PixelCNN++: Improving the PixelCNN with Discretized Logistic Mixture Likelihood and Other Modifications

PixelCNNs are a recently proposed class of powerful generative models with likelihood. Here we discuss our implementation of PixelCNNs which we available at https: //github. com/openai/pixel-cnn. Our implementation a number of modifications to the original model that both simplify its and improve its performance. 1) We use a discretized logistic mixture on the pixels, rather than a 256-way softmax, which we find to speed training. 2) We condition on whole pixels, rather than R/G/B sub-pixels, the model structure. 3) We use downsampling to efficiently capture at multiple resolutions. 4) We introduce additional short-cut to further speed up optimization. 5) We regularize the model using. Finally, we present state-of-the-art log likelihood results on-10 to demonstrate the usefulness of these modifications.