unused hidden variables of LSTM on deep_learning_nlp_tutorial.py examples

[antran89]

Hi,

I am just curious about line 1197.

self.hidden = self.init_hidden()

This value has not been used to update parameters of network in the program. But you are saying also, we need to clear out the hidden state of the LSTM, detaching it from its history on the last instance. Is it correct?

Best regards,

[chsasank]

@rguthrie3, would you mind answering the above?

PS: After reorganisation of tutorial, part with this code is here

[Varal7]

I believe this is due to a typo in line 178.

As we can see in Pytorch's code, the nn.lstm model's arguments (h0, c_0) are optional.
If no arguments are passed, then the model creates a hidden state and cell state based on input.

In this case, we want our hidden state to be passed to the next iteration. Hence, #63.

[antran89]

@Varal7 Thank for clarification.

[rguthrie3]

I fixed this bug in my version of the repo, but I think that the version in this repo was copied over before I fixed that bug. the call to the lstm should take the self.hidden state as an argument, and update the hidden state after each call.