TP-GAN

pytorch replicate of TP-GAN "Beyond Face Rotation: Global and Local Perception GAN for Photorealistic and Identity Preserving Frontal View Synthesis"

what's different from the official code

• I use wasserstein-GP as adversial loss
• I tried adopting modified ReNet18 or MobilNetV2 to extract features to compute perceptual loss (idendity preserving loss in the original paper)
• remove batch normalization layers
• remove the last tanh activation in generator
• change the first conv and the first residual block in decoder of generator's kernel size from 2 to 3

requirements

• python3
• tensorboardX
• pytorch 0.3.1

usage

to train feature extract models

vim pretrain_config.py #set options
python pretrain.py

to train TP-GAN

vim config.py #set options
python train.py

to test TP-GAN

python test.py $args

some other implementations

• official tensorflow implementation
• tensorflow replicate