Unsupervised Image to Image Translation Networks

This is the TensorFlow Implementation of the NIPS 2017 paper "Unsupervised Image to Image Translation Networks" by Harry Yang.

Disclaimer: This was our own research project but it shares the same idea with the paper so we are making the code publicly available.

Introduction

In particular, we tried the 'pix2pix' model which is the auto-encoder model described in the paper, and also the 'resnet' model made up of 9 blocks of resnet (middle blocks are shared). We found that the 'resnet' model gives better result than auto-encoder and (slightly) better results than CycleGAN.

Below is a snapshot of our result at the 44th epoch using 'resnet' model with default parameters on the horse-zebra dataset. From left to right is: real horse, real zebra, fake zebra, fake horse, cycle horse, cycle zebra.

Getting Started

Prepare dataset

Download horse-zebra dataset (or any other dataset) and create a csv file containing the paths of image pairs in the dataset. -Download an image dataset (e.g. horse-zebra) and unzip:

bash ./download_datasets.sh horse2zebra

-Create csv files

python -m dataset.create_single_domain_filelist --input_path=path/to/horse/train --output_file=/data/img2img/horse.csv
python -m dataset.create_single_domain_filelist --input_path=path/to/zebra/train --output_file=/data/img2img/zebra.csv
python -m dataset.create_image_pair_list --first_dataset=/data/img2img/horse.csv --second_dataset=/data/img2img/zebra.csv --output_file=/data/img2img/horse_zebra.csv

• Modify config.py based on the location of the csv file (the trained models will also be saved in the folder).

Train CycleGAN

• Use resnet model with default parameters:

python -m main --split_name='horse_zebra' --cycle_lambda=15 --rec_lambda=1 --num_separate_layers_g=2 --num_separate_layers_d=5 --num_no_skip_layers=0 --lsgan_lambda_a=1 --lsgan_lambda_b=1 --network_structure='resnet'

• Use autoencoder model with default parameters:

python -m main --split_name='horse_zebra' --cycle_lambda=30 --rec_lambda=10 --num_separate_layers_g=3 --num_separate_layers_d=5 --num_no_skip_layers=0 --lsgan_lambda_a=2 --lsgan_lambda_b=2

Restore from previous checkpoints

python -m main --split_name='horse_zebra' --cycle_lambda=15 --rec_lambda=1 --num_separate_layers_g=2 --num_separate_layers_d=5 --num_no_skip_layers=0 --lsgan_lambda_a=1 --lsgan_lambda_b=1 --network_structure='resnet' --checkpoint_dir=path/to/saved/checkpoint

TensorBoard Output

Visualization

Each epoch saves an html file for better visualization.