In my endeavour to learn Pytorch, I have come up with this simple implementation of a PixelCNN model (sequence based completion using convolutions only, introduced Oord et al 2016) on two very small datasets:
- A toy generative dataset of LCD digits https://github.com/pilipolio/learn-pytorch/blob/master/201708_ToyPixelCNN.ipynb
- the ~800 FontAwesome icons https://github.com/pilipolio/learn-pytorch/blob/master/201708_IconPixelCNN.ipynb
I would be very happy to translate these into a Pytorch tutorial if there is appetite for it.