Removing NaNs not working in 0.3.0 anymore

1

Hi guys,

In version 0.2.0_post3 I used to do the following operation to remove NaNs that were being produced by a division by zero:

out[out != out] = 0

and it used to work!

With version 0.3.0_post4 I’m getting the following error message:

assert not ctx.needs_input_grad[1], "MaskedFill can't differentiate the mask"
AssertionError: MaskedFill can't differentiate the mask

Any tips / workarounds?

Cheers,
Miguel

2

Hey, the following should work.
The problem is that the result of out != out has requires_grad=True while the indexing operation cannot differentiate wrt the mask.

out[(out != out).detach()] = 0
1 Like
3

Will this create any sort of problem in the remaining part of the model?

4

It won’t as the original out is not modified, just the temporary Variable which is the result of out != out.

1 Like