biggan_cifar

LeCam Regularization for training BigGAN on CIFAR

Implementation of our regularization method for training the BigGAN model under the limited CIFAR dataset.

Installation

Clone this repository

git clone https://github.com/google/lecam-gan.git
cd lecam-gan/biggan_cifar

Install packages, refer to the Pytorch webpage for installing with different CUDA versions

conda create --name lcgan_pytorch python=3.6
conda activate lcgan_pytorch
conda install pytorch torchvision cudatoolkit=10.1 -c pytorch
conda install -c anaconda tensorflow-gpu==1.14
pip install -r requirements.txt

Training

Please refer to the scripts we provide in the scripts folder. For example, training the model on the 20% CIFAR-10 dataset with our regularization:

CUDA_VISIBLE_DEVICES=0,1 bash ./scripts/lc-biggan-cifar10-0.2.sh

Testing

Calculating the IS/FID scores with three evaluation runs:

CUDA_VISIBLE_DEVICES=0,1 python eval.py --repeat 3 --dataset C10 --network
weights/lc-biggan-cifar10-0.2/G_ema_best.pth

You can change the trained model file and dataset using the --network and --dataset commands.

Notes

This repository is built based on the implementation from DiffAug.