Exploring GAN: Improving training techniques

Environment setup & dependency installation

git clone https://github.com/rikenmehta03/gan_mine.git
cd gan_mine
./install 

This commands will setup a virtual environment using python3 and install the required packages in the same environment. Install script will also create an alias for activating the virtual environment: cv_env

Repository structure

Every module will be a directory containing __init__.py file and other subfiles. Below are the initial modules we need to write.

• gan : This module contains trainer classes including generic gan trainer and BigGAN trainer.
• utils : This module contains utility functions or classes we write. For example Logger class in logger.py file.
• model : This module contain all the different architecure we tried for discriminator or generator.
• data_loader : This module provides wrapper for data-loader class for different datasets.
• evaluation_metric : Utility function and evaluation class used to evaluate various results.
• scripts : Ad-hoc scripts written to perform various experiments

Training script

Run these commands to run the training scripts

python main.py --model dcgan --dataset imagenet --data_path dataset/ --batch_size 64 --iters 100000 --log_iter 1000 --sn true --device cuda #train with spactral normalization
python main.py --dataset imagenet --data_path /var/www/dataset --batch_size 64 --iters 100000 --log_iter 1000 --sn true --device cuda --gpus 0,1,2,3 #Train with data parallel

Evaluation script

Training process logs the sample images every 500 iterations. Model is saved according to the parameter log_iter. The results are saved in logs folder. To run the evaluation on the results, run the command

evaluate.py --dataset imagenet --eval_folder imagenet_dcgan_sn_64_bs_64_18_12_2018_16:59:50 --device cuda --data_path dataset/ 

All other configurable parameters are available in parameters.py file.

Results

• DCGAN on celeb-HQ

• DCGAN on ImageNet*

• BigGAN on LSUN (Church outdoor & Bridge)