https://github.com/mazzzystar/wavegan-pytorch

PyTorch implementation of " Synthesizing Audio with Generative Adversarial Networks"
https://github.com/mazzzystar/wavegan-pytorch

ai-music gan generative-model music-generation wavegan wavegan-pytorch

Last synced: 2 months ago
JSON representation

PyTorch implementation of " Synthesizing Audio with Generative Adversarial Networks"

• Host: GitHub
• URL: https://github.com/mazzzystar/wavegan-pytorch
• Owner: mazzzystar
• Created: 2018-07-16T12:03:08.000Z (over 6 years ago)
• Default Branch: master
• Last Pushed: 2020-06-08T12:37:57.000Z (over 4 years ago)
• Last Synced: 2024-08-05T06:05:32.811Z (6 months ago)
• Topics: ai-music, gan, generative-model, music-generation, wavegan, wavegan-pytorch
• Language: Python
• Homepage: https://arxiv.org/abs/1802.04208
• Size: 439 KB
• Stars: 64
• Watchers: 6
• Forks: 15
• Open Issues: 4
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

        
# WaveGAN-pytorch

PyTorch implementation of [Synthesizing Audio with Generative Adversarial Networks(Chris Donahue, Feb 2018)](https://arxiv.org/abs/1802.04208).

Befor running, make sure you have the `sc09` dataset, and put that dataset under your current filepath.

## Quick Start:

1. Installation

```

sudo apt-get install libav-tools

```

2. Download dataset

* `sc09`: [sc09 raw WAV files](http://deepyeti.ucsd.edu/cdonahue/sc09.tar.gz), utterances of spoken english words '0'-'9'

* `piano`: [Piano raw WAV files](http://deepyeti.ucsd.edu/cdonahue/mancini_piano.tar.gz)

3. Run

For `sc09` task, **make sure `sc09` dataset under your current project filepath befor run your code.**

```

$ python train.py

```

#### Training time

* For `SC09` dataset, 4 X Tesla P40 takes nearly 2 days to get reasonable result.

* For `piano` piano dataset, 2 X Tesla P40 takes 3-6 hours to get reasonable result.

* Increase the `BATCH_SIZE` from 10 to 32 or 64 can acquire shorter per-epoch time on multiple-GPU but slower gradient descent learning rate.

## Results

Generated "0-9": https://soundcloud.com/mazzzystar/sets/dcgan-sc09

Generated piano: https://soundcloud.com/mazzzystar/sets/wavegan-piano

Loss curve:

![](imgs/loss_curve.png)

## Architecture

![](imgs/archi.png)

## TODO

* [ ] Add some evaluation experiments, eg. inception score.

## Contributions

This repo is based on [chrisdonahue's](https://github.com/chrisdonahue/wavegan) and [jtcramer's](https://github.com/jtcramer/wavegan) implementation.