Closed Pierre-Bartet closed 1 month ago
Thanks. Will be fixed in https://github.com/rusty1s/pytorch_scatter/pull/445.
Thanks for the quick fix (and the whole library)!
Looking at the code I see that there is a return out.nan_to_num_(neginf=0.0)
, but - inf is a perfectly legit output value.
Is there something preventing the correct handling of these cases https://github.com/rusty1s/pytorch_scatter/pull/426 and https://github.com/rusty1s/pytorch_scatter/issues/407 ?
The following code incorrectly set an unused output element to 0 instead of letting it to its original value (here -10):
which mean
scatter_logsumexp
only works in the corner case where all outputs are affected by the scatter operation.