ICCV paper of DualGAN

DualGAN: unsupervised dual learning for image-to-image translation

please cite the paper, if the codes has been used for your research.

architecture of DualGAN

How to setup

Prerequisites

• Linux

• Python (2.7 or later)

• numpy

• scipy

• NVIDIA GPU + CUDA 8.0 + CuDNN v5.1

• TensorFlow 1.0 or later

Getting Started

steps

• clone this repo:

git clone https://github.com/duxingren14/DualGAN.git

cd DualGAN

• download datasets (e.g., sketch-photo), run:

bash ./datasets/download_dataset.sh sketch-photo

• download pre-trained model (e.g., sketch-photo), run:

bash ./checkpoint/download_ckpt.sh sketch-photo

• train the model:

python main.py --phase train --dataset_name sketch-photo --image_size 256 --lambda_A 1000.0 --lambda_B 1000.0 --epoch 100

• test the model:

python main.py --phase test --dataset_name sketch-photo --image_size 256 --lambda_A 1000.0 --lambda_B 1000.0 --epoch 100

optional

Similarly, run experiments on facades dataset with the following commands:

bash ./datasets/download_dataset.sh facades

python main.py --phase train --dataset_name facades --lambda_A 1000.0 --lambda_B 1000.0 --epoch 100

python main.py --phase test --dataset_name facades --lambda_A 1000.0 --lambda_B 1000.0 --epoch 100

For thoese who cannot download datasets or pretrained models using the scripts, please try manual downloading from the link as below:

all datasets from google drive

pretrained models from google drive

Experimental results:

Acknowledgments

Codes are built on the top of pix2pix-tensorflow and DCGAN-tensorflow. Thanks for their precedent contributions!