PixelCNN++

A Pytorch Implementation of PixelCNN++.

Main work taken from the official implementation

Pre-trained models are available here

I kept the code structure to facilitate comparison with the official code.

The code achieves 2.95 BPD on test set, compared to 2.92 BPD on the official tensorflow implementation.

Running the code

python main.py

Differences with official implementation

1. No data dependant weight initialization
2. No exponential moving average of past models for test set evalutation

Contact

For questions / comments / requests, feel free to send me an email.
Happy generative modelling :)