Progressive Growing of GANs for Improved Quality, Stability, and Variation

This is PyTorch implementation of ProgressiveGAN described in paper "Progressive Growing of GANs for Improved Quality, Stability, and Variation".

Work is in progress. Equivalent lr layers still don't work well.

Usage

Config

Use config.json file for set up model before training.

• exp_name - model name
• batch - batch size
• latent - size of latent space vector
• isize - final generating image size
• epochs - number of epochs
• lr_d - lerning rate of discriminator
• lr_g - lerning rate of generator
• lr_decay_epoch - []
• weights - using in generate.py script

Training

To begin trainig use train.py script.

• device_ids - GPU ids. Use list for initialize
• device - use GPU or CPU for training
• data_path - path of dataset

Train on cpu:

python train.py -c config.json -d cpu

Train on 2 gpus:

python train.py -c config.json -d cuda --dev_ids 0 1

Runing

To use generator run generate.py script:

python generate.py -o out/test/ -c config.json -n 20

Example of generating cats

Compatability

• Python 3.7.3
• PyTorch 1.7.1
• CUDA 10.1
• CUDNN 7.6.3

Acknowledgement

[1] https://arxiv.org/abs/1710.10196 [2] https://github.com/nashory/pggan-pytorch