Any tips for for updating the Flux example/test for Torch.jl for NNlib >= v0.8?
Currently affected by an issue which leads to trouble with differentiating a foreigncall expression: https://github.com/stemann/Torch.jl/blob/feature/torch_1.10.2/test/flux_tests.jl#L23
Call stack:
Torch.Wrapper.at_dim
: https://buildkite.com/julialang/torch-dot-jl/builds/269#0193115b-342e-4b34-bf9c-48cbf9f37eef/339-544gradient
call for a Metalhead.ResNet(18)
: https://buildkite.com/julialang/torch-dot-jl/builds/269#0193115b-342e-4b34-bf9c-48cbf9f37eef/339-591I'm guessing some functions need to be marked as @non_differentiable
Related question: Is the current approach with ZygoteRules.@adjoint
up to date or should they be replaced with ChainRulesCore.rrule
s (at some point)?
IMO little is gained by replacing @adjoint
with rrule
now, so I wouldn't bother.
The only reason @adjoint
still exists is that nobody has had the bandwidth to try deprecating it properly.
Last updated: Nov 22 2024 at 04:41 UTC