Potential bug in weighting trigger candidate losses

[davides]

First, thanks for publishing your implementation of this technique. It's been very helpful!

While stepping through the code, I think I may have found a small issue. The goal of this code seems to be: re-weight the loss contributed by each batch by the fraction of unmasked tokens it contains.

If that's the case, shouldn't curr_num_elements count all elements != constants.PAD_TOKEN_ID (-100), instead of -1?

Thanks!