"nan" error occurs for the valuable "logprobs" in L87 of translator.py

[JunjieHu]

Hi, I am running an experiment for unsupervised Slovak-English translation. I get the "nan" error during beam search decoding. That is, the valuable "logprobs" in L87 of translator.py contains all "nan". https://github.com/artetxem/undreamt/blob/master/undreamt/ translator.py#L87

I follow the same setting in your paper (train bilingual wordings by vecmap and run for 300k iterations). Does anyone know how to fix it? Thanks!

[artetxem]

It looks like the error is on greedy decoding, not beam search. Does this happen at training or test time?

[JunjieHu]

Hi sorry for the late response. I know the error happens during the test time. But I didn't print out the logprobs during the training, I am not sure whether this value becomes "nan" or not during training time (but I guess it happens).

[masonreznov]

Hello @JunjieHu, did you solve the NAN problem?

[JunjieHu]

Hi @masonreznov unfortunately I didn't solve the NAN problem. Maybe @artetxem could suggest a fix to this.

[mlschomaker]

I have the same issue, this happens during training as well for me