-
the network output is actually a mixture of logistics following the PixelCNN++ paper by Salimans et al.
-
I took the code for loss computation from the OpenAI pixelcnn_pp repository for stable loss computation but I implemented it myself as well
- for reasons of stability the OpenAI loss provides better samples
-
here's a sample from a network trained with the OpenAI loss on the CIFAR-10 dataset:
