PSA for rnn_cell_backward from C5W1A1E5

I got all of the dimensions of the gradients correct, but their values seemed to be a bit off. It turned out that I replaced tanh(np.square(Wax @ xt + Waa @ a_prev + ba) ) with np.tanh(np.square(a_next)) instead of just np.square(a_next). So the bottom line is to remeber how a_next was calculated and avoid excessive tanh calculations.

1 Like

HI @Tom_Pham

Thanks for sharing. It is great that you have resolved the problem.

1 Like