I’m looking for some pointers as to how one might go about sharing input-hidden weights for both directions of a recurrence with Bidirectional RNNs. Does anyone have suggestions about how that might be done?
Any guidance greatly appreciated. Thanks!
I think you would have to write your own RNNCell and do a bidirectional yourself.
Hello! I had to implement a BiRNN with shared input matrices (W_ih) and I did it successfully. However, I can not write such a class that uses the PackedSequence object.
Any recommendations?