os.path.abspath(os.path.join(os.path.dirname(__file__), "..")) +sys.path.append(parent_dir) + +import torch +from alias_free_activation.cuda import activation1d +from activations import SnakeBeta + + +def test_load_fused_kernels(): + try: + print("[Success] load_fused_kernels") + except ImportError as e: + print("[Fail] load_fused_kernels") + raise e + + +def test_anti_alias_activation(): + data = torch.rand((10, 10, 200), device="cuda") + + # Check activations, Snake CUDA vs. Torch + fused_anti_alias_activation = activation1d.Activation1d( + activation=SnakeBeta(10), fused=True + ).cuda() + fused_activation_output = fused_anti_alias_activation(data) + + torch_anti_alias_activation = activation1d.Activation1d( + activation=SnakeBeta(10), fused=False + ).cuda() + torch_activation_output = torch_anti_alias_activation(data) + + test_result = (fused_activation_output - torch_activation_output).abs() + + while test_result.dim() != 1: + test_result = test_result.mean(dim=-1) + + diff = test_result.mean(dim=-1) + + if diff <= 1e-3: + print( + f"\n[Success] test_fused_anti_alias_activation" + f"\n > mean_difference={diff}" + f"\n > fused_values={fused_activation_output[-1][-1][:].tolist()}" + f"\n > torch_values={torch_activation_output[-1][-1][:].tolist()}" + ) + else: + print( + f"\n[Fail] test_fused_anti_alias_activation" + f"\n > mean_difference={diff}, " + f"\n > fused_values={fused_activation_output[-1][-1][:].tolist()}, " + f"\n > torch_values={torch_activation_output[-1][-1][:].tolist()}" + ) + + + +if __name__ == "__main__": + from alias_free_activation.cuda import load + + load.load() + test_load_fused_kernels() + test_anti_alias_activation() diff --git a/BigVGAN/tests/test_cuda_vs_torch_model.py b/BigVGAN/tests/test_cuda_vs_torch_model.py new file mode 100644 index 0000000000000000000000000000000000000000..86ad0518d5925a7d14afb6713435b91ba851c1d6 --- /dev/null +++ b/BigVGAN/tests/test_cuda_vs_torch_model.py @@ -0,0 +1,221 @@ +# Copyright (c) 2024 NVIDIA CORPORATION. +# Licensed under the MIT license. + +import os +import sys + +# to import modules from parent_dir +parent_dir = os.path.abspath(os.path.join(os.path.dirname(__file__), "..")) +sys.path.append(parent_dir) + +import torch +import json +from env import AttrDict +from bigvgan import BigVGAN +from time import time +from tqdm import tqdm +from meldataset import mel_spectrogram, MAX_WAV_VALUE +from scipy.io.wavfile import write +import numpy as np + +import argparse + +torch.backends.cudnn.benchmark = True + +# For easier debugging +torch.set_printoptions(linewidth=200, threshold=10_000) + + +def generate_soundwave(duration=5.0, sr=24000): + t = np.linspace(0, duration, int(sr * duration), False, dtype=np.float32) + + modulation = np.sin(2 * np.pi * t / duration) + + min_freq = 220 + max_freq = 1760 + frequencies = min_freq + (max_freq - min_freq) * (modulation + 1) / 2 + soundwave = np.sin(2 * np.pi * frequencies * t) + + soundwave = soundwave / np.max(np.abs(soundwave)) * 0.95 + + return soundwave, sr + + +def get_mel(x, h): + return mel_spectrogram( + x, h.n_fft, h.num_mels, h.sampling_rate, h.hop_size, h.win_size, h.fmin, h.fmax + ) + + +def load_checkpoint(filepath, device): + assert os.path.isfile(filepath) + print(f"Loading '{filepath}'") + checkpoint_dict = torch.load(filepath, map_location=device) + print("Complete.") + return checkpoint_dict + + +if __name__ == "__main__": + parser = argparse.ArgumentParser( + description="Test script to check CUDA kernel correctness." + ) + parser.add_argument( + "--checkpoint_file", + type=str, + required=True, + help="Path to the checkpoint file. Assumes config.json exists in the directory.", + ) + + args = parser.parse_args() + + config_file = os.path.join(os.path.split(args.checkpoint_file)[0], "config.json") + with open(config_file) as f: + config = f.read() + json_config = json.loads(config) + h = AttrDict({**json_config}) + + print("loading plain Pytorch BigVGAN") + generator_original = BigVGAN(h).to("cuda") + print("loading CUDA kernel BigVGAN with auto-build") + generator_cuda_kernel = BigVGAN(h, use_cuda_kernel=True).to("cuda") + + state_dict_g = load_checkpoint(args.checkpoint_file, "cuda") + generator_original.load_state_dict(state_dict_g["generator"]) + generator_cuda_kernel.load_state_dict(state_dict_g["generator"]) + + generator_original.remove_weight_norm() + generator_original.eval() + generator_cuda_kernel.remove_weight_norm() + generator_cuda_kernel.eval() + + # define number of samples and length of mel frame to benchmark + num_sample = 10 + num_mel_frame = 16384 + + # CUDA kernel correctness check + diff = 0.0 + for i in tqdm(range(num_sample)): + # Random mel + data = torch.rand((1, h.num_mels, num_mel_frame), device="cuda") + + with torch.inference_mode(): + audio_original = generator_original(data) + + with torch.inference_mode(): + audio_cuda_kernel = generator_cuda_kernel(data) + + # Both outputs should be (almost) the same + test_result = (audio_original - audio_cuda_kernel).abs() + diff += test_result.mean(dim=-1).item() + + diff /= num_sample + if ( + diff <= 2e-3 + ): # We can expect a small difference (~1e-3) which does not affect perceptual quality + print( + f"\n[Success] test CUDA fused vs. plain torch BigVGAN inference" + f"\n > mean_difference={diff}" + f"\n > fused_values={audio_cuda_kernel[-1][-1][-30:].tolist()}" + f"\n > torch_values={audio_original[-1][-1][-30:].tolist()}" + ) + else: + print( + f"\n[Fail] test CUDA fused vs. plain torch BigVGAN inference" + f"\n > mean_difference={diff}" + f"\n > fused_values={audio_cuda_kernel[-1][-1][-30:].tolist()}, " + f"\n > torch_values={audio_original[-1][-1][-30:].tolist()}" + ) + + del data, audio_original, audio_cuda_kernel + + # Variables for tracking total time and VRAM usage + toc_total_original = 0 + toc_total_cuda_kernel = 0 + vram_used_original_total = 0 + vram_used_cuda_kernel_total = 0 + audio_length_total = 0 + + # Measure Original inference in isolation + for i in tqdm(range(num_sample)): + torch.cuda.reset_peak_memory_stats(device="cuda") + data = torch.rand((1, h.num_mels, num_mel_frame), device="cuda") + torch.cuda.synchronize() + tic = time() + with torch.inference_mode(): + audio_original = generator_original(data) + torch.cuda.synchronize() + toc = time() - tic + toc_total_original += toc + + vram_used_original_total += torch.cuda.max_memory_allocated(device="cuda") + + del data, audio_original + torch.cuda.empty_cache() + + # Measure CUDA kernel inference in isolation + for i in tqdm(range(num_sample)): + torch.cuda.reset_peak_memory_stats(device="cuda") + data = torch.rand((1, h.num_mels, num_mel_frame), device="cuda") + torch.cuda.synchronize() + tic = time() + with torch.inference_mode(): + audio_cuda_kernel = generator_cuda_kernel(data) + torch.cuda.synchronize() + toc = time() - tic + toc_total_cuda_kernel += toc + + audio_length_total += audio_cuda_kernel.shape[-1] + + vram_used_cuda_kernel_total += torch.cuda.max_memory_allocated(device="cuda") + + del data, audio_cuda_kernel + torch.cuda.empty_cache() + + # Calculate metrics + audio_second = audio_length_total / h.sampling_rate + khz_original = audio_length_total / toc_total_original / 1000 + khz_cuda_kernel = audio_length_total / toc_total_cuda_kernel / 1000 + vram_used_original_gb = vram_used_original_total / num_sample / (1024 ** 3) + vram_used_cuda_kernel_gb = vram_used_cuda_kernel_total / num_sample / (1024 ** 3) + + # Print results + print( + f"Original BigVGAN: took {toc_total_original:.2f} seconds to generate {audio_second:.2f} seconds of audio, {khz_original:.1f}kHz, {audio_second / toc_total_original:.1f} faster than realtime, VRAM used {vram_used_original_gb:.1f} GB" + ) + print( + f"CUDA kernel BigVGAN: took {toc_total_cuda_kernel:.2f} seconds to generate {audio_second:.2f} seconds of audio, {khz_cuda_kernel:.1f}kHz, {audio_second / toc_total_cuda_kernel:.1f} faster than realtime, VRAM used {vram_used_cuda_kernel_gb:.1f} GB" + ) + print(f"speedup of CUDA kernel: {khz_cuda_kernel / khz_original}") + print(f"VRAM saving of CUDA kernel: {vram_used_original_gb / vram_used_cuda_kernel_gb}") + + # Use artificial sine waves for inference test + audio_real, sr = generate_soundwave(duration=5.0, sr=h.sampling_rate) + audio_real = torch.tensor(audio_real).to("cuda") + # Compute mel spectrogram from the ground truth audio + x = get_mel(audio_real.unsqueeze(0), h) + + with torch.inference_mode(): + y_g_hat_original = generator_original(x) + y_g_hat_cuda_kernel = generator_cuda_kernel(x) + + audio_real = audio_real.squeeze() + audio_real = audio_real * MAX_WAV_VALUE + audio_real = audio_real.cpu().numpy().astype("int16") + + audio_original = y_g_hat_original.squeeze() + audio_original = audio_original * MAX_WAV_VALUE + audio_original = audio_original.cpu().numpy().astype("int16") + + audio_cuda_kernel = y_g_hat_cuda_kernel.squeeze() + audio_cuda_kernel = audio_cuda_kernel * MAX_WAV_VALUE + audio_cuda_kernel = audio_cuda_kernel.cpu().numpy().astype("int16") + + os.makedirs("tmp", exist_ok=True) + output_file_real = os.path.join("tmp", "audio_real.wav") + output_file_original = os.path.join("tmp", "audio_generated_original.wav") + output_file_cuda_kernel = os.path.join("tmp", "audio_generated_cuda_kernel.wav") + write(output_file_real, h.sampling_rate, audio_real) + write(output_file_original, h.sampling_rate, audio_original) + write(output_file_cuda_kernel, h.sampling_rate, audio_cuda_kernel) + print("Example generated audios of original vs. fused CUDA kernel written to tmp!") + print("Done") diff --git a/BigVGAN/train.py b/BigVGAN/train.py new file mode 100644 index 0000000000000000000000000000000000000000..01eeb095c9994895868a40905713b5d043eec396 --- /dev/null +++ b/BigVGAN/train.py @@ -0,0 +1,777 @@ +# Copyright (c) 2024 NVIDIA CORPORATION. +# Licensed under the MIT license. + +# Adapted from https://github.com/jik876/hifi-gan under the MIT license. +# LICENSE is in incl_licenses directory. + + +import warnings + +warnings.simplefilter(action="ignore", category=FutureWarning) +import itertools +import os +import time +import argparse +import json +import torch +import torch.nn.functional as F +from torch.utils.tensorboard import SummaryWriter +from torch.utils.data import DistributedSampler, DataLoader +import torch.multiprocessing as mp +from torch.distributed import init_process_group +from torch.nn.parallel import DistributedDataParallel +from env import AttrDict, build_env +from meldataset import MelDataset, mel_spectrogram, get_dataset_filelist, MAX_WAV_VALUE + +from bigvgan import BigVGAN +from discriminators import ( + MultiPeriodDiscriminator, + MultiResolutionDiscriminator, + MultiBandDiscriminator, + MultiScaleSubbandCQTDiscriminator, +) +from loss import ( + feature_loss, + generator_loss, + discriminator_loss, + MultiScaleMelSpectrogramLoss, +) + +from utils import ( + plot_spectrogram, + plot_spectrogram_clipped, + scan_checkpoint, + load_checkpoint, + save_checkpoint, + save_audio, +) +import torchaudio as ta +from pesq import pesq +from tqdm import tqdm +import auraloss + +torch.backends.cudnn.benchmark = False + + +def train(rank, a, h): + if h.num_gpus > 1: + # initialize distributed + init_process_group( + backend=h.dist_config["dist_backend"], + init_method=h.dist_config["dist_url"], + world_size=h.dist_config["world_size"] * h.num_gpus, + rank=rank, + ) + + # Set seed and device + torch.cuda.manual_seed(h.seed) + torch.cuda.set_device(rank) + device = torch.device(f"cuda:{rank:d}") + + # Define BigVGAN generator + generator = BigVGAN(h).to(device) + + # Define discriminators. MPD is used by default + mpd = MultiPeriodDiscriminator(h).to(device) + + # Define additional discriminators. BigVGAN-v1 uses UnivNet's MRD as default + # New in BigVGAN-v2: option to switch to new discriminators: MultiBandDiscriminator / MultiScaleSubbandCQTDiscriminator + if h.get("use_mbd_instead_of_mrd", False): # Switch to MBD + print( + "[INFO] using MultiBandDiscriminator of BigVGAN-v2 instead of MultiResolutionDiscriminator" + ) + # Variable name is kept as "mrd" for backward compatibility & minimal code change + mrd = MultiBandDiscriminator(h).to(device) + elif h.get("use_cqtd_instead_of_mrd", False): # Switch to CQTD + print( + "[INFO] using MultiScaleSubbandCQTDiscriminator of BigVGAN-v2 instead of MultiResolutionDiscriminator" + ) + mrd = MultiScaleSubbandCQTDiscriminator(h).to(device) + else: # Fallback to original MRD in BigVGAN-v1 + mrd = MultiResolutionDiscriminator(h).to(device) + + # New in BigVGAN-v2: option to switch to multi-scale L1 mel loss + if h.get("use_multiscale_melloss", False): + print( + "[INFO] using multi-scale Mel l1 loss of BigVGAN-v2 instead of the original single-scale loss" + ) + fn_mel_loss_multiscale = MultiScaleMelSpectrogramLoss( + sampling_rate=h.sampling_rate + ) # NOTE: accepts waveform as input + else: + fn_mel_loss_singlescale = F.l1_loss + + # Print the model & number of parameters, and create or scan the latest checkpoint from checkpoints directory + if rank == 0: + print(generator) + print(mpd) + print(mrd) + print(f"Generator params: {sum(p.numel() for p in generator.parameters())}") + print(f"Discriminator mpd params: {sum(p.numel() for p in mpd.parameters())}") + print(f"Discriminator mrd params: {sum(p.numel() for p in mrd.parameters())}") + os.makedirs(a.checkpoint_path, exist_ok=True) + print(f"Checkpoints directory: {a.checkpoint_path}") + + if os.path.isdir(a.checkpoint_path): + # New in v2.1: If the step prefix pattern-based checkpoints are not found, also check for renamed files in Hugging Face Hub to resume training + cp_g = scan_checkpoint( + a.checkpoint_path, prefix="g_", renamed_file="bigvgan_generator.pt" + ) + cp_do = scan_checkpoint( + a.checkpoint_path, + prefix="do_", + renamed_file="bigvgan_discriminator_optimizer.pt", + ) + + # Load the latest checkpoint if exists + steps = 0 + if cp_g is None or cp_do is None: + state_dict_do = None + last_epoch = -1 + else: + state_dict_g = load_checkpoint(cp_g, device) + state_dict_do = load_checkpoint(cp_do, device) + generator.load_state_dict(state_dict_g["generator"]) + mpd.load_state_dict(state_dict_do["mpd"]) + mrd.load_state_dict(state_dict_do["mrd"]) + steps = state_dict_do["steps"] + 1 + last_epoch = state_dict_do["epoch"] + + # Initialize DDP, optimizers, and schedulers + if h.num_gpus > 1: + generator = DistributedDataParallel(generator, device_ids=[rank]).to(device) + mpd = DistributedDataParallel(mpd, device_ids=[rank]).to(device) + mrd = DistributedDataParallel(mrd, device_ids=[rank]).to(device) + + optim_g = torch.optim.AdamW( + generator.parameters(), h.learning_rate, betas=[h.adam_b1, h.adam_b2] + ) + optim_d = torch.optim.AdamW( + itertools.chain(mrd.parameters(), mpd.parameters()), + h.learning_rate, + betas=[h.adam_b1, h.adam_b2], + ) + + if state_dict_do is not None: + optim_g.load_state_dict(state_dict_do["optim_g"]) + optim_d.load_state_dict(state_dict_do["optim_d"]) + + scheduler_g = torch.optim.lr_scheduler.ExponentialLR( + optim_g, gamma=h.lr_decay, last_epoch=last_epoch + ) + scheduler_d = torch.optim.lr_scheduler.ExponentialLR( + optim_d, gamma=h.lr_decay, last_epoch=last_epoch + ) + + # Define training and validation datasets + + """ + unseen_validation_filelist will contain sample filepaths outside the seen training & validation dataset + Example: trained on LibriTTS, validate on VCTK + """ + training_filelist, validation_filelist, list_unseen_validation_filelist = ( + get_dataset_filelist(a) + ) + + trainset = MelDataset( + training_filelist, + h, + h.segment_size, + h.n_fft, + h.num_mels, + h.hop_size, + h.win_size, + h.sampling_rate, + h.fmin, + h.fmax, + shuffle=False if h.num_gpus > 1 else True, + fmax_loss=h.fmax_for_loss, + device=device, + fine_tuning=a.fine_tuning, + base_mels_path=a.input_mels_dir, + is_seen=True, + ) + + train_sampler = DistributedSampler(trainset) if h.num_gpus > 1 else None + + train_loader = DataLoader( + trainset, + num_workers=h.num_workers, + shuffle=False, + sampler=train_sampler, + batch_size=h.batch_size, + pin_memory=True, + drop_last=True, + ) + + if rank == 0: + validset = MelDataset( + validation_filelist, + h, + h.segment_size, + h.n_fft, + h.num_mels, + h.hop_size, + h.win_size, + h.sampling_rate, + h.fmin, + h.fmax, + False, + False, + fmax_loss=h.fmax_for_loss, + device=device, + fine_tuning=a.fine_tuning, + base_mels_path=a.input_mels_dir, + is_seen=True, + ) + validation_loader = DataLoader( + validset, + num_workers=1, + shuffle=False, + sampler=None, + batch_size=1, + pin_memory=True, + drop_last=True, + ) + + list_unseen_validset = [] + list_unseen_validation_loader = [] + for i in range(len(list_unseen_validation_filelist)): + unseen_validset = MelDataset( + list_unseen_validation_filelist[i], + h, + h.segment_size, + h.n_fft, + h.num_mels, + h.hop_size, + h.win_size, + h.sampling_rate, + h.fmin, + h.fmax, + False, + False, + fmax_loss=h.fmax_for_loss, + device=device, + fine_tuning=a.fine_tuning, + base_mels_path=a.input_mels_dir, + is_seen=False, + ) + unseen_validation_loader = DataLoader( + unseen_validset, + num_workers=1, + shuffle=False, + sampler=None, + batch_size=1, + pin_memory=True, + drop_last=True, + ) + list_unseen_validset.append(unseen_validset) + list_unseen_validation_loader.append(unseen_validation_loader) + + # Tensorboard logger + sw = SummaryWriter(os.path.join(a.checkpoint_path, "logs")) + if a.save_audio: # Also save audio to disk if --save_audio is set to True + os.makedirs(os.path.join(a.checkpoint_path, "samples"), exist_ok=True) + + """ + Validation loop, "mode" parameter is automatically defined as (seen or unseen)_(name of the dataset). + If the name of the dataset contains "nonspeech", it skips PESQ calculation to prevent errors + """ + + def validate(rank, a, h, loader, mode="seen"): + assert rank == 0, "validate should only run on rank=0" + generator.eval() + torch.cuda.empty_cache() + + val_err_tot = 0 + val_pesq_tot = 0 + val_mrstft_tot = 0 + + # Modules for evaluation metrics + pesq_resampler = ta.transforms.Resample(h.sampling_rate, 16000).cuda() + loss_mrstft = auraloss.freq.MultiResolutionSTFTLoss(device="cuda") + + if a.save_audio: # Also save audio to disk if --save_audio is set to True + os.makedirs( + os.path.join(a.checkpoint_path, "samples", f"gt_{mode}"), + exist_ok=True, + ) + os.makedirs( + os.path.join(a.checkpoint_path, "samples", f"{mode}_{steps:08d}"), + exist_ok=True, + ) + + with torch.no_grad(): + print(f"step {steps} {mode} speaker validation...") + + # Loop over validation set and compute metrics + for j, batch in enumerate(tqdm(loader)): + x, y, _, y_mel = batch + y = y.to(device) + if hasattr(generator, "module"): + y_g_hat = generator.module(x.to(device)) + else: + y_g_hat = generator(x.to(device)) + y_mel = y_mel.to(device, non_blocking=True) + y_g_hat_mel = mel_spectrogram( + y_g_hat.squeeze(1), + h.n_fft, + h.num_mels, + h.sampling_rate, + h.hop_size, + h.win_size, + h.fmin, + h.fmax_for_loss, + ) + min_t = min(y_mel.size(-1), y_g_hat_mel.size(-1)) + val_err_tot += F.l1_loss(y_mel[...,:min_t], y_g_hat_mel[...,:min_t]).item() + + # PESQ calculation. only evaluate PESQ if it's speech signal (nonspeech PESQ will error out) + if ( + not "nonspeech" in mode + ): # Skips if the name of dataset (in mode string) contains "nonspeech" + + # Resample to 16000 for pesq + y_16k = pesq_resampler(y) + y_g_hat_16k = pesq_resampler(y_g_hat.squeeze(1)) + y_int_16k = (y_16k[0] * MAX_WAV_VALUE).short().cpu().numpy() + y_g_hat_int_16k = ( + (y_g_hat_16k[0] * MAX_WAV_VALUE).short().cpu().numpy() + ) + val_pesq_tot += pesq(16000, y_int_16k, y_g_hat_int_16k, "wb") + + # MRSTFT calculation + min_t = min(y.size(-1), y_g_hat.size(-1)) + val_mrstft_tot += loss_mrstft(y_g_hat[...,:min_t], y[...,:min_t]).item() + + # Log audio and figures to Tensorboard + if j % a.eval_subsample == 0: # Subsample every nth from validation set + if steps >= 0: + sw.add_audio(f"gt_{mode}/y_{j}", y[0], steps, h.sampling_rate) + if ( + a.save_audio + ): # Also save audio to disk if --save_audio is set to True + save_audio( + y[0], + os.path.join( + a.checkpoint_path, + "samples", + f"gt_{mode}", + f"{j:04d}.wav", + ), + h.sampling_rate, + ) + sw.add_figure( + f"gt_{mode}/y_spec_{j}", + plot_spectrogram(x[0]), + steps, + ) + + sw.add_audio( + f"generated_{mode}/y_hat_{j}", + y_g_hat[0], + steps, + h.sampling_rate, + ) + if ( + a.save_audio + ): # Also save audio to disk if --save_audio is set to True + save_audio( + y_g_hat[0, 0], + os.path.join( + a.checkpoint_path, + "samples", + f"{mode}_{steps:08d}", + f"{j:04d}.wav", + ), + h.sampling_rate, + ) + # Spectrogram of synthesized audio + y_hat_spec = mel_spectrogram( + y_g_hat.squeeze(1), + h.n_fft, + h.num_mels, + h.sampling_rate, + h.hop_size, + h.win_size, + h.fmin, + h.fmax, + ) + sw.add_figure( + f"generated_{mode}/y_hat_spec_{j}", + plot_spectrogram(y_hat_spec.squeeze(0).cpu().numpy()), + steps, + ) + + """ + Visualization of spectrogram difference between GT and synthesized audio, difference higher than 1 is clipped for better visualization. + """ + spec_delta = torch.clamp( + torch.abs(x[0] - y_hat_spec.squeeze(0).cpu()), + min=1e-6, + max=1.0, + ) + sw.add_figure( + f"delta_dclip1_{mode}/spec_{j}", + plot_spectrogram_clipped(spec_delta.numpy(), clip_max=1.0), + steps, + ) + + val_err = val_err_tot / (j + 1) + val_pesq = val_pesq_tot / (j + 1) + val_mrstft = val_mrstft_tot / (j + 1) + # Log evaluation metrics to Tensorboard + sw.add_scalar(f"validation_{mode}/mel_spec_error", val_err, steps) + sw.add_scalar(f"validation_{mode}/pesq", val_pesq, steps) + sw.add_scalar(f"validation_{mode}/mrstft", val_mrstft, steps) + + generator.train() + + # If the checkpoint is loaded, start with validation loop + if steps != 0 and rank == 0 and not a.debug: + if not a.skip_seen: + validate( + rank, + a, + h, + validation_loader, + mode=f"seen_{train_loader.dataset.name}", + ) + for i in range(len(list_unseen_validation_loader)): + validate( + rank, + a, + h, + list_unseen_validation_loader[i], + mode=f"unseen_{list_unseen_validation_loader[i].dataset.name}", + ) + # Exit the script if --evaluate is set to True + if a.evaluate: + exit() + + # Main training loop + generator.train() + mpd.train() + mrd.train() + for epoch in range(max(0, last_epoch), a.training_epochs): + if rank == 0: + start = time.time() + print(f"Epoch: {epoch + 1}") + + if h.num_gpus > 1: + train_sampler.set_epoch(epoch) + + for i, batch in enumerate(train_loader): + if rank == 0: + start_b = time.time() + x, y, _, y_mel = batch + + x = x.to(device, non_blocking=True) + y = y.to(device, non_blocking=True) + y_mel = y_mel.to(device, non_blocking=True) + y = y.unsqueeze(1) + + y_g_hat = generator(x) + y_g_hat_mel = mel_spectrogram( + y_g_hat.squeeze(1), + h.n_fft, + h.num_mels, + h.sampling_rate, + h.hop_size, + h.win_size, + h.fmin, + h.fmax_for_loss, + ) + + optim_d.zero_grad() + + # MPD + y_df_hat_r, y_df_hat_g, _, _ = mpd(y, y_g_hat.detach()) + loss_disc_f, losses_disc_f_r, losses_disc_f_g = discriminator_loss( + y_df_hat_r, y_df_hat_g + ) + + # MRD + y_ds_hat_r, y_ds_hat_g, _, _ = mrd(y, y_g_hat.detach()) + loss_disc_s, losses_disc_s_r, losses_disc_s_g = discriminator_loss( + y_ds_hat_r, y_ds_hat_g + ) + + loss_disc_all = loss_disc_s + loss_disc_f + + # Set clip_grad_norm value + clip_grad_norm = h.get("clip_grad_norm", 1000.0) # Default to 1000 + + # Whether to freeze D for initial training steps + if steps >= a.freeze_step: + loss_disc_all.backward() + grad_norm_mpd = torch.nn.utils.clip_grad_norm_( + mpd.parameters(), clip_grad_norm + ) + grad_norm_mrd = torch.nn.utils.clip_grad_norm_( + mrd.parameters(), clip_grad_norm + ) + optim_d.step() + else: + print( + f"[WARNING] skipping D training for the first {a.freeze_step} steps" + ) + grad_norm_mpd = 0.0 + grad_norm_mrd = 0.0 + + # Generator + optim_g.zero_grad() + + # L1 Mel-Spectrogram Loss + lambda_melloss = h.get( + "lambda_melloss", 45.0 + ) # Defaults to 45 in BigVGAN-v1 if not set + if h.get("use_multiscale_melloss", False): # uses wav for loss + loss_mel = fn_mel_loss_multiscale(y, y_g_hat) * lambda_melloss + else: # Uses mel for loss + loss_mel = fn_mel_loss_singlescale(y_mel, y_g_hat_mel) * lambda_melloss + + # MPD loss + y_df_hat_r, y_df_hat_g, fmap_f_r, fmap_f_g = mpd(y, y_g_hat) + loss_fm_f = feature_loss(fmap_f_r, fmap_f_g) + loss_gen_f, losses_gen_f = generator_loss(y_df_hat_g) + + # MRD loss + y_ds_hat_r, y_ds_hat_g, fmap_s_r, fmap_s_g = mrd(y, y_g_hat) + loss_fm_s = feature_loss(fmap_s_r, fmap_s_g) + loss_gen_s, losses_gen_s = generator_loss(y_ds_hat_g) + + if steps >= a.freeze_step: + loss_gen_all = ( + loss_gen_s + loss_gen_f + loss_fm_s + loss_fm_f + loss_mel + ) + else: + print( + f"[WARNING] using regression loss only for G for the first {a.freeze_step} steps" + ) + loss_gen_all = loss_mel + + loss_gen_all.backward() + grad_norm_g = torch.nn.utils.clip_grad_norm_( + generator.parameters(), clip_grad_norm + ) + optim_g.step() + + if rank == 0: + # STDOUT logging + if steps % a.stdout_interval == 0: + mel_error = ( + loss_mel.item() / lambda_melloss + ) # Log training mel regression loss to stdout + print( + f"Steps: {steps:d}, " + f"Gen Loss Total: {loss_gen_all:4.3f}, " + f"Mel Error: {mel_error:4.3f}, " + f"s/b: {time.time() - start_b:4.3f} " + f"lr: {optim_g.param_groups[0]['lr']:4.7f} " + f"grad_norm_g: {grad_norm_g:4.3f}" + ) + + # Checkpointing + if steps % a.checkpoint_interval == 0 and steps != 0: + checkpoint_path = f"{a.checkpoint_path}/g_{steps:08d}" + save_checkpoint( + checkpoint_path, + { + "generator": ( + generator.module if h.num_gpus > 1 else generator + ).state_dict() + }, + ) + checkpoint_path = f"{a.checkpoint_path}/do_{steps:08d}" + save_checkpoint( + checkpoint_path, + { + "mpd": (mpd.module if h.num_gpus > 1 else mpd).state_dict(), + "mrd": (mrd.module if h.num_gpus > 1 else mrd).state_dict(), + "optim_g": optim_g.state_dict(), + "optim_d": optim_d.state_dict(), + "steps": steps, + "epoch": epoch, + }, + ) + + # Tensorboard summary logging + if steps % a.summary_interval == 0: + mel_error = ( + loss_mel.item() / lambda_melloss + ) # Log training mel regression loss to tensorboard + sw.add_scalar("training/gen_loss_total", loss_gen_all.item(), steps) + sw.add_scalar("training/mel_spec_error", mel_error, steps) + sw.add_scalar("training/fm_loss_mpd", loss_fm_f.item(), steps) + sw.add_scalar("training/gen_loss_mpd", loss_gen_f.item(), steps) + sw.add_scalar("training/disc_loss_mpd", loss_disc_f.item(), steps) + sw.add_scalar("training/grad_norm_mpd", grad_norm_mpd, steps) + sw.add_scalar("training/fm_loss_mrd", loss_fm_s.item(), steps) + sw.add_scalar("training/gen_loss_mrd", loss_gen_s.item(), steps) + sw.add_scalar("training/disc_loss_mrd", loss_disc_s.item(), steps) + sw.add_scalar("training/grad_norm_mrd", grad_norm_mrd, steps) + sw.add_scalar("training/grad_norm_g", grad_norm_g, steps) + sw.add_scalar( + "training/learning_rate_d", scheduler_d.get_last_lr()[0], steps + ) + sw.add_scalar( + "training/learning_rate_g", scheduler_g.get_last_lr()[0], steps + ) + sw.add_scalar("training/epoch", epoch + 1, steps) + + # Validation + if steps % a.validation_interval == 0: + # Plot training input x so far used + for i_x in range(x.shape[0]): + sw.add_figure( + f"training_input/x_{i_x}", + plot_spectrogram(x[i_x].cpu()), + steps, + ) + sw.add_audio( + f"training_input/y_{i_x}", + y[i_x][0], + steps, + h.sampling_rate, + ) + + # Seen and unseen speakers validation loops + if not a.debug and steps != 0: + validate( + rank, + a, + h, + validation_loader, + mode=f"seen_{train_loader.dataset.name}", + ) + for i in range(len(list_unseen_validation_loader)): + validate( + rank, + a, + h, + list_unseen_validation_loader[i], + mode=f"unseen_{list_unseen_validation_loader[i].dataset.name}", + ) + steps += 1 + + # BigVGAN-v2 learning rate scheduler is changed from epoch-level to step-level + scheduler_g.step() + scheduler_d.step() + + if rank == 0: + print( + f"Time taken for epoch {epoch + 1} is {int(time.time() - start)} sec\n" + ) + + +def main(): + print("Initializing Training Process..") + + parser = argparse.ArgumentParser() + + parser.add_argument("--group_name", default=None) + + parser.add_argument("--input_wavs_dir", default="LibriTTS") + parser.add_argument("--input_mels_dir", default="ft_dataset") + parser.add_argument( + "--input_training_file", default="tests/LibriTTS/train-full.txt" + ) + parser.add_argument( + "--input_validation_file", default="tests/LibriTTS/val-full.txt" + ) + + parser.add_argument( + "--list_input_unseen_wavs_dir", + nargs="+", + default=["tests/LibriTTS", "tests/LibriTTS"], + ) + parser.add_argument( + "--list_input_unseen_validation_file", + nargs="+", + default=["tests/LibriTTS/dev-clean.txt", "tests/LibriTTS/dev-other.txt"], + ) + + parser.add_argument("--checkpoint_path", default="exp/bigvgan") + parser.add_argument("--config", default="") + + parser.add_argument("--training_epochs", default=100000, type=int) + parser.add_argument("--stdout_interval", default=5, type=int) + parser.add_argument("--checkpoint_interval", default=50000, type=int) + parser.add_argument("--summary_interval", default=100, type=int) + parser.add_argument("--validation_interval", default=50000, type=int) + + parser.add_argument( + "--freeze_step", + default=0, + type=int, + help="freeze D for the first specified steps. G only uses regression loss for these steps.", + ) + + parser.add_argument("--fine_tuning", default=False, type=bool) + + parser.add_argument( + "--debug", + default=False, + type=bool, + help="debug mode. skips validation loop throughout training", + ) + parser.add_argument( + "--evaluate", + default=False, + type=bool, + help="only run evaluation from checkpoint and exit", + ) + parser.add_argument( + "--eval_subsample", + default=5, + type=int, + help="subsampling during evaluation loop", + ) + parser.add_argument( + "--skip_seen", + default=False, + type=bool, + help="skip seen dataset. useful for test set inference", + ) + parser.add_argument( + "--save_audio", + default=False, + type=bool, + help="save audio of test set inference to disk", + ) + + a = parser.parse_args() + + with open(a.config) as f: + data = f.read() + + json_config = json.loads(data) + h = AttrDict(json_config) + + build_env(a.config, "config.json", a.checkpoint_path) + + torch.manual_seed(h.seed) + if torch.cuda.is_available(): + torch.cuda.manual_seed(h.seed) + h.num_gpus = torch.cuda.device_count() + h.batch_size = int(h.batch_size / h.num_gpus) + print(f"Batch size per GPU: {h.batch_size}") + else: + pass + + if h.num_gpus > 1: + mp.spawn( + train, + nprocs=h.num_gpus, + args=( + a, + h, + ), + ) + else: + train(0, a, h) + + +if __name__ == "__main__": + main() diff --git a/BigVGAN/utils.py b/BigVGAN/utils.py new file mode 100644 index 0000000000000000000000000000000000000000..da98a24cf1447778305563f8e909f30b06e06b26 --- /dev/null +++ b/BigVGAN/utils.py @@ -0,0 +1,99 @@ +# Adapted from https://github.com/jik876/hifi-gan under the MIT license. +# LICENSE is in incl_licenses directory. + +import glob +import os +import matplotlib +import torch +from torch.nn.utils import weight_norm + +matplotlib.use("Agg") +import matplotlib.pylab as plt +from .meldataset import MAX_WAV_VALUE +from scipy.io.wavfile import write + + +def plot_spectrogram(spectrogram): + fig, ax = plt.subplots(figsize=(10, 2)) + im = ax.imshow(spectrogram, aspect="auto", origin="lower", interpolation="none") + plt.colorbar(im, ax=ax) + + fig.canvas.draw() + plt.close() + + return fig + + +def plot_spectrogram_clipped(spectrogram, clip_max=2.0): + fig, ax = plt.subplots(figsize=(10, 2)) + im = ax.imshow( + spectrogram, + aspect="auto", + origin="lower", + interpolation="none", + vmin=1e-6, + vmax=clip_max, + ) + plt.colorbar(im, ax=ax) + + fig.canvas.draw() + plt.close() + + return fig + + +def init_weights(m, mean=0.0, std=0.01): + classname = m.__class__.__name__ + if classname.find("Conv") != -1: + m.weight.data.normal_(mean, std) + + +def apply_weight_norm(m): + classname = m.__class__.