Sasha Rush

Yup, seems like a bug. It seems like it is not finding any valid solution at all when you set length < t. Is there anyway you could avoid using...

Hi Tim, Sorry for the slow reply. I started looking at the issue, but I am still not happy with the speed of this particular function and am worried it...

I'm so sorry, things have been a bit crazy. I will try to get to this this week.

Hi, I finally got around to making you a full example: https://github.com/harvardnlp/pytorch-struct/blob/master/notebooks/CTC_with_padding.ipynb  Let me know if it is clear. I will eventually get his checked in under the lengths...

Awesome. I am going to leave it open so I remember to build it in.

So just to unify terminology The way we compute linear-chain CRF in O(log N) time is by viewing it as a single balanced binary tree of height `log N`. Each...

You can do O(log N) [parallel] backtracking. The trick in the code is that we never implement the backward / backpointer step. We really on the fact that in pytorch...

Neat! Yeah this should result in the same outcome. My assertion though is that you do not need to write the backward pass manually. It might lead to some speed-ups...

Btw, your repo looks really nice. If you wanted to build some more transformer-backed, structured prediction models, I would be happy to collaborate. Would be nice to have a single...

Neat! Me too. General hypergraphs would be really interesting. But I think we would need to write that in CUDA or TVM manually. The one thing I care a lot...