Ecosyste.ms: Awesome
An open API service indexing awesome lists of open source software.
https://github.com/andreaferretti/wgan
Wasserstein GAN with Gradient Penalty
https://github.com/andreaferretti/wgan
gan pytorch wasserstein-gan
Last synced: 10 days ago
JSON representation
Wasserstein GAN with Gradient Penalty
- Host: GitHub
- URL: https://github.com/andreaferretti/wgan
- Owner: andreaferretti
- License: apache-2.0
- Created: 2019-08-22T09:37:00.000Z (over 5 years ago)
- Default Branch: master
- Last Pushed: 2019-09-09T14:38:09.000Z (over 5 years ago)
- Last Synced: 2024-12-02T07:31:07.316Z (2 months ago)
- Topics: gan, pytorch, wasserstein-gan
- Language: Python
- Size: 15.6 KB
- Stars: 1
- Watchers: 3
- Forks: 2
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
README
# Wasserstein GAN with Gradient Penalty in Pytorch
This is a cleanup of [improved-wgan-pytorch](https://github.com/jalola/improved-wgan-pytorch),
which implements methods from [Improved Training of Wasserstein GANs](https://github.com/igul222/improved_wgan_training)
to train Wasserstein GAN with Gradient Penalty.
## Prerequisites
This project uses PyTorch (any recent version will do, we are using 1.2),
and [LMDB](https://lmdb.readthedocs.io/en/release/) (the latter is needed to
load data in [LSUN](https://github.com/fyu/lsun) format).
Optionally, it can log events using [tensorboardX](https://github.com/lanpa/tensorboardX)
that can be displayed using [TensorBoard](https://www.tensorflow.org/tensorboard).
For instance, you can create an [Anaconda](https://anaconda.org) environment
to run this project using something like
```
conda create -n wgan python=3.6
source activate wgan
conda install pytorch torchvision cudatoolkit=9.0 -c pytorch
pip install lmdb
pip install tensorboardX # optional
```
## Usage
If you have a folder of images, structured like this
```
.
└── images
├── class1
│ ├── img001.png
│ ├── img002.png
│ └── ...
├── class2
│ └── ...
└── ...
```
use `--dataset raw --data-dir images`, while for the LSUN dataset
```
.
└── LSUN
├── bedroom_train_lmdb
│ ├── data.mdb
│ └── lock.mdb
├── bedroom_val_lmdb
│ ├── data.mdb
│ └── lock.mdb
├── bridge_train_lmdb
│ ├── data.mdb
│ └── lock.mdb
├── bridge_val_lmdb
│ ├── data.mdb
│ └── lock.mdb
└── ...
```
use something like `--dataset lsun --data-dir LSUN --image-class bedroom`.
All options are described in the help below
```
usage: train.py [-h] --data-dir DATA_DIR --output-dir OUTPUT_DIR --model-dir
MODEL_DIR [--image-class IMAGE_CLASS] [--dataset {raw,lsun}]
[--restore] [--image-size IMAGE_SIZE]
[--state-size STATE_SIZE] [--batch-size BATCH_SIZE]
[--epochs EPOCHS]
[--generator-iterations GENERATOR_ITERATIONS]
[--critic-iterations CRITIC_ITERATIONS]
[--sample-every SAMPLE_EVERY]
[--gradient-penalty GRADIENT_PENALTY]
[--generator-lr GENERATOR_LR] [--critic-lr CRITIC_LR]
Wasserstein GAN
optional arguments:
-h, --help show this help message and exit
--data-dir DATA_DIR directory with the images
--output-dir OUTPUT_DIR
directory where to store the generated images
--model-dir MODEL_DIR
directory where to store the models
--image-class IMAGE_CLASS
class to train on, only for LSUN
--dataset {raw,lsun} format of the dataset
--restore restart training from the saved models
--image-size IMAGE_SIZE
image dimension
--state-size STATE_SIZE
state size
--batch-size BATCH_SIZE
batch size
--epochs EPOCHS number of epochs
--generator-iterations GENERATOR_ITERATIONS
number of iterations for the generator
--critic-iterations CRITIC_ITERATIONS
number of iterations for the critic
--sample-every SAMPLE_EVERY
how often to sample images
--gradient-penalty GRADIENT_PENALTY
gradient penalty
--generator-lr GENERATOR_LR
learning rate for the generator
--critic-lr CRITIC_LR
learning rate for the critic
```
## TensorboardX
Results such as costs, generated images for tensorboard will be written to `./runs` folder.
To display the results to tensorboard, run: `tensorboard --logdir runs`
## Acknowledgements
* [github.com/jalola/improved-wgan-pytorch](https://github.com/jalola/improved-wgan-pytorch)
* [igul222/improved_wgan_training](https://github.com/igul222/improved_wgan_training)
* [caogang/wgan-gp](https://github.com/caogang/wgan-gp)
* [LayerNorm](https://github.com/pytorch/pytorch/issues/1959)
## License
This project is licensed under [Apache2](https://opensource.org/licenses/Apache-2.0).
It is based on [improved-wgan-pytorch](https://github.com/jalola/improved-wgan-pytorch),
which is itself licensed under [MIT](https://opensource.org/licenses/MIT).