WGAN-QC

PyTorch implementation of Wasserstein GAN With Quadratic Transport Cost, ICCV 2019.
The code is build under Python 3.6, PyTorch 1.0

Usage

DATASET: CelebA/CelebA-HQ/LSUN/ImageNet dog
DATASET_PATH: path to the root folder of a dataset
OUTPUT_DIR: the directory of your output
GPU_ID: gpu id

1. Install dependency: CVXOPT for linear programming.
   
2. Download your dataset, unzip it and put it in DATASET_PATH.
   

On the CelebA dataset, run

python wgan_qc_resnet2.py --dataset celeba --dataroot DATASET_PATH --output_dir OUTPUT_DIR --batchSize 64 --imageSize 64 --Giters 60000 --gamma 0.1 --EMA_startIter 55000 --gpu_ids GPU_ID

On the CelebA-HQ dataset, run

python wgan_qc_resnet1.py --dataset celebahq --dataroot DATASET_PATH --output_dir OUTPUT_DIR --batchSize 16 --imageSize 256 --Giters 125000 --gamma 0.05 --EMA_startIter 120000 --genImg_num 16 --gpu_ids GPU_ID

On the LSUN dataset, run

python wgan_qc_resnet2.py --dataset lsun --dataroot DATASET_PATH --output_dir OUTPUT_DIR --batchSize 64 --imageSize 64 --Giters 100000 --gamma 0.1 --EMA_startIter 95000 --pin_mem --gpu_ids GPU_ID

On the ImageNet dog subset, run

python wgan_qc_resnet2.py --dataset dog --dataroot DATASET_PATH --output_dir OUTPUT_DIR --batchSize 64 --imageSize 128 --Giters 150000 --gamma 0.1 --EMA_startIter 145000 --lr_anneal 0.1 --milestones 120000,200000 --gpu_ids GPU_ID

For other dataset that the pairwise data distance is small, set

--K 1.0

and increase --gamma

If you have severe checkerboard artifact issues, try to add the following options

--res_ratio 1.0 --DOptIters 5

For any problem, please contact Huidong Liu at [email protected] or [email protected]

You are welcome to cite our work using:

@InProceedings{Liu_2019_ICCV,
author = {Liu, Huidong and Gu, Xianfeng and Samaras, Dimitris},
title = {Wasserstein GAN With Quadratic Transport Cost},
booktitle = {The IEEE International Conference on Computer Vision (ICCV)},
month = {October},
year = {2019}
}