A question on rnn_layer about cs231n-2017

[Burton2000]

Looking at that bit of code everything seems fine there. Do the tests for rnn_step_backward function all pass? Also is dprev_h initialized to zeros before the for loop?

[nihuo76]

Hi, thank you for your response. The tests for rnn_step_backward is successful and returns a error at e-9. I also intialize dprev_h to be zero before the for loop.

[Burton2000]

After the for loop do you return the last hidden state gradient, so you set dh0 = dprev_h

[nihuo76]

Hi, I did that. But a weird result happens that all gradient with respect to x,Wx,Wh are wrong in my version. It is just so weird.