__name__ + if classname.find("Conv") != -1: + weight_norm(m) + + +def get_padding(kernel_size, dilation=1): + return int((kernel_size * dilation - dilation) / 2) + + +def load_checkpoint(filepath, device): + assert os.path.isfile(filepath) + print(f"Loading '{filepath}'") + checkpoint_dict = torch.load(filepath, map_location=device) + print("Complete.") + return checkpoint_dict + + +def save_checkpoint(filepath, obj): + print(f"Saving checkpoint to {filepath}") + torch.save(obj, filepath) + print("Complete.") + + +def scan_checkpoint(cp_dir, prefix, renamed_file=None): + # Fallback to original scanning logic first + pattern = os.path.join(cp_dir, prefix + "????????") + cp_list = glob.glob(pattern) + + if len(cp_list) > 0: + last_checkpoint_path = sorted(cp_list)[-1] + print(f"[INFO] Resuming from checkpoint: '{last_checkpoint_path}'") + return last_checkpoint_path + + # If no pattern-based checkpoints are found, check for renamed file + if renamed_file: + renamed_path = os.path.join(cp_dir, renamed_file) + if os.path.isfile(renamed_path): + print(f"[INFO] Resuming from renamed checkpoint: '{renamed_file}'") + return renamed_path + + return None + + +def save_audio(audio, path, sr): + # wav: torch with 1d shape + audio = audio * MAX_WAV_VALUE + audio = audio.cpu().numpy().astype("int16") + write(path, sr, audio) diff --git a/LICENSE b/LICENSE new file mode 100644 index 0000000000000000000000000000000000000000..cb908c072d8d2caef6cc27b8aecfc69349b4a3b2 --- /dev/null +++ b/LICENSE @@ -0,0 +1,21 @@ +MIT License + +Copyright (c) 2024 Stan Kirdey + +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/README.md b/README.md new file mode 100644 index 0000000000000000000000000000000000000000..b886f8f92385cc4ffb07d03d089b3feea5b2e3bc --- /dev/null +++ b/README.md @@ -0,0 +1,120 @@ +# VoiceRestore: Flow-Matching Transformers for Speech Recording Quality Restoration + +VoiceRestore is a cutting-edge speech restoration model designed to significantly enhance the quality of degraded voice recordings. Leveraging flow-matching transformers, this model excels at addressing a wide range of audio imperfections commonly found in speech, including background noise, reverberation, distortion, and signal loss. + +Demo of audio restorations: [VoiceRestore](https://sparkling-rabanadas-3082be.netlify.app/) + +Credits: This repository is based on the [E2-TTS implementation by Lucidrains](https://github.com/lucidrains/e2-tts-pytorch) + +## Example +### Degraded Input: + +![Degraded Input](./imgs/degraded.png "Degraded Input") + +Degraded audio (reverberation, distortion, noise, random cut): + +**Note**: Adjust your volume before playing the degraded audio sample, as it may contain distortions. + +https://github.com/user-attachments/assets/0c030274-60b5-41a4-abe6-59a3f1bc934b + +--- +### Restored (steps=32, cfg=1.0): + +![Restored](./imgs/restored.png "Restored") + +Restored audio - 16 steps, strength 0.5: + +https://github.com/user-attachments/assets/fdbbb988-9bd2-4750-bddd-32bd5153d254 + +--- +### Ground Truth: + +![Ground Truth](./imgs/ground_truth.png "Ground Truth") + +--- +## Key Features + +- **Universal Restoration**: The model can handle any level and type of voice recording degradation. Pure magic. +- **Easy to Use**: Simple interface for processing degraded audio files. +- **Pretrained Model**: Includes a 301 million parameter transformer model with pre-trained weights. (Model is still in the process of training, there will be further checkpoint updates) + +--- +## Quick Start + +1. Clone the repository: + ```bash + git clone --recurse-submodules https://github.com/skirdey/voicerestore.git + cd VoiceRestore + ``` + + if you did not clone with `--recurse-submodules`, you can run: + ```bash + git submodule update --init --recursive + ``` + +2. Install dependencies: + ```bash + pip install -r requirements.txt + ``` + +3. Download the [pre-trained model](https://drive.google.com/drive/folders/1uBJNp4mrPJQY9WEaiTI9u09IsRg1lAPR?usp=sharing) and place it in the `checkpoints` folder. + +4. Run a test restoration: + ```bash + python inference_short.py --checkpoint ./checkpoints/voice-restore-20d-16h-optim.pt --input test_input.wav --output test_output.wav --steps 32 --cfg_strength 0.5 + ``` + This will process `test_input.wav` and save the result as `test_output.wav`. + +5. Run a long form restoration, it uses window chunking: + ```bash + python inference_long.py --checkpoint ./checkpoints/voice-restore-20d-16h-optim.pt --input test_input_long.wav --output test_output_long.wav --steps 32 --cfg_strength 0.5 --window_size_sec 10.0 --overlap 0.25 + ``` + This will process `test_input_long.wav` (you need to provide it) and save the result as `test_output_long.wav`. + +## Usage + +To restore your own audio files: + +```python +from model import OptimizedAudioRestorationModel + +model = OptimizedAudioRestorationModel() +restored_audio = model.forward(input_audio, steps=32, cfg_strength=0.5) +``` + + + +## Model Details + +- **Architecture**: Flow-matching transformer +- **Parameters**: 300M+ parameters +- **Input**: Degraded speech audio (various formats supported) +- **Output**: Restored speech + +## Limitations and Future Work + +- Current model is optimized for speech; may not perform optimally on music or other audio types. +- Ongoing research to improve performance on extreme degradations. +- Future updates may include real-time processing capabilities. + +## Citation + +If you use VoiceRestore in your research, please cite our paper: + +``` +@article{kirdey2024voicerestore, + title={VoiceRestore: Flow-Matching Transformers for Speech Recording Quality Restoration}, + author={Kirdey, Stanislav}, + journal={arXiv}, + year={2024} +} +``` + +## License + +This project is licensed under the MIT License - see the [LICENSE](LICENSE) file for details. + +## Acknowledgments + +- Based on the [E2-TTS implementation by Lucidrains](https://github.com/lucidrains/e2-tts-pytorch) +- Special thanks to the open-source community for their invaluable contributions. diff --git a/audio/degraded.mp4 b/audio/degraded.mp4 new file mode 100644 index 0000000000000000000000000000000000000000..e75c31650df404149d5179222326ff6c3f36aeab Binary files /dev/null and b/audio/degraded.mp4 differ diff --git a/audio/restored.mp4 b/audio/restored.mp4 new file mode 100644 index 0000000000000000000000000000000000000000..e421360683f127058e2a1d987bd8737fd0618d26 Binary files /dev/null and b/audio/restored.mp4 differ diff --git a/config.json b/config.json new file mode 100644 index 0000000000000000000000000000000000000000..4a1a83714d81f8a87a47df244eae9a54732a9cd7 --- /dev/null +++ b/config.json @@ -0,0 +1,18 @@ +{ + "architectures": [ + "VoiceRestore" + ], + "model_type": "voice_restore", + "model_name_or_path": "content/voicerestore", + "steps": 16, + "cfg_strength": 0.5, + "window_size_sec": 5.0, + "overlap": 0.5, + "torch_dtype": "float32", + "transformers_version": "4.42.4", + "auto_map": { + "AutoConfig": "modeling.VoiceRestoreConfig", + "AutoModel": "modeling.VoiceRestore" + }, + "preprocessor_class": null +} diff --git a/imgs/degraded.png b/imgs/degraded.png new file mode 100644 index 0000000000000000000000000000000000000000..27f49780f771ceea332a80c7342dc6dfc02648b8 Binary files /dev/null and b/imgs/degraded.png differ diff --git a/imgs/ground_truth.png b/imgs/ground_truth.png new file mode 100644 index 0000000000000000000000000000000000000000..5c73230016157b64332f0c1bef62a88ed4bb0090 Binary files /dev/null and b/imgs/ground_truth.png differ diff --git a/imgs/restored.png b/imgs/restored.png new file mode 100644 index 0000000000000000000000000000000000000000..ec090c833ec91c7defc3f1e1a3bfdbfc088958fb Binary files /dev/null and b/imgs/restored.png differ diff --git a/inference_long.py b/inference_long.py new file mode 100644 index 0000000000000000000000000000000000000000..d83654f57f762f9f8eee3c0a964f0756caea395c --- /dev/null +++ b/inference_long.py @@ -0,0 +1,187 @@ +import sys +sys.path.append('./BigVGAN') + +import time +import torch +import torchaudio +import argparse +from tqdm import tqdm +import librosa +from BigVGAN import bigvgan +from BigVGAN.meldataset import get_mel_spectrogram +from model import OptimizedAudioRestorationModel + + +# Set the device handle macbooks with M1 chip + +device = 'cuda' if torch.cuda.is_available() else 'cpu' + +# Initialize BigVGAN model +bigvgan_model = bigvgan.BigVGAN.from_pretrained( + 'nvidia/bigvgan_v2_24khz_100band_256x', + use_cuda_kernel=False, + force_download=False +).to(device) +bigvgan_model.remove_weight_norm() + +def measure_gpu_memory(): + if device == 'cuda': + torch.cuda.synchronize() + return torch.cuda.max_memory_allocated() / (1024 ** 2) # Convert to MB + return 0 + + + +def apply_overlap_windowing_waveform(waveform, window_size_samples, overlap): + step_size = int(window_size_samples * (1 - overlap)) + num_chunks = (waveform.shape[-1] - window_size_samples) // step_size + 1 + windows = [] + + for i in range(num_chunks): + start_idx = i * step_size + end_idx = start_idx + window_size_samples + chunk = waveform[..., start_idx:end_idx] + windows.append(chunk) + + return torch.stack(windows) + +def reconstruct_waveform_from_windows(windows, window_size_samples, overlap): + step_size = int(window_size_samples * (1 - overlap)) + shape = windows.shape + if len(shape) == 2: + # windows.shape == (num_windows, window_len) + num_windows, window_len = shape + channels = 1 + windows = windows.unsqueeze(1) # Now windows.shape == (num_windows, 1, window_len) + elif len(shape) == 3: + num_windows, channels, window_len = shape + else: + raise ValueError(f"Unexpected windows.shape: {windows.shape}") + + output_length = (num_windows - 1) * step_size + window_size_samples + + reconstructed = torch.zeros((channels, output_length)) + window_sums = torch.zeros((channels, output_length)) + + for i in range(num_windows): + start_idx = i * step_size + end_idx = start_idx + window_len + reconstructed[:, start_idx:end_idx] += windows[i] + window_sums[:, start_idx:end_idx] += 1 + + reconstructed = reconstructed / window_sums.clamp(min=1e-6) + if channels == 1: + reconstructed = reconstructed.squeeze(0) # Remove channel dimension if single channel + return reconstructed + +def load_model(save_path): + """ + Load the optimized audio restoration model. + + Parameters: + - save_path: Path to the checkpoint file. + """ + optimized_model = OptimizedAudioRestorationModel(device=device, bigvgan_model=bigvgan_model) + state_dict = torch.load(save_path, map_location=device) + + if 'model_state_dict' in state_dict: + state_dict = state_dict['model_state_dict'] + optimized_model.voice_restore.load_state_dict(state_dict, strict=True) + + return optimized_model + +def restore_audio(model, input_path, output_path, steps=16, cfg_strength=0.5, window_size_sec=5.0, overlap=0.5): + # Load the audio file + start_time = time.time() + initial_gpu_memory = measure_gpu_memory() + wav, sr = librosa.load(input_path, sr=24000, mono=True) + wav = torch.FloatTensor(wav).unsqueeze(0) # Shape: [1, num_samples] + + window_size_samples = int(window_size_sec * sr) + step_size = int(window_size_samples * (1 - overlap)) + + # Apply overlapping windowing to the waveform + wav_windows = apply_overlap_windowing_waveform(wav, window_size_samples, overlap) + + restored_wav_windows = [] + + for wav_window in tqdm(wav_windows): + wav_window = wav_window.to(device) # Shape: [1, window_size_samples] + + # Convert to Mel-spectrogram + processed_mel = get_mel_spectrogram(wav_window, bigvgan_model.h).to(device) + + # Restore audio + with torch.no_grad(): + with torch.autocast(device): + restored_mel = model.voice_restore.sample(processed_mel.transpose(1, 2), steps=steps, cfg_strength=cfg_strength) + restored_mel = restored_mel.squeeze(0).transpose(0, 1) + + # Convert restored mel-spectrogram to waveform + with torch.no_grad(): + with torch.autocast(device): + restored_wav = bigvgan_model(restored_mel.unsqueeze(0)).squeeze(0).float().cpu() # Shape: [num_samples] + + # Debug: Print shapes + # print(f"restored_wav.shape: {restored_wav.shape}") + + restored_wav_windows.append(restored_wav) + del wav_window, processed_mel, restored_mel, restored_wav + torch.cuda.empty_cache() + + restored_wav_windows = torch.stack(restored_wav_windows) # Shape: [num_windows, num_samples] + + # Debug: Print shapes + # print(f"restored_wav_windows.shape: {restored_wav_windows.shape}") + + # Reconstruct the full waveform from the processed windows + restored_wav = reconstruct_waveform_from_windows(restored_wav_windows, window_size_samples, overlap) + + # Ensure the restored_wav has correct dimensions for saving + if restored_wav.dim() == 1: + restored_wav = restored_wav.unsqueeze(0) # Shape: [1, num_samples] + + # Save the restored audio + torchaudio.save(output_path, restored_wav, 24000) + + end_time = time.time() + total_time = end_time - start_time + peak_gpu_memory = measure_gpu_memory() + gpu_memory_used = peak_gpu_memory - initial_gpu_memory + + print(f"Total inference time: {total_time:.2f} seconds") + print(f"Peak GPU memory usage: {peak_gpu_memory:.2f} MB") + print(f"GPU memory used: {gpu_memory_used:.2f} MB") + +if __name__ == "__main__": + # Argument parser setup + parser = argparse.ArgumentParser(description="Audio restoration using OptimizedAudioRestorationModel for long-form audio.") + parser.add_argument('--checkpoint', type=str, required=True, help="Path to the checkpoint file") + parser.add_argument('--input', type=str, required=True, help="Path to the input audio file") + parser.add_argument('--output', type=str, required=True, help="Path to save the restored audio file") + parser.add_argument('--steps', type=int, default=16, help="Number of sampling steps") + parser.add_argument('--cfg_strength', type=float, default=0.5, help="CFG strength value") + parser.add_argument('--window_size_sec', type=float, default=5.0, help="Window size in seconds for overlapping") + parser.add_argument('--overlap', type=float, default=0.5, help="Overlap ratio for windowing") + + # Parse arguments + args = parser.parse_args() + + # Load the optimized model + optimized_model = load_model(args.checkpoint) + + if device == 'cuda': + optimized_model.bfloat16() + optimized_model.eval() + optimized_model.to(device) + + # Use the model to restore audio + restore_audio( + optimized_model, + args.input, + args.output, + steps=args.steps, + cfg_strength=args.cfg_strength, + window_size_sec=args.window_size_sec, + overlap=args.overlap + ) diff --git a/inference_short.py b/inference_short.py new file mode 100644 index 0000000000000000000000000000000000000000..d0af58da2a2c950412944fa4e9dfc91855918b2a --- /dev/null +++ b/inference_short.py @@ -0,0 +1,73 @@ +import sys +sys.path.append('./BigVGAN') + +import torch +import torch.nn as nn +import torchaudio +import argparse +from BigVGAN import bigvgan +from BigVGAN.meldataset import get_mel_spectrogram +from model import OptimizedAudioRestorationModel + + +device = 'cuda' if torch.cuda.is_available() else 'cpu' + +# If running on non-windows system, you can try using cuda kernel for faster processing `use_cuda_kernel=True` +bigvgan_model = bigvgan.BigVGAN.from_pretrained('nvidia/bigvgan_v2_24khz_100band_256x', use_cuda_kernel=False).to(device) +bigvgan_model.remove_weight_norm() +example_input = torch.randn(1, 16000) # Example input waveform +example_spec = get_mel_spectrogram(example_input, bigvgan_model.h) + + +def load_model(save_path): + """ + Load the model. + + Parameters: + - save_path: The file path where the optimized model is saved. + """ + + optimized_model = OptimizedAudioRestorationModel(device=device, bigvgan_model=bigvgan_model) + state_dict = torch.load(save_path, map_location=torch.device(device)) + if 'model_state_dict' in state_dict: + state_dict = state_dict['model_state_dict'] + optimized_model.voice_restore.load_state_dict(state_dict, strict=True) + + return optimized_model + + +def restore_audio(model, input_path, output_path, steps=16, cfg_strength=0.5): + audio, sr = torchaudio.load(input_path) + + if sr != model.target_sample_rate: + audio = torchaudio.functional.resample(audio, sr, model.target_sample_rate) + + audio = audio.mean(dim=0, keepdim=True) if audio.dim() > 1 else audio # Convert to mono if stereo + + with torch.inference_mode(): + with torch.autocast(device): + restored_wav = model(audio, steps=steps, cfg_strength=cfg_strength) + restored_wav = restored_wav.squeeze(0).float().cpu() # Move to CPU after processing + + torchaudio.save(output_path, restored_wav, model.target_sample_rate) + + +if __name__ == "__main__": + # Argument parser setup + parser = argparse.ArgumentParser(description="Audio restoration using OptimizedAudioRestorationModel") + parser.add_argument('--checkpoint', type=str, required=True, help="Path to the checkpoint file") + parser.add_argument('--input', type=str, required=True, help="Path to the input audio file") + parser.add_argument('--output', type=str, required=True, help="Path to save the restored audio file") + parser.add_argument('--steps', type=int, default=16, help="Number of sampling steps") + parser.add_argument('--cfg_strength', type=float, default=0.5, help="CFG strength value") + + # Parse arguments + args = parser.parse_args() + + # Load the optimized model + optimized_model = load_model(args.checkpoint) + optimized_model.eval() + optimized_model.to(device) + + # Use the model to restore audio + restore_audio(optimized_model, args.input, args.output, steps=args.steps, cfg_strength=args.cfg_strength) diff --git a/long_form_sample.ogg b/long_form_sample.ogg new file mode 100644 index 0000000000000000000000000000000000000000..ecce4c53ff333fc6a15e54f88864d0c6944565e6 --- /dev/null +++ b/long_form_sample.ogg @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:043ed59b76cf8f41cfaad3be67033f5bc29a530805431b0d97383f5b2f708fe8 +size 2948029 diff --git a/model.py b/model.py new file mode 100644 index 0000000000000000000000000000000000000000..43365d3c87f7c7f6b3a2110d24eb506b42663a8c --- /dev/null +++ b/model.py @@ -0,0 +1,52 @@ +import torch +from BigVGAN.meldataset import get_mel_spectrogram +from voice_restore import VoiceRestore + + +class OptimizedAudioRestorationModel(torch.nn.Module): + def __init__(self, target_sample_rate=24000, device=None, bigvgan_model=None): + super().