Carlo Lucibello
Carlo Lucibello
It is impossible to have efficient iteration by index order with the current implementation of `Network`. Probably this is issue is a `won't fix`
CI algorithm implemented in #23 CoreHD still to implement
it would be great to be able to cross-reference equations in a notebook, I hope this issue can be solved
A workaround was provided in this discussion https://discourse.julialang.org/t/pluto-how-to-number-equations-in-latex-mode/59900/9
> We do need to return the nothing there, so the "fix" would be to let it fall back to the nt_nothing function and accumulate there. can you post a...
This problem is still present ```julia julia> d = Dict("x"=>rand(2)) Dict{String, Vector{Float64}} with 1 entry: "x" => [0.626974, 0.519716] julia> gradient(x -> sum(x["x"]), d) #OK (Dict{Any, Any}("x" => 2-element Fill{Float64}:...
I think we should be correct and avoid silent bugs before aiming at being fast. Maybe taking a fast path after a "has duplicates" check can help performance a bit?
Since Zygote doesn't depend on CUDA and NNlib I'm not sure where we should put this. Here but inside a `@require`? In `NNlib/NNlibCUDA`?
Can we have some benchmarks here?