DiscoGAN in PyTorch

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

Requirements

Usage

First download datasets (from pix2pix) with:

$ bash ./data/download_dataset.sh dataset_name

facades: 400 images from CMP Facades dataset.
cityscapes: 2975 images from the Cityscapes training set.
maps: 1096 training images scraped from Google Maps
edges2shoes: 50k training images from UT Zappos50K dataset.
edges2handbags: 137K Amazon Handbag images from iGAN project.

or you can use your own dataset by placing images like:

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

To train a model:

$ python main.py --dataset=edges2shoes --num_gpu=1
$ python main.py --dataset=YOUR_DATASET_NAME --num_gpu=4

To test a model (use your load_path):

$ python main.py --dataset=edges2handbags --load_path=logs/edges2handbags_2017-03-18_10-55-37 --num_gpu=0 --is_train=False

Results

1. Toy dataset

Result of samples from 2-dimentional Gaussian mixture models. IPython notebook

# iteration: 0:

# iteration: 10000:

2. Edges2shoes dataset

# iteration: 9600:

x_A -> G_AB(x_A) -> G_BA(G_AB(x_A)) (color -> sketch -> color)

x_B -> G_BA(x_B) -> G_AB(G_BA(x_B)) (sketch -> color -> sketch)

3. Edges2handbags dataset

# iteration: 9500:

x_A -> G_AB(x_A) -> G_BA(G_AB(x_A)) (color -> sketch -> color)

x_B -> G_BA(x_B) -> G_AB(G_BA(x_B)) (sketch -> color -> sketch)

4. Facades dataset

Reconstruction looks quite weird which is not included in the paper.

# iteration: 19450:

x_B -> G_BA(x_B) -> G_AB(G_BA(x_B)) (image -> segmentation -> image)

x_A -> G_AB(x_A) -> G_BA(G_AB(x_A)) (segmentation -> image -> segmentation)

5. Cityscapes dataset

# iteration: 8350: