The data set would be astronomy sub-images that are either bad (edge of chip artifacts, bright star saturation and spikes, internal reflections, chip flaws) or good (populated with fuzzy-dot stars and galaxies and asteroids and stuff). Let's say the typical image is 512x512 but it varies a lot. Because the bad features tend to be big, I'd probably like to bin the images down to say 64x64 for compactness and speed. It has to run fast on tens of thousands of images. I'm sort of tempted by the solution of adopting PlaidML as my back end (if I understand what its role is), because it can compile the problem for many architectures, like CUDA, CPU-only, OpenCL.