Whitening and Coloring transform for GANs

Check out our paper

Unsupervised Cifar-10 (IS 8.66)

Supervised Cifar-10 (IS 9.06)

Requirments

• python2
• numpy
• scipy
• skimage
• pandas
• tensorflow == 1.5.0 (I have not testted with other versions)
• keras == 2.0.8 (I have tried latter versions, but they throw a bug. Not sure from where it came from)
• tqdm

For the commands reproducing experiments from the paper check scripts folder.

All scripts has the following name: (name of the dataset) + (architecture type (resnet or dcgan)) + (discriminator normalization (sn or wgan_gp)) + (conditional of unconditional) + (if conditional use soft assigment (sa)).

For example:

CUDA_VISIBLE_DEVICES=0 scripts/cifar10_resnet_sn_uncond.sh

will train GAN for cifar10 dataset, with resnet architecture, spectral normalized discriminator in unconditional case.

All dataset except for imagenet downloaded and trained at the same time.

Imagenet

0. This will consume a loot of memory. Because dataset is packed into numpy files for sequential reads.
1. Download imagenet ILSVRC2012. Train and val. Put train to ../ILSVRC2012/train, and val to ../ILSVRC2012/val/val (val/val is important)
2. Preprocess imagenet train:

bash preprocess.sh ../ILSVRC2012/train ../imagenet-resized

3. Preprocess imagenet val:

bash preprocess.sh ../ILSVRC2012/val ../imagenet-resized-val

4. Now you can remove ILSVRC2012
5. CUDA_VISIBLE_DEVICES=0 scripts/imagenet_resnet_sn_cond_sa.sh This will first pack imagenet into numpy files, and then start traning.

Citation:

@inproceedings{
	siarohin2018whitening,
	title={Whitening and Coloring transform for {GAN}s},
	author={Aliaksandr Siarohin and Enver Sangineto and Nicu Sebe},
	booktitle={ICLR},
	year={2019},
	url={https://openreview.net/forum?id=S1x2Fj0qKQ}
}