Is that true the source and target embedding space should have the same dimensionality?

[yuchenlin]

Hi Mikel,

I found that you didn't mention the case that the source embeddings and target embeddings can have different dimensionality (for example, english 300d, and Chinese 100d). Did I miss something in the paper or the code? Or you only focus on the situation that the two embedding spaces must have the same dimensionality?

Thanks and regards, Bill

[artetxem]

You are right, we currently do not support both spaces having different dimensionality. It should not be difficult to implement it, but I do not think that it would be particularly useful either. I will leave the issue open for now.

[shizhediao]

what if the source embedding is 200000300 while the target embedding is 199999300, could it be trained correctly by this code? thanks!

[shizhediao]

@artetxem

[artetxem]

@shizhediao that should work ok.