Possible error in reimplementation

[vvigilante]

With reference to the original implementation
https://github.com/kracwarlock/action-recognition-visual-attention/blob/6738a0e2240df45ba79e87d24a174f53adb4f29b/src/actrec.py#L111

it looks like to me that they use this structure for decoding the lstm:
Dense(100) -> tanh -> Dense(n_classes) -> softmax
while you implemented in the function _decode_lstm:
Dense(n_classes) -> tanh -> Dense(n_classes) -> softmax

I think that the first, hidden fully connected layer has too few neurons, and that is it different from the original implementation.