Vincent Nguyen

@francoishernandez are you ok with the first solution ? it looks ok to me.

well so far nobody screamed for the reported scores .... also I have a suggestion: let's take length_penalty average and add alpha in the "avg" mode using cur_len ** alpha....

So for now, my suggestion is: default length_penalty=avg default alpha=1 change the lp with **alpha we remove the normalization in the _report_score() function and later on we can always add...

Hi Ben, welcome back. Up to now, we tried to make the code python2 compatible (which is a requirement in Travis as you can see). I do understand it is...

okay. It seems that it used to be implemented and then removed by this commit: https://github.com/OpenNMT/OpenNMT-py/commit/3052e3890fdb0249a1f37e7b40f8e36a1dabf3a4#diff-bb5f81e1cd6e5e13578730e07d26f151L182 Anyway, it needs to be reimplemented and if possible with the same output as...

@bpopeters di you have time to work on this ?

kind reminder if you have time.

@ziorufus do you have time to test #1240 ?

Before we dive into torch.cuda.amp, should we expect a behavior change versus this issue https://github.com/NVIDIA/apex/issues/475 ? thanks.

Are you using a length penalty ? I think this is expected with the new exit condition but the actual translations (text) is it different ?