RuntimeError: index out of range at /pytorch/aten/src/TH/generic/THTensorEvenMoreMath.cpp:191

[loretoparisi]

We have this error occasionally. It's hard to reproduce it and get significative logging. A very similar error has been found and fixed here:
chenxijun1029/DeepFM_with_PyTorch#1

This may help to find a possibile solution for LASER too.

[hoschwenk]

Well, without any information when and where the error happens, it is indeed to find a solutio nin LASER. Looking at the link, it seems to happen in the word embedding layer.
You could try to print the tokens before the error is raised.
However, I've used embed.py many time, to literally embed billions of sentences, and I never had this error.

[loretoparisi]

@hoschwenk thank you, yes in fact it is pretty hard to reproduce, we are going to send you more logging info.

[simonefrancia]

That is the error that we randomly faced with and we cannot reproduce it systematically(we don't have an specific input that causes the error):

The expanded size of the tensor (13) must match the existing size (12) at non-singleton dimension 1.  Target sizes: [3, 13].  Tensor sizes: [12]

We also opened an issue (pytorch/pytorch#23652) on Pytorch Repo, because we think that it could be linked with loading LASER model with that. In fact this error occurs in the first inference of LASER embedding.
Thank you