Refactor old

[r0mainK]

Hey !

So first off thanks for the repo, it really helped having a tested CRF in pytorch. I spent the past couple cycles improving your code, and thought I would PR. Here is the list of changes I PR ed here (high level, we can talk of each commit during the review):

• remove useless methods: you had a lot of methods needlessely encapsulating code, I removed them as it made it harder to read. Specifically, I removed reset_parameters because it is like creating a new layer, I removed _compute_score and _compute_normalize to have only one loop, and because separating thr code made it harder to understand, and I removed _viterbi_decode because it was useless
• refactor validation: ValueError and assert were competing for validating inputs, but there was a ton of redundance and ValueError were more informative, so I removed all the asserts. I also added a check on the mask, and moved some code around for efficiency
• optimize decoding: I made the second part of the viterbi decoding in batch, and now return a tensor. this allows for an important speedup (even for small batch sizes). It's better to never convert back to list in code, it should be upt to the user to do this conversion

There is also some minor changes, like adding pytorch to the setup.py. I think the commits are well split so it should be no problem to review. I tested on my computer with pytorch 1.0, 1.1 and 1.4, all work.

PS: I tried to keep your coding style as much as possible, but I would recommend removing the line comments. Your code is good enough so that is impedes comprehension in my opinion.

EDIT: Added optimization of numerator computation

[coveralls]

Coverage remained the same at 100.0% when pulling acae768de081a2739a0b6e7e84b921755c73344b on r0mainK:refactor-old into 1b7ced20c01352ccf70b7ae7bb99a58366de8e48 on kmkurn:master.

[kmkurn]

Hi, thank you for the PR! Really appreciate it. However my hands are a bit full at the moment, and this is not a small PR, so it might take a while before I can review. But from skimming the code:

• I prefer that the version is not bumped (yet) because I'd rather do that when I'm about to release the next version
• Please bring back reset_parameters to keep it aligned with PyTorch's convention where each layer has that method (which is still the case last time I checked)

Thanks again for the PR!

[r0mainK]

No problem, I'm not in a rush :)

I'll remove both commits, indeed I hadn't seen that the reset_parameters was idiomatic, I had checked only the superclass nn.Module but not the actual instances !