You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Stopping/dropping gradients is quite a common feature for a lot of applications. It would be great to have that in Tapir as well. Currently, ChainRulesCore provides ignore_gradient and non_differentiable as a unified interface, but pretty much only Zygote supports it. However, ReverseDiff provides its own macro: ReverseDiff.@skip.
The text was updated successfully, but these errors were encountered:
This should be straightforward, albeit slightly different from how ChainRules / Zygote do it, because we do a lot of in-place incrementation and propagation of memory locations at which to increment on the forwards-pass of AD.
I think this could be done by adding a rule which
replaces the fdata associated to the thing whose gradient we're dropping with a copy that doesn't live at the same memory address, and
Hi,
Stopping/dropping gradients is quite a common feature for a lot of applications. It would be great to have that in Tapir as well. Currently,
ChainRulesCore
providesignore_gradient
andnon_differentiable
as a unified interface, but pretty much onlyZygote
supports it. However,ReverseDiff
provides its own macro:ReverseDiff.@skip
.The text was updated successfully, but these errors were encountered: