Hello, I want to modify the LSTM with a custom gate. I understand how to modify the LSTM cell. What I don’t understand is how would the back-propagation go through this modified cell, without me integrating it back with the pytorch LSTM layer? If it doesn’t, how do I integrate my changes into the original code?
If you’re written a
forward method using the Python API, you don’t really have to do much. Autograd will take care of backprop for you.