hadasah/btm

btm's Issues

`expert_probs` in btm/ensemble_eval_lm.py: 104

Hi,

Thanks for open sourcing this great work! I have some questions on how to calculate the posterior probability for experts. From this line, it seems that the expert probabilities are calculated inside the sequence_scorer.py. I didn't find any expert_probs in this file but fount it in another repo. It seems that the expert probs are the probabilities of the last position in the sequence weights[:, :, -1], why is that the case? I imagine that the last position's probability does not have any meanings. It would be very helpful if you could clarify how to calculate the posterior probability in more detail :) Thanks!

Unable to find the contents of the fairseq directory

Hello.

I am very interested in the contributions your paper has shown.
However, I cannot try to reproduce the experiment because the fairseq directory is empty.
Could you please add some information about the fairseq directory?