Balanced Training for Sparse GANs [Paper]

Yite Wang*, Jing Wu*, Naira Hovakimyan, Ruoyu Sun

$*$ denotes equal contribution.

In NeurIPS'2023

Requirements:

The code is tested using Redhat system with python 3.9. NVIDIA V100 and NVIDIA RTX 2080TI are used to run all the experiments. To install required packages, please find the requirements.txt file.

Prepare dataset

1. CIFAR-10 and STL-10 datasets will download automatically.

2. Modify folder location of IS computation MODEL_DIR under sparselearning/gan_utils/inception_score.py.

3. Download FID statistics from this repo of GNGAN.

Run our code:

Please see the scripts in scripts folder to run our code. For more information, please refer to main.py and sparselearning/core.py.

For example, to run the baseline:

chmod +x scripts/baseline1.sh
scripts/baseline1.sh

Acknowlegement:

Our code is mainly based on :

ITOP and GAN ticket.

Contact:

Yite Wang (yitew2@illinois.edu)

Jing Wu (jingwu6@illinois.edu)