__init__() + + # Initialize VoiceRestore + self.voice_restore = VoiceRestore( + sigma=0.0, + transformer={ + 'dim': 768, + 'depth': 20, + 'heads': 16, + 'dim_head': 64, + 'skip_connect_type': 'concat', + 'max_seq_len': 2000, + }, + num_channels=100 + ) + + self.device = device + if self.device == 'cuda': + self.voice_restore.bfloat16() + self.voice_restore.eval() + self.voice_restore.to(self.device) + self.target_sample_rate = target_sample_rate + self.bigvgan_model = bigvgan_model + + + + def forward(self, audio, steps=32, cfg_strength=0.5): + # Convert to Mel-spectrogram + + if self.bigvgan_model is None: + raise ValueError("BigVGAN model is not provided. Please provide the BigVGAN model.") + + if self.device is None: + raise ValueError("Device is not provided. Please provide the device (cuda, cpu or mps).") + + processed_mel = get_mel_spectrogram(audio, self.bigvgan_model.h).to(self.device) + + # Restore audio + restored_mel = self.voice_restore.sample(processed_mel.transpose(1, 2), steps=steps, cfg_strength=cfg_strength) + restored_mel = restored_mel.squeeze(0).transpose(0, 1) + + # Convert restored mel-spectrogram to waveform + restored_wav = self.bigvgan_model(restored_mel.unsqueeze(0)) + + return restored_wav \ No newline at end of file diff --git a/modeling.py b/modeling.py new file mode 100644 index 0000000000000000000000000000000000000000..0fcd16eea86ad71db8a6ad7702b40e7eeeadccff --- /dev/null +++ b/modeling.py @@ -0,0 +1,128 @@ +import torch +import torchaudio +import torch.nn as nn +from transformers import PreTrainedModel +import torch +from BigVGAN import bigvgan +from BigVGAN.meldataset import get_mel_spectrogram +from voice_restore import VoiceRestore +import argparse +from model import OptimizedAudioRestorationModel +import librosa +from inference_long import apply_overlap_windowing_waveform, reconstruct_waveform_from_windows + +device = torch.device("cuda" if torch.cuda.is_available() else "cpu") + +# Configuration class for VoiceRestore +class VoiceRestoreConfig(PretrainedConfig): + model_type = "voice_restore" + + def __init__(self, **kwargs): + super().__init__(**kwargs) + self.steps = kwargs.get("steps", 16) + self.cfg_strength = kwargs.get("cfg_strength", 0.5) + self.window_size_sec = kwargs.get("window_size_sec", 5.0) + self.overlap = kwargs.get("overlap", 0.5) + +# Model class for VoiceRestore +class VoiceRestore(PreTrainedModel): + config_class = VoiceRestoreConfig + + def __init__(self, config: VoiceRestoreConfig): + super().__init__(config) + self.steps = config.steps + self.cfg_strength = config.cfg_strength + self.window_size_sec = config.window_size_sec + self.overlap = config.overlap + + # Initialize BigVGAN model + self.bigvgan_model = bigvgan.BigVGAN.from_pretrained( + 'nvidia/bigvgan_v2_24khz_100band_256x', + use_cuda_kernel=False, + force_download=False + ).to(device) + self.bigvgan_model.remove_weight_norm() + + # Optimized restoration model + self.optimized_model = OptimizedAudioRestorationModel(device=device, bigvgan_model=self.bigvgan_model) + save_path = "/content/voicerestore/checkpoints/voice-restore-20d-16h-optim.pt" + state_dict = torch.load(save_path, map_location=torch.device(device)) + print("loaded") + if 'model_state_dict' in state_dict: + state_dict = state_dict['model_state_dict'] + print("change keys") + + self.optimized_model.voice_restore.load_state_dict(state_dict, strict=True) + self.optimized_model.eval() + + def forward(self, input_path, output_path, short=True): + # Restore the audio using the parameters from the config + if short: + self.restore_audio_short(self.optimized_model, input_path, output_path, self.steps, self.cfg_strength) + else: + self.restore_audio_long(self.optimized_model, input_path, output_path, self.steps, self.cfg_strength, self.window_size_sec, self.overlap) + + def restore_audio_short(self, model, input_path, output_path, steps, cfg_strength): + """ + Short inference for audio restoration. + """ + # Load the audio file + device_type = device.type + audio, sr = torchaudio.load(input_path) + if sr != model.target_sample_rate: + audio = torchaudio.functional.resample(audio, sr, model.target_sample_rate) + + audio = audio.mean(dim=0, keepdim=True) if audio.dim() > 1 else audio # Convert to mono if stereo + + with torch.inference_mode(): + with torch.autocast(device_type): + restored_wav = model(audio, steps=steps, cfg_strength=cfg_strength) + restored_wav = restored_wav.squeeze(0).float().cpu() # Move to CPU after processing + + # Save the restored audio + torchaudio.save(output_path, restored_wav, model.target_sample_rate) + + def restore_audio_long(self, model, input_path, output_path, steps, cfg_strength, window_size_sec, overlap): + """ + Long inference for audio restoration using overlapping windows. + """ + # Load the audio file + wav, sr = librosa.load(input_path, sr=24000, mono=True) + wav = torch.FloatTensor(wav).unsqueeze(0) # Shape: [1, num_samples] + + window_size_samples = int(window_size_sec * sr) + wav_windows = apply_overlap_windowing_waveform(wav, window_size_samples, overlap) + + restored_wav_windows = [] + for wav_window in wav_windows: + wav_window = wav_window.to(device) + processed_mel = get_mel_spectrogram(wav_window, self.bigvgan_model.h).to(device) + + # Restore audio + with torch.no_grad(): + with torch.autocast(device): + restored_mel = model.voice_restore.sample(processed_mel.transpose(1, 2), steps=steps, cfg_strength=cfg_strength) + restored_mel = restored_mel.squeeze(0).transpose(0, 1) + + restored_wav = self.bigvgan_model(restored_mel.unsqueeze(0)).squeeze(0).float().cpu() + restored_wav_windows.append(restored_wav) + + torch.cuda.empty_cache() + + restored_wav_windows = torch.stack(restored_wav_windows) + restored_wav = reconstruct_waveform_from_windows(restored_wav_windows, window_size_samples, overlap) + + # Save the restored audio + torchaudio.save(output_path, restored_wav.unsqueeze(0), 24000) + + +# # Function to load the model using AutoModel +# from transformers import AutoModel + +# def load_voice_restore_model(checkpoint_path: str): +# model = AutoModel.from_pretrained(checkpoint_path, config=VoiceRestoreConfig()) +# return model + +# # Example Usage +# model = load_voice_restore_model("./checkpoints/voice-restore-20d-16h-optim.pt") +# model("test_input.wav", "test_output.wav") diff --git a/pytorch_model.bin b/pytorch_model.bin new file mode 100644 index 0000000000000000000000000000000000000000..557b7fa7ab766dc65867abb4c48c8acb4dc37d8f --- /dev/null +++ b/pytorch_model.