Runtime error on trainng about muse HOT 4 CLOSED

Okay it looks like the built dictionary is empty.
This is possible if you have dico_build set to S2T&T2S, it means that there is no reciprocity in the nearest neighbors at all (if x is translation of y then y is not translation of x). This happens when the alignment is very bad, and that the model failed to make a good initial alignment. The reason for that can be that:

• you did not train the model long enough and it did not have time to align the emebddings
• the embeddings are too hard to align and the model just can't find a way to align it

You can avoid this issue by defining dico_build to S2T. This should fix the bug, but the model still might not converge.

from muse.

What are the sizes of the source and target vocabularies?

from muse.

source-en: 243003,
target-es: 872827.
max_vocab = 200000

from muse.

Thanks for the help. @glample

from muse.