Pytorch implementation of A Style-Based Generator Architecture for Generative Adversarial Network

Requirements

• Python3
• Pytorch >= 1.0.0
• TensorBoardX
• fire
• apex [optional]
• pyyaml

Usage

train

python main.py 
    --config_file=path_to_config_file
    --checkpoint=path_to_config_file[default='']

inference

python main.py 
    --config_file=path_to_config_file
    --run_type=inference

Default configuration file is located in config directory.

Currently completed task

• [x] Progressive method
• [x] Tuning
• [x] Add mapping and styles
• [x] Remove traditional input
• [x] Add noise inputs
• [x] Mixing regularization

Fake image and real image score graph

fp32 precision

mixed precision

There seems to be no difference in the score.

Discriminator loss

fp32 precision

mixed precision

There is a problem with R1 regularization, so training does not work properly. This also affects image samples. It would be better not to use it now.

Train speed

There seems to be a clear speed difference depending on the precision, but it seems to be meaningless because the mixed precision training isn't done properly.

Inference Images

8x8 images

16x16 images

32x32 images

64x64 images

128x128 images

256x256 images

Pretrained checkpoint

~256x256~ See #1