Sasha Rush

Nice, that is similar in spirit to this code which we have been working on https://github.com/harvardnlp/pytorch-struct/pull/81 . We can integrate them both in to the library. There might also be...

Very neat. So I think that instead of first computing marginals, we can apply this approach in the backward operation of the semiring itself. This is how I compute unbiased...

Awesome sounds like we have three different methods. The one in my PR is from Yao's NeurIPS work https://arxiv.org/abs/2011.14244 which is unbiased forward and biased backward. Maybe we should have...

Is there a reason you closed this? It looks cool.

Very neat. I'll take a read. Do you find that this gives a speedup? Seems hard to parallelize.

Yes, A Neural Attention Model for Abstractive Sentence Summarization , Rush et al http://aclweb.org/anthology/D15-1044 On Sat, Sep 3, 2016 at 9:46 AM, chenwangliangguo [email protected] wrote: > Is there any paper...

This means that the cuda extension packaged with the library failed to install. Was their an error message when you installed?

What a great question... If there are two specific spans that you need, then your "dirty trick" is the right way to do it. If you want to do it...

I think this is a nice reference for bayes nets https://dl.acm.org/doi/pdf/10.1145/765568.765570  Alternatively you can think of CRF as exponential families and therefore the log-partition generates moments: https://www.cs.cmu.edu/~epxing/Class/10708-14/scribe_notes/scribe_note_lecture6.pdf I can't...

Yes, would love a PR. I think you can do this in the StructDistribution class https://github.com/harvardnlp/pytorch-struct/blob/master/torch_struct/distributions.py. Should in theory work for all the distributions in the library. The main thing...