Kemal Kurniawan

Thanks for the explanation. I guess there's no easy way to solve this. It's not a problem for me any longer so feel free to close this issue. But if...

Yay, a logo! I think it'd look better without the halo though. Sorry! 🙏

Sorry for responding so late. I like your solution. I've merged #104 so you can implement the solution now.

I'm merging this despite the build error. The error seems caused by [Python 3.6 that no longer exists in newer Ubuntu](https://github.com/actions/setup-python/issues/544).

Hi, sorry for jumping in, but the issue with `NonProjectiveDependencyCRF` is I believe the line below: https://github.com/harvardnlp/pytorch-struct/blob/9f5a0f5b5a41de9eb64b6abf23246677ce5a02ae/torch_struct/deptree.py#L196 It should've used `lap.logdet()` to avoid numerical error for large matrices. Slightly off...

With the actual data I’m using, yes it becomes much larger. I keep getting OOM, and this is at inference time where I use torch.no_grad(). The high memory consumption seems...

@atul47B You can use something like ```python is_pad = tags == -100 tags.masked_fill_(is_pad, 0) loss = -crf(emissions, tags, mask=~is_pad) ``` The `crf` forward computation will ignore positions where mask is...

Closing because the issue is resolved.

I actually have a working code for variable-length matrix-tree for labeled tree: https://gist.github.com/kmkurn/fb34ad169466b050dddac1de2d1f7b93. This also handles overflow bc of exp when the potentials are extremely positive. I did try to...

Great! For now, I hope my code can help those who need it quickly. Another thing, what can be done easily in the meantime is maybe raise an error/warning if...