"fill_value" missed from current API

First congratulations @rusty1s for merging torch-scatter into pytorch master. I noticed that the fill_value param is missed from the current API, so what’s their default values now for different reduction method (max, min, mean, add)?

I ask because the previous default fill_value for max and min reduction methods are not appropriate (numpy.finfo(numpy.float32).max and numpy.finfo(numpy.float32).min) for me. I’d prefer them set to zero in my case.

[gasteigerjo]

Okay, I’ll just write a wrapper for that and then all will be good again in my code world. 😉

Thank you!

[gasteigerjo]

I went with

mask = ptr[1:] == ptr[:-1]
out[mask] = fill_value

Which works, but it’s slower than segment_coo. So I’ll just stick with that.