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:556a83f119c828d0a6ee47bc84149403fd4428eb755bcc222169a68eaa4443dd +size 1204445470 diff --git a/requirements.txt b/requirements.txt new file mode 100644 index 0000000000000000000000000000000000000000..da3eb78c6b356dc2896c143a495a24b7bb92e2cd --- /dev/null +++ b/requirements.txt @@ -0,0 +1,19 @@ +torch==2.4.1 +torchaudio==2.4.1 +torchvision==0.19.1 +librosa==0.10.2.post1 +torchdiffeq==0.2.4 +jaxtyping==0.2.34 +x-transformers==1.34.0 +gateloop-transformer==0.2.5 +matplotlib==3.9.2 +huggingface-hub==0.24.6 +matplotlib +tqdm==4.66.5 +scipy==1.14.1 +scikit-learn==1.5.1 +rotary-embedding-torch==0.8.3 +soundfile==0.12.1 +einops +ninja +tqdm \ No newline at end of file diff --git a/tensor_typing.py b/tensor_typing.py new file mode 100644 index 0000000000000000000000000000000000000000..598500733685166aafd97a46bb27ee1d98117315 --- /dev/null +++ b/tensor_typing.py @@ -0,0 +1,26 @@ +from torch import Tensor + +from jaxtyping import ( + Float, + Int, + Bool +) + +# jaxtyping is a misnomer, works for pytorch + +class TorchTyping: + def __init__(self, abstract_dtype): + self.abstract_dtype = abstract_dtype + + def __getitem__(self, shapes: str): + return self.abstract_dtype[Tensor, shapes] + +Float = TorchTyping(Float) +Int = TorchTyping(Int) +Bool = TorchTyping(Bool) + +__all__ = [ + Float, + Int, + Bool +] diff --git a/test_input.wav b/test_input.wav new file mode 100644 index 0000000000000000000000000000000000000000..8ef9628ad893f9416d6be0f50c3678a161136e54 Binary files /dev/null and b/test_input.wav differ diff --git a/test_output_1.wav b/test_output_1.wav new file mode 100644 index 0000000000000000000000000000000000000000..2bbc2397531aa624139767f704669b47db87c4b9 Binary files /dev/null and b/test_output_1.wav differ diff --git a/voice_restore.py b/voice_restore.py new file mode 100644 index 0000000000000000000000000000000000000000..6196ad9c22313c77fd084f42c3731fca5a960cff --- /dev/null +++ b/voice_restore.py @@ -0,0 +1,238 @@ +# +# Code is adapted from https://github.com/lucidrains/e2-tts-pytorch +# + +""" +ein notation: +b - batch +n - sequence +nt - text sequence +nw - raw wave length +d - dimension +""" + + +from __future__ import annotations +from typing import Dict, Any, Optional +from functools import partial + +import torch +from torch import nn +from torch.nn import Module, ModuleList, Sequential, Linear +import torch.nn.functional as F + +from torchdiffeq import odeint +from einops.layers.torch import Rearrange +from einops import rearrange, repeat, pack, unpack +from x_transformers import Attention, FeedForward, RMSNorm, AdaptiveRMSNorm +from x_transformers.x_transformers import RotaryEmbedding +from gateloop_transformer import SimpleGateLoopLayer + +from tensor_typing import Float + +class Identity(Module): + def forward(self, x, **kwargs): + return x + +class AdaLNZero(Module): + def __init__(self, dim: int, dim_condition: Optional[int] = None, init_bias_value: float = -2.): + super().__init__() + dim_condition = dim_condition or dim + self.to_gamma = nn.Linear(dim_condition, dim) + nn.init.zeros_(self.to_gamma.weight) + nn.init.constant_(self.to_gamma.bias, init_bias_value) + + def forward(self, x: torch.Tensor, *, condition: torch.Tensor) -> torch.Tensor: + if condition.ndim == 2: + condition = rearrange(condition, 'b d -> b 1 d') + gamma = self.to_gamma(condition).sigmoid() + return x * gamma + +def exists(v: Any) -> bool: + return v is not None + +def default(v: Any, d: Any) -> Any: + return v if exists(v) else d + +def divisible_by(num: int, den: int) -> bool: + return (num % den) == 0 + +class Transformer(Module): + def __init__( + self, + *, + dim: int, + depth: int = 8, + cond_on_time: bool = True, + skip_connect_type: str = 'concat', + abs_pos_emb: bool = True, + max_seq_len: int = 8192, + heads: int = 8, + dim_head: int = 64, + num_gateloop_layers: int = 1, + dropout: float = 0.1, + num_registers: int = 32, + attn_kwargs: Dict[str, Any] = dict(gate_value_heads=True, softclamp_logits=True), + ff_kwargs: Dict[str, Any] = dict() + ): + super().__init__() + assert divisible_by(depth, 2), 'depth needs to be even' + + self.max_seq_len = max_seq_len + self.abs_pos_emb = nn.Embedding(max_seq_len, dim) if abs_pos_emb else None + self.dim = dim + self.skip_connect_type = skip_connect_type + needs_skip_proj = skip_connect_type == 'concat' + self.depth = depth + self.layers = ModuleList([]) + + self.num_registers = num_registers + self.registers = nn.Parameter(torch.zeros(num_registers, dim)) + nn.init.normal_(self.registers, std=0.02) + + self.rotary_emb = RotaryEmbedding(dim_head) + self.cond_on_time = cond_on_time + rmsnorm_klass = AdaptiveRMSNorm if cond_on_time else RMSNorm + postbranch_klass = partial(AdaLNZero, dim=dim) if cond_on_time else Identity + + self.time_cond_mlp = Sequential( + Rearrange('... -> ... 1'), + Linear(1, dim), + nn.SiLU() + ) if cond_on_time else nn.Identity() + + for ind in range(depth): + is_later_half = ind >= (depth // 2) + gateloop = SimpleGateLoopLayer(dim=dim) + attn_norm = rmsnorm_klass(dim) + attn = Attention(dim=dim, heads=heads, dim_head=dim_head, dropout=dropout, **attn_kwargs) + attn_adaln_zero = postbranch_klass() + ff_norm = rmsnorm_klass(dim) + ff = FeedForward(dim=dim, glu=True, dropout=dropout, **ff_kwargs) + ff_adaln_zero = postbranch_klass() + skip_proj = Linear(dim * 2, dim, bias=False) if needs_skip_proj and is_later_half else None + + self.layers.append(ModuleList([ + gateloop, skip_proj, attn_norm, attn, attn_adaln_zero, + ff_norm, ff, ff_adaln_zero + ])) + + self.final_norm = RMSNorm(dim) + + def forward( + self, + x: Float['b n d'], + times: Optional[Float['b'] | Float['']] = None, + ) -> torch.Tensor: + batch, seq_len, device = *x.shape[:2], x.device + + assert not (exists(times) ^ self.cond_on_time), '`times` must be passed in if `cond_on_time` is set to `True` and vice versa' + + norm_kwargs = {} + + if exists(self.abs_pos_emb): + # assert seq_len <= self.max_seq_len, f'{seq_len} exceeds the set `max_seq_len` ({self.max_seq_len}) on Transformer' + seq = torch.arange(seq_len, device=device) + x = x + self.abs_pos_emb(seq) + + if exists(times): + if times.ndim == 0: + times = repeat(times, ' -> b', b=batch) + times = self.time_cond_mlp(times) + norm_kwargs['condition'] = times + + registers = repeat(self.registers, 'r d -> b r d', b=batch) + x, registers_packed_shape = pack((registers, x), 'b * d') + + rotary_pos_emb = self.rotary_emb.forward_from_seq_len(x.shape[-2]) + + skips = [] + + for ind, ( + gateloop, maybe_skip_proj, attn_norm, attn, maybe_attn_adaln_zero, + ff_norm, ff, maybe_ff_adaln_zero + ) in enumerate(self.layers): + layer = ind + 1 + is_first_half = layer <= (self.depth // 2) + + if is_first_half: + skips.append(x) + else: + skip = skips.pop() + if self.skip_connect_type == 'concat': + x = torch.cat((x, skip), dim=-1) + x = maybe_skip_proj(x) + + x = gateloop(x) + x + + attn_out = attn(attn_norm(x, **norm_kwargs), rotary_pos_emb=rotary_pos_emb) + x = x + maybe_attn_adaln_zero(attn_out, **norm_kwargs) + + ff_out = ff(ff_norm(x, **norm_kwargs)) + x = x + maybe_ff_adaln_zero(ff_out, **norm_kwargs) + + assert len(skips) == 0 + + _, x = unpack(x, registers_packed_shape, 'b * d') + + return self.final_norm(x) + +class VoiceRestore(nn.Module): + def __init__( + self, + sigma: float = 0.0, + transformer: Optional[Dict[str, Any]] = None, + odeint_kwargs: Optional[Dict[str, Any]] = None, + num_channels: int = 100, + ): + super().__init__() + self.sigma = sigma + self.num_channels = num_channels + + self.transformer = Transformer(**transformer, cond_on_time=True) + + self.odeint_kwargs = odeint_kwargs or {'atol': 1e-5, 'rtol': 1e-5, 'method': 'midpoint'} + + self.proj_in = nn.Linear(num_channels, self.transformer.dim) + self.cond_proj = nn.Linear(num_channels, self.transformer.dim) + self.to_pred = nn.Linear(self.transformer.dim, num_channels) + + def transformer_with_pred_head(self, x: torch.Tensor, times: torch.Tensor, cond: Optional[torch.Tensor] = None) -> torch.Tensor: + x = self.proj_in(x) + if cond is not None: + cond_proj = self.cond_proj(cond) + x = x + cond_proj + attended = self.transformer(x, times=times) + return self.to_pred(attended) + + def cfg_transformer_with_pred_head( + self, + *args, + cond=None, + mask=None, + cfg_strength: float = 0.5, + **kwargs, + ): + pred = self.transformer_with_pred_head(*args, **kwargs, cond=cond) + + if cfg_strength < 1e-5: + return pred * mask.unsqueeze(-1) if mask is not None else pred + + null_pred = self.transformer_with_pred_head(*args, **kwargs, cond=None) + + result = pred + (pred - null_pred) * cfg_strength + return result * mask.unsqueeze(-1) if mask is not None else result + + + @torch.no_grad() + def sample(self, processed: torch.Tensor, steps: int = 32, cfg_strength: float = 0.5) -> torch.Tensor: + self.eval() + times = torch.linspace(0, 1, steps, device=processed.device) + + def ode_fn(t: torch.Tensor, x: torch.Tensor) -> torch.Tensor: + return self.cfg_transformer_with_pred_head(x, times=t, cond=processed, cfg_strength=cfg_strength) + + y0 = torch.randn_like(processed) + trajectory = odeint(ode_fn, y0, times, **self.odeint_kwargs) + restored = trajectory[-1] + return restored \ No newline at end of file