Localized Generative Adversarial Nets (LGAN) for Image Generation with Diversity

Author:Liheng Zhang, Date: 11/21/2017

This is the project for the following technical report:

Guo-Jun Qi, Liheng Zhang, Hao Hu. Global versus Localized Generative Adversarial Nets. arXiv: 1711.06020 [[pdf] (https://arxiv.org/pdf/1711.06020.pdf)]

Questions about the source codes can be directed to Liheng Zhang at lihengzhang1993@knights.ucf.edu.

Requirements

• Python == 2.7
• Pytorch == 0.2.0_4

For celebA dataset

1. Setup and download dataset

mkdir celebA; cd celebA

Download img_align_celeba.zip from http://mmlab.ie.cuhk.edu.hk/projects/CelebA.html under the link "Align&Cropped Images".

unzip img_align_celeba.zip; cd ..

2. Train LGAN

python train.py --dataset folder --dataroot ./celebA --imageSize 64 --nz 32 --nc 3 --cuda --outf "./results/celebA"

3. Densely sample images with diversity

python dense_sample.py --dataset folder --dataroot ./celebA --imageSize 64 --nz 32 --nc 3 --cuda --netG "./results/celebA/netG_epoch_24.pth" --outf "./results/celebA" 

For mnist dataset

1. Train LGAN

python train.py --dataset mnist --dataroot ./mnist --imageSize 32 --nz 10 --nc 1 --lrD 0.0001 --lrG 0.0005 --cuda --outf "./results/mnist"

2. Densely sample images with diversity

python dense_sample.py --dataset mnist --dataroot ./mnist --imagesSize 32 --nz 10 --nc 1 --cuda --netG "./results/mnist/netG_epoch_24.pth" --outf "./results/mnist"

Acknowledge

Parts of codes are reused from DCGAN at https://github.com/pytorch/examples/tree/master/dcgan