Magic line for visualising attention

[cipri-tom]

Hello,

Thank you for releasing your code ! It is a great contribution and I see it helps a lot of people (including me!).

I have a question WRT the visualize_attention() method, namely this line:

Attention-OCR/src/model/model.py

Line 447 in 88cff37

attention_orig = np.convolve(attention_orig, [0.199547,0.200226,0.200454,0.200226,0.199547], mode='same')

I believe its purpose is to highlight the important part and drive the rest to zero, similar to a softmax. I'm curious as to how you got the numbers for the filter. Any insight would be greatly appreciated !

Thanks !