Michael Abbott
Michael Abbott
If you print the status of the example in the docs at the readme, this makes it about 4 screenfuls, instead of 20something. But shows a few numbers so that...
Example: ```julia using Optimisers, Functors, Zygote struct TwoThirds a; b; c; end # from the tests Functors.@functor TwoThirds (a, c) Optimisers.trainable(x::TwoThirds) = (a = x.a,) mtt = TwoThirds([1.0, 2.0], [3.0],...
This tries to fix #62 by using `functor(typeof(x), dx)` only for gradients, and `functor(aux)` for the tree of offsets made by `destructure`. But it doesn't quite work, as for an...
This adds tests which should detect problems with CuArrays, which can run on your laptop (and on CI) using a fake GPUArray. Similar to https://github.com/FluxML/OneHotArrays.jl/pull/9 ~~WIP as it should still...
This is simpler than the version in https://github.com/FluxML/Flux.jl/pull/969, as it has no special handling for momentum, or not yet. It's unusual in that I think it needs to be written...
Notice that Tracker does not behave like Zygote here: ```julia julia> arr = [3.0, 4.0]; julia> Tracker.withgradient(nt -> sum(sum(x) for x in nt), (a=arr, b=arr, c=copy(arr))) (val = 21.0 (tracked),...
Moving y to "gpu" inside loss causes its gradient to be lost: ```julia julia> using Tracker, JLArrays julia> JLArrays.allowscalar(false) julia> Tracker.withgradient((x,y) -> sum(x[1:2] + jl(y))^2, jl([1,2,3.0]), [4,5.0]) (val = 144.0,...
Closes #880.
This PR repairs a small slowdown due to https://github.com/JuliaLang/julia/pull/37898. When all matrices are the same size, that wastes a few ns thinking about their sizes, while this PR restores the...
It's surprising that `splitobs` and `DataLoader` make views, when they mention only `getobs` in their docstrings, which does not: ```julia help?> splitobs splitobs(data; at, shuffle=false) -> Tuple Split the data...