Sasha Rush

hi, sorry for the long delay here. I'm going to try to add some tests to make sure it is returning the right values. I don't have a great sense...

that sounds right. I will turn off chart by default. Also now the backward on marginals approach works with fastlogsemiring.

Hmm, this is worrisome, we are pretty careful about precision but it is possible we are losing some bits somewhere. I will try to replicate. This is on CUDA right?...

That is strange. If possible, do you think you could send me a minimal example, perhaps as a colab? Is it plausible that your gold tree is invalid somehow (either...

I see, so I think there are two things happening here : 1) minor: the marginal is having a slight underflow, 2) major: the score is very off because of...

I wouldn't recommend switching between the two. NonProj will be faster and more general, but less accurate in cases where the languages 99% projective. Unfortunately there is currently no easy...

No ideas yet, but I will look into both issues.

@kmkurn ah fantastic, I changed that line and made another issue about variable length.

@wangxinyu0922 thanks for the repro, I will see what I can do to fix.

@wangxinyu0922 I think your code had a bug in it.  Your mask is zeroing out the potentials for the 0th head word.