Alexander Sikorski

I wonder what the current progress on this is, as I need to plot information related to an image (from Images.jl) on top of that image. Is this already possible...

@timholy That would indeed be a possibility, though I couldn't view the result in IJulia anymore. Can I export the annotated picture back to an Image to finally write it...

This realy works when plotting to SVG like you wrote, but I get the following error when trying that from IJulia: ``` Embedding bitmaps in Cairo backends (i.e. PNG, PDF,...

I tried writing the cairo backend part, but didn't get far. Unfortunately [`cairo_image_surface_create_from_png`](http://cairographics.org/manual/cairo-PNG-Support.html#cairo-image-surface-create-from-png) just takes a path to a png file, which seems to restrictive. I then found two methods...

Assuming it was possible to embed the bitmaps into a Compose context, how would Compose/Gadfly decide which layer (i.e. the bitmap or some other plot) should be on top over...

I'm stuck with the same problem, how did you solve yours? https://discourse.julialang.org/t/flux-higher-order-derivatives-and-forward-mode/38805/3 seems to have a similar problem but the workaround there (extracting the NN paramteres and optimzing externally) doesn't...

So what is the actual problem preventing Zygote from computing higher order derivatives?

I actually thought the problem lied elsewhere, but using `sum` instead of `getindex` seems to work here. ```julia x = rand(2) m = Chain(Dense(2,1)) Flux.gradient(params(m)) do gradient(m,x) |> sum |>...

A cleaner way is to extract the gradient by tuple destructuring (is it called that?) `dx, = gradient(m, x)`

So out of #77 we would just need the `@adjoint ∇getindex` part to circumvent the `setindex` call is that correct?