Skip to content

Latest commit

 

History

History
21 lines (18 loc) · 846 Bytes

README.md

File metadata and controls

21 lines (18 loc) · 846 Bytes

PixelCNN

PixelCNN on grayscale celeb/mnist/cifar 28x28.

python train.py --ds mnist
python train.py --ds cifar
python train.py --ds celeb # assumes celeb faces can be loaded from 'celeb.npz' 

Code is based on this repository.

Changes

  • uses command line arguments with argparse instead of config files
  • added cifar/celeb as additional datasets to mnist.
  • integrated with tensorboard
  • wrote code for inpainting
  • made the above animation of generation process (code for animation not included).
  • wrote test case for likelihood computations.
  • changed hyperparameter slightly (increased nlayers and ksize)
  • simplified code and moved into a single file (besides dataloading that also has its own file).