David Chiang

Maybe just call torch.expand on the arguments beforehand?

Or is it possible that it just works already, but the dimension checks are too strict?

Made this no longer based on #5.

Oh I see. But still, shouldn't it depend on whether the current output word is EOS?

`beam_scores[i][j]` contains the score resulting from extending hypothesis `i` with word `j`. If `j` is EOS, then the length of the hypothesis is `time_step`, so that's the lenght that should...

OK, that makes sense. I think there's a slight disadvantage this way -- imagine that the length penalty is very large (i.e., strongly prefers short sentences), and there are two...

I mean that although "the cat sat on the mat today" is still an incomplete hypothesis, we know it must have at least 8 words including EOS. With a very...

> But then what do we do about the first step and the second step? Both of them have length 1? The first step and second step shouldn't be any...

This appears to work: ``` %% g(g) g(g[gv:\hfill]) (h+) ```

This is because `\baselineskip` tries to make the space between baselines equal, and the baseline of a staff is the baseline of the lowest text. If there is no translation,...