GP-GAN: Gender-Preserving-GAN-for-Synthesizing-Faces-from-Landmarks

This repository is about out ICPR work, GP-GAN: Gender Preserving GAN for Synthesizing Faces from Landmarks

Prerequisites

This code has tested on Ubuntu 16/18 with Pytorch 0.4 and cuda 9.0/8.0

Getting Started

Clone this repo:

• git clone https://github.com/DetionDX/GP-GAN-GenderPreserving-GAN-for-Synthesizing-Faces-from-Landmarks.git
• cd GP-GAN-GenderPreserving-GAN-for-Synthesizing-Faces-from-Landmarks

Download dataset

• cd datasets
• bash download_lfw_landmark_dataset.sh

Training

change directory into the cloned folder

• cd ..
• python train.py --dataroot ./datasets/lfw/ --which_direction BtoA --fineSize 64 --loadSize 64 --no_flip --name lfw_gpgan

To view training results and loss plots, run:

• python -m visdom.server

and click the URL http://localhost:8097

If you have prblem to download the pretrained vgg16.t7 file. You can download here: and put it into directory: ./models

Testing

python test.py --dataroot ./datasets/lfw/ --which_direction BtoA --fineSize 64 --loadSize 64 --name lfw_gpgan --which_epoch 10

Reference

@INPROCEEDINGS{di2018gp, 
author={Xing Di and Vishwanath A. Sindagi and Vishal M. Patel}, 
booktitle={2018 24th International Conference on Pattern Recognition (ICPR)}, 
title={GP-GAN: Gender Preserving GAN for Synthesizing Faces from Landmarks}, 
year={2018},
pages={1079-1084}, 
month={Aug}
}

Acknowledgments

This is work is highlg inspired by pytorch-CycleGAN-and-pix2pix

We highly thanks to He Zhang for his discussing about DenseUnet part