Training a PixelCNN on Flickr8k images

Regenerate a portion of an image using PixelCNN.

Example Image

Flickr8k dataset, 8000 images 64x64 pixels

The batch size is set to 4 for a 2GB GPU. It should be increased to 32 or greater to leverage the full size of your GPU memory.

Execute training:

python pixelcnn_flickr8k.py

Execute to view loss chart and audio synthesis:

tensorboard --logdir=.

Select an image from the dataset, zero out the lower half and use the trained model to synthesize it.

import pixelcnn_flickr8k
pixelcnn_flickr8k.generate()