RNN gradients are wrong

[eaplatanios]

This line should be 𝛁cell += new𝛁cell instead. There is also an issue with the initial state initialization to a zeros tensor with batch size 1 that does not broadcast. I don't have time to open a PR right now because I'm traveling but will try to open one once I get a chance.

[rxwei]

Nice catch!

[Shashi456]

Should I take care of this @eaplatanios

[eaplatanios]

@Shashi456 thanks! Feel free to fix the += bug. The initial state issue is a less trivial API design question that I haven't thought about much yet. There are a couple of possible solutions that I used in swift-rl, so I'll try to look into it sometime soon.

[sgugger]

I think this has been solved by the two PRs merged. Feel free to reopen if there is still a problem.

[Shashi456]

@sgugger the tests against these changes were remaining so I thought we'd keep this PR open. #555, #554 track the tests. But since we have separate issues for them maybe it's okay.

[sgugger]

Like you said there are already two issues tracking this. I wanted to do a bit of clean-up, but if anyone feels strongly about this issue remaining open, they can click the button :)