[samcodes]

The hard part is that after the convolutions you want a fully connected layer or two, and to get those dimensions right you need to know the input dimensions. But, pytorch is building the graph at runtime, so maybe you could do this...