redheadm/dr-gan

Disentangled Representation Learning GAN for Pose-Invariant Face Recognition

* Still under Development and haven't check the result of this implimentation (2017/9/25)

• 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")
• Powered by DL HACKS

Requirements

• python 3.6
• pytorch 0.2.0
• numpy 1.13.1
• scipy 0.18.1
• matplotlib 2.0.0

How to use

Single-Image DR-GAN

1. modify DataLoader function at main.py to define dataloader which is applicable to your data

   • data needs to have ID and pose lables corresponds to each image
   • if you don't have, "-random" option allow you to see how the code works with meanless random data.

   python main.py -random

2. Run main.py to train models

   • trained models and Loss_log will be saved at "DR_GAN/snapshot/Single" by default

   python main.py -random

3. Generate Image with arbitrary pose

   • use "-generate" option
   • specify leaned model by "-snapshot" option
   • generated images will be saved at specified sanpshot directory

   python main.py -random -generate -snapshot=snapshot/Single/2017-09-22_20-31-08/epoch1

Multi-Image DR-GAN

1. modify DataLoader function at main.py to define dataloader which is applicable to your data

   • data needs to have ID and pose lables corresponds to each image
   • if you don't have, "-random" option allow you to see how the code works with meanless random data.

   python main.py -multi-DRGAN -images-perID=4 -random

2. Run main.py with "-multi-DRGAN" and "-images-perID" option

   • Multi-Image DR-GAN assumes input data to have N images per person and in my code, they should be sequentially aligned. So change N depends on your data.
   • input data size have to be divisible by batch size
   • batch size have to be divisible by images_perID
   • trained modles and Loss_log will be saved at "DR_GAN/snapshot/Multi" by default

   python main.py -multi-DRGAN -images-perID=4 -random

3. Generate Image with arbitrary pose

   • use "-generate" option
   • specify leaned model by "-snapshot" option
   • generated images will be saved at specified sanpshot directory

   python main.py -random -multi-DRGAN -generate -images-perID=4 -snapshot=snapshot/Multi/2017-09-22_23-03-50/epoch5