revsic / tf-diffwave

Tensorflow implementation of DiffWave: A Versatile Diffusion Model for Audio Synthesis

Home Page:https://revsic.github.io/tf-diffwave

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

tf-diffwave

(Unofficial) Tensorflow implementation of DiffWave (Zhifeng Kong et al., 2020)

  • DiffWave: A Versatile Diffusion Model for Audio Synthesis, Zhifeng Kong et al., 2020. [arXiv:2009.09761]

Requirements

Tested in python 3.7.3 conda environment, requirements.txt

Usage

To download LJ-Speech dataset, run under script.

Dataset will be downloaded in '~/tensorflow_datasets' in tfrecord format. If you want to change the download directory, specify data_dir parameter of LJSpeech initializer.

from dataset import LJSpeech
from dataset.config import Config

config = Config()
# lj = LJSpeech(config, data_dir=path, download=True)
lj = LJSpeech(config, download=True) 

To train model, run train.py.

Checkpoint will be written on TrainConfig.ckpt, tensorboard summary on TrainConfig.log.

python train.py
tensorboard --logdir ./log/

If you want to train model from raw audio, specify audio directory and turn on the flag --from-raw.

python .\train.py --data-dir D:\LJSpeech-1.1\wavs --from-raw

To start to train from previous checkpoint, --load-step is available.

python .\train.py --load-step 416 --config ./ckpt/q1.json

For experiments, reference expr.ipynb.

To inference test set, run inference.py.

python .\inference.py

Pretrained checkpoints are relased on releases.

To use pretrained model, download files and unzip it. Checkout git repository to proper commit tags and followings are sample script.

with open('l1.json') as f:
    config = Config.load(json.load(f))

diffwave = DiffWave(config.model)
diffwave.restore('./l1/l1_1000000.ckpt-1').expect_partial()

Learning Curve

res.channels=64, T=20, train 1M steps.

loss

Samples

Reference https://revsic.github.io/tf-diffwave.

About

Tensorflow implementation of DiffWave: A Versatile Diffusion Model for Audio Synthesis

https://revsic.github.io/tf-diffwave

License:MIT License


Languages

Language:Jupyter Notebook 88.3%Language:Python 11.7%