Closed gdalle closed 9 months ago
I recall this came up before in a different issue/PR thread, and the comment then was that this returns a tuple of "value" and "pullback function" rather than a "value and pullback"-returning function. Certainly the former seems to make more sense when we're talking about split reverse mode AD?
Regardless of what the best interpretation would be (I personally find this ambiguity horrifying), I think it would be good to fix the docstring first instead of changing the behavior. If you wanna review the associated PR #125 I can merge it afterwards :)
You'll find no complaints from me there, not sure why it wasn't called e.g. value_and_pullback
in the first place!
It returns
val, (tangents -> pullbacks)
instead oftangents -> (val, pullbacks)
For comparison, the forward equivalent does
tangents -> (val, pushforwards)
which is what it promises