Deviations from the paper

[arsenious]

Hi
Nice work with the code
Just wanted to document how your code differs from the paper.
For example, you have an extra a_t formed using W_add, b_add and h_t
The paper simply adds the key k_t in place of a_t
Also, it's not clear how they actually got k_t.

Also I assumed they had used CNNs to get features before going to the LSTM controller, since it was an image input