DR-GAN-by-pytorch

Disentangled Representation Learning GAN for Pose-Invariant Face Recognition

• Authors: Luan Tran, Xi Yin, Xiaoming Liu
• CVPR2017: http://cvlab.cse.msu.edu/pdfs/Tran_Yin_Liu_CVPR2017.pdf
• Pytorch implimentation of DR-GAN (updated version in "Representation Learning by Rotating Your Faces")
• Added a pretrained ResNet18 to offer a feature loss in order to improve Generator's performance. (Only in Multi_DRGAN)

Requirements

• python 3.x
• pytorch 0.2
• torchvision
• numpy
• scipy
• matplotlib
• pillow
• tensorboardX

How to use

Single-Image DR-GAN

1. Modify model function at base_options.py to define single model.

   • Data needs to have ID and pose lables corresponds to each image.
   • If you don't have, default dataset is CFP_dataset. Modify dataroot function at base_options.py.

2. Run train.py to train models

   • Trained models and Loss_log will be saved at "checkpoints" by default. Generated pictures will be saved at "result".

     python train.py

   • You can also use tensorboard to watch the loss graphs in real-time. (Install tensorboard before doing it.)

     tensorboard --logdir=/home/zhangjunhao/logs (Or the address of dir 'logs' in your folder.）

3. Generate Image with arbitrary pose

   • Change the "save_path" in base_model.py.
   • Specify leaned model's filename by "--pretrained_G" option in base_options.py.
   • Generated images will be saved at specified result directory.

     python test.py

Multi-Image DR-GAN

1. Modify model function at base_options.py to define multi model.

   • Data needs to have ID and pose lables corresponds to each image.
   • If you don't have, default dataset is CFP_dataset. Modify dataroot function at base_options.py.

2. Run train.py to train models

   • Trained models and Loss_log will be saved at "checkpoints" by default. Generated pictures will be saved at "result".

     python train.py

   • You can also use tensorboard to watch the loss graphs in real-time. (Install tensorboard before doing it.)

     tensorboard --logdir=/home/zhangjunhao/logs (Or the address of dir 'logs' in your folder.）

3. Generate Image with arbitrary pose

   • Change the "save_path" in base_model.py.
   • Specify leaned model's filename by "--pretrained_G" option in base_options.py.
   • Generated images will be saved at specified result directory.

     python test.py