off by one error in seq2seq seminar?

[g0djan]

Hello, I've been working on seq2seq and spotted that here
https://github.com/yandexdataschool/Practical_RL/blob/master/week07_seq2seq/basic_model_torch.py#L36
We're getting a hidden state from the first token after eos. Because 2 lines earlier infer_length is called with include_eos=True so the returned length points to the token after eos.

If it's intentional can you please explain what's the point of getting hidden state that includes information about first token that isn't part of the sequence? Otherwise I believe 1 should be just subtracted from every element of end_index not only from the elements that go out of range.

[g0djan]

Though I didn't see much change for easy mode yet, maybe less wobbly in the end
The run without changes

The run where I set include_eos=False so the returned index points to eos and not the first token after that