This repository collects chainer implementation of state-of-the-art GAN algorithms.
These codes are evaluated with the inception score on Cifar-10 dataset.
Note that our codes are not faithful re-implementation of the original paper.
Install the requirements first:
pip install -r requirements.txt
This implementation has been tested with the following versions.
python 3.5.2
chainer 2.0.0
cupy 1.0.0.1
tensorflow 1.2.0 # only for downloading inception model
numpy 1.11.1
Download the inception score module forked from https://github.com/hvy/chainer-inception-score.
git submodule update -i
Download the inception model.
cd common/inception
python download.py --outfile inception_score.model
You can start training with train.py.
python train.py --gpu 0 --algorithm dcgan --out result_dcgan
Please see example.sh to train other algorithms.
| Inception | Inception (Official) | FID | |
|---|---|---|---|
| Real data | 12.0 | 11.24 | 3.2 (train vs test) |
| SN-DCGAN | 7.5 | 7.41 | 23.6 |
| WGAN-GP | 6.8 | 7.86 (ResNet) | 28.2 |
| DFM | 7.3 | 7.72 | 30.1 |
| Cramer GAN | 6.4 | - | 32.7 |
| DRAGAN | 7.1 | 6.90 | 31.5 |
| DCGAN-vanilla | 6.7 | 6.16 [WGAN2] 6.99 [DRAGAN] | 34.3 |
| minibatch discrimination | 7.0 | 6.86 (-L+HA) | 31.3 |
| BEGAN | 5.4 | 5.62 | 84.0 |
Inception scores are calculated by average of 10 evaluation with 5000 samples.
FIDs are calculated with 50000 train dataset and 10000 generated samples.
- SN-DCGAN
- WGAN-GP
- DFM
- Cramer GAN
- DRAGAN
- DCGAN
- Minibatch discrimination
- BEGAN
MIT License. Please see the LICENSE file for details.