improve occlusion_heatmap() performance

[rldw]

I'd like to use the occlusion_heatmap() function in visualize.py to drastically improve my prediction performance. However, for a 45x90px CMYK image this takes roughly 2s. Considering that I have a couple of thousand images in my test set this is not really practical. So is there a way to improve the function's performance?

I'm not too familiar with numpy, but I assume that one could vectorize the nested for-loop?

for i in range(s0): for j in range(s1):

[dnouri]

I haven't used occlusion_heatmap() too much. I assume you're running the code on the GPU? You could try to profile it, or just sprinkle some print(time() - t0) over the code to understand better where time is spent.

[BenjaminBossan]

Vectorizing the nested for-loop would not help since the slow part - the call to predict_proba - is already outside the inner loop. The problem is that for each image, you have 45x90 predictions, so I would expect the function call to be approximately as slow as calling predict_proba on 4050 samples (if not, please report).

What could be done? I see two solutions:

1. Don't predict for all images - I at least just look at a couple of them to get a feel.
2. Don't use occlusion_heatmap but instead propagate the error back on the image; this does not provide the same information, but it might still help and at least it should be much faster.

[dnouri]

By the way, the second option is being implemented in #223.