Equivalent of Keras's `mask` arg in the forward pass of an RNN

1

I’m looking for the equivalent of the mask arg for the forward pass from Keras’s RNN implementation in PyTorch. It seems that PyTorch has the pack_padded_sequence function to skip over the end of padded sequences, but Keras has some additional functionality that I need, which is the ability to skip RNN computations at arbitrary time points and propagate the previous time step to the next time step according to the mask arg.

1 Like
2

Let me know if you wind a solution or workaround to this. I am actually struggling with exactly the same.