discogan

DiscoGAN

Because the authors call their method DiscoGAN!

PyTorch implementation of Learning to Discover Cross-Domain Relations with Generative Adversarial Networks.

Generates an image in one domain given another image in another domain, without explicitly paired training data.

Cross-domain relations are often natural to humans. For example, we recognize the relationship between an English sentence and its translated sentence in French. We also choose a suit jacket with pants or shoes in the same style to wear. Can machines also achieve a similar ability to relate two different image domains?

This GAN-based model trains with two independently collected sets of images and learns how to map two domains without any extra label. For example, the model takes a handbag image as an input, and generates its corresponding shoe image in the same style.

The core of the model is based on two different GANs coupled together – each of them ensures our generative functions can map each domain to its counterpart domain. A key intuition is to constraint all images in one domain to be representable by images in the other domain. For example, when learning to generate a shoe image based on each handbag image, we force this generated image to be an image-based representation of the handbag image (and hence reconstruct the handbag image) through a reconstruction loss, and to be as close to images in the shoe domain as possible through a GAN loss.

Creating your own dataset

You can create your own dataset by organizing images as:

data
└── YOUR_DATASET_NAME
    ├── A
    │   ├── xxx.jpg (name doesn't matter)
    │   ├── yyy.jpg
    │   └── ...
    └── B
        ├── zzz.jpg
        ├── www.jpg
        └── ...

All images in each dataset should have same size:

# for Ubuntu
$ sudo apt-get install imagemagick
$ mogrify -resize 256x256! -quality 100 -path YOUR_DATASET_NAME/A/*.jpg
$ mogrify -resize 256x256! -quality 100 -path YOUR_DATASET_NAME/B/*.jpg

# for Mac
$ brew install imagemagick
$ mogrify -resize 256x256! -quality 100 -path YOUR_DATASET_NAME/A/*.jpg
$ mogrify -resize 256x256! -quality 100 -path YOUR_DATASET_NAME/B/*.jpg

# for scale and center crop
$ mogrify -resize 256x256^ -gravity center -crop 256x256+0+0 -quality 100 -path ../A/*.jpg

To train a model (from the project root):

export PYTHONPATH=src
python src/pytorch_model/discogan/__init__.py --dataset=edges2shoes --num_gpu=1
python src/pytorch_model/discogan/__init__.py --dataset=YOUR_DATASET_NAME --num_gpu=4

To test a model, use your load_path:

export PYTHONPATH=src
python src/pytorch_model/discogan/__init__.py --dataset=edges2handbags \
    --load_path=src/pytorch_model/discogan/logs/edges2handbags_2017-07-28_10-55-37 \
    --num_gpu=0 ==is_train=False