gumpfly / cross-domain-disen

Cross-domain disentanglement network

Cross-domain disentantanglement network

Code for the paper "Image-to-image translation for cross-domain disentanglement", NeurIPS 2018.

Based on this pix2pix implementation by Christopher Hesse, extensively explained in this article.

Setup

Please follow the setup described here. Tested with Tensorflow 1.8.0.

See DATA/MNISTCDCB/ for example images of our MNIST-CD/CB dataset.

Usage

In order to train a MODEL using DATA, run

python run_cross_domain_disen.py \
  --mode train \ 
  --output_dir checkpoints/MODEL \ 
  --input_dir DATA/train/  

Once the model finished training, it can be tested by running

python run_cross_domain_disen.py \ 
  --mode test \ 
  --output_dir test/MODEL \
  --checkpoint checkpoints/MODEL \
  --input_dir DATA/test/  

In order to extract disentangled features for other tasks (e.g. cross-domain retrieval), run

python run_cross_domain_disen.py \ 
  --mode features \ 
  --output_dir features/MODEL \ 
  --checkpoint checkpoints/MODEL \ 
  --input_dir DATA/test/  

Pre-trained models

MNIST-CD/CB (w/ GRL)

MNIST-CD/CB (w/o GRL)

Cars

Chairs

Facades

Maps

Dataset

MNIST-CD/CB

Citation

Please, cite the following paper if you use this code:

@inproceedings{gonzalez-garcia2018NeurIPS,
  title={Image-to-image translation for cross-domain disentanglement},
  author={Gonzalez-Garcia, Abel and van de Weijer, Joost and Bengio, Yoshua},
  booktile={NeurIPS},
  year={2018}
}