Skip to content

Latest commit

 

History

History
37 lines (30 loc) · 1.12 KB

README.md

File metadata and controls

37 lines (30 loc) · 1.12 KB

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