RuntimeError: Expected object of scalar type Long but got scalar type Float for argument #3 'index' in call to _th_index_select about codexglue

[marnaoty]

I think this bug is due to changes in recent pytorch versions.

I think the buggy line is:
prevK = bestScoresId / numWords
inside the beam advance function.

I think the original implementation was seeking an integer division here in this line, but pytorch no longer consider this as an integer division. Probably you can solve this by using // instead.

I hope we can get an answer from the original contributers.

[meikerwang]

@marnaoty Thank you for your reply and these details~
I modify the buggy line by using integer division "//" to replace float division "/" and run the code successfully
Also can use floor_divide instead of true_divide

I think this bug is due to changes in recent pytorch versions.

I think the buggy line is:
prevK = bestScoresId / numWords
inside the beam advance function.

I think the original implementation was seeking an integer division here in this line, but pytorch no longer consider this as an integer division. Probably you can solve this by using // instead.

I hope we can get an answer from the original contributers.