Jan Weidner

Thanks, @juliohm for bringing this up! To me the question really is how often do people want unitful plotting. Personally, I use it quite a bit, but it might be...

> Unitful.jl is a super lightweight dependency It does add 0.2s load time, I would not call this super lightweight. > I wonder if it wouldn't be much easier for...

> Alternatively if they are just recipes, they could go into Unitful.jl directly? This is a very good suggestion. I imagine that most Plots.jl users don't use Unitful.jl, but that...

> Did you mean for your PyTorch code to be `y = loss(net(x))`? :smile: you are right!

> Any updates on the memory differences from then? Yes I already updated the post.

@DhairyaLGandhi I added timings, pytorch is 2.5x faster.

@mcabbott very cool, that you are working on this. With your branch I get: ```julia x = CUDA.randn(128,128,128,10) Zygote.gradient(loss, x) #warmup CUDA.@time for _ in 1:100 Zygote.gradient(loss, x) end #...

@mcabbott here is an example more close to my actual use case. Here runtime of pytorch is ~1.6x faster and memory usage is ~4x lower. ## Zygote Zygote implementation ```julia...

> Am I reading correctly that there are no nonlinearities here, you never broadcast relu etc? That step was something which looked like there was room to fuse things &...

I updated my post to include non linearities and multiple Zygote commits.