Open howardgriffin opened 7 years ago
Yes, the paper also includes a PixelCNN++ density estimator. This seems to be part of the reason this code does not reach state of the art performance (14% test error on CIFAR-10).
I invite contributors to add this feature, as without it this code isn't properly reproducing the paper.
I think they are using PixelCNN++ density estimator for MNIST only.
The Generator formula seems to be different from the paper which only contains entropy(pull away).