Felix Wechsler
Felix Wechsler
But `cis`has good performance? And the linked issue was only for CUDA
Uh, I believe the problem is more general than posted above. Several basic functions are broken: ```julia julia> f_exp(x) = sum(real(exp.(x))) f_exp (generic function with 1 method) julia> Zygote.gradient(f_exp, x_c)...
Without CUDA it seems to work fine, I haven't encountered such an issue yet. Above the examples are with and without CUDA. The only one that fails, is Zygote +...
Hm, I tried to do some debugging. `T = Broadcast.combine_eltypes(f, args)` is always `Float32` independent whether we have a `CuArray{ Zygote.dual_function(abs2).(CuArray(randn(ComplexF32, (2,2)))) 2×2 CuArray{Float32, 2}: 2.71749 0.407771 0.735002 0.240149 ```...
Hm, I'm confused why the code distinguishes between complex and real numbers. [DualNumbers.jl doesn't](https://github.com/JuliaDiff/DualNumbers.jl/blob/master/src/dual.jl) Edit: OK reading some discussions, that's not exactly the same as in ForwardDiff. I believe that's...
That's still an issue, isn't it? I might want to use the creation of the CAs during the forward pass of the function and hence it needs to be rather...
Hi! I stumbled over the same issue today. Is there any progress? Best, Felix
Thanks a lot! But no hurry, I just was interested in the progress :laughing: Because of #48 it might be also a bad idea to put the creation in the...
Is that affected by #86?
Thanks for your code, I think I can follow it conceptually. Without the @generated I also can inspect the generated code, it looks exactly like what I want to have....