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).