Robust evaluation

[wwarriner]

Currently evaluation occurs only on structured, non-overlapping patches. The "right" way to perform evaluation is by consensus of overlapping patches.

Based on the paper here: https://arxiv.org/abs/1504.07947 the maximum benefit/cost ratio is simple voting or max pooling over a range of patches that include a given pixel.

So the plan would be to pad the image, then evaluate the model on every possible patch centered on a pixel in the original image. We would probably not want to evaluate on every single pixel, that would mean for each N-pixel image we would need N evaluations on the GPU, which could be very expensive. We can mitigate this issue by adding a small-pixel stride so we only have N/K^2 patches to evaluate, where K is the stride length in pixels.

Storing the intermediate results is tricky, and there are a few options. The fastest may be a dict of dicts or ragged array of nested lists for each pixel subscript. Then in each we store the intermediate classifications/probabilities for the appropriate pixel.

Then to collapse the results we loop over a blank image copy. For each subscript, access the associated intermediate class/prob list and deal with it appropriately to produce a single class, and store that in the blank image copy.