Is there an API-based workaround to avoid recreating the tape when extra arguments change their value? about reversediff.jl HOT 3 CLOSED

P.S. What I don't understand very well yet, after reading the documentation, is whether GradientConfig can help make my above example more efficient in some way.

from reversediff.jl.

If I understand things well, it seems that it is impossible to use a tape indeed in this example, simply because the tape needs to change at each re-evaluation (iteration in my own loop), but still, I can at least exploit GradientConfig, as follows:

rcfg = ReverseDiff.GradientConfig(x);

z.a = 4.9;

ForwardDiff.gradient!(fresult, f, x);

ReverseDiff.gradient!(rresult, f, x, rcfg);

DiffBase.value(fresult)
DiffBase.value(rresult)

DiffBase.gradient(fresult)
DiffBase.gradient(rresult)

Although a tape wouldn't be of use in this situation, using GradientConfig avoids some re-allocation at least. If this is correct, maybe we can close the issue. Thanks for the support and advice.

from reversediff.jl.

The naive answer is no at the moment; I believe this issue is a dupe of #36. For now, you either have to re-record every time (which you can do with a pre-allocated GradientConfig to save a little bit of memory) or include the parameters you don't actually care about in the differentiation.

Note that in your original example, you're using a custom type; it's not likely that we'll ever be able to support pre-recording with non-array/non-scalar types unless Julia allows getfield overloading.

Closing as a dupe of #36.

from reversediff.jl.