Skip to content

PyTorch implementation of "Conditional Image Generation with PixelCNN Decoders" by van den Oord et al. 2016

License

Notifications You must be signed in to change notification settings

kkleidal/GatedPixelCNNPyTorch

Repository files navigation

PyTorch: Gated PixelCNN Conditional Autoregressive Decoder

Implementation of "Conditional Image Generation with PixelCNN Decoders" by van den Oord et al. 2016 with some modifications suggested by "Generating Interpretable Images with Controllable Structure" by Reed et al. 2017.

NOTE: currently, this project does not support masking and autoregression across multiple channels. As such, it can only model the generation of 1 channel 2D tensors. I may implement this in the future and contributions are welcome.

The implementation is heavily based on the Chainer implementation by Sergei Turukin and his helpful blog post.

Example usage and samples from a trained model will be added eventually; I'm in the process of using this for a class final project and need to focus on that for the time being but will come back and flesh out this repo when I have time in a few weeks.

Important note on masks

I use slightly different convolution masks than van den Oord and Turukin. Please take a look at this notebook to see my explanation of the masks and how they prevent blindspots.

About

PyTorch implementation of "Conditional Image Generation with PixelCNN Decoders" by van den Oord et al. 2016

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published