cossio

Is there a way for Tusk to support custom proxy settings?

In the latest release ablines throws an error in the presence of axis transforms, though in principle this should work. ``` value = ArgumentError: ABLines is only defined for the...

`one'(x)` errors, although `gradient(one, 0.0) === nothing` does not

1

```julia using Zygote julia> one'(0.0) ERROR: MethodError: no method matching getindex(::Nothing, ::Int64) Stacktrace: [1] (::Zygote.var"#59#60"{typeof(one)})(x::Float64) @ Zygote ~/.julia/packages/Zygote/FPUm3/src/compiler/interface.jl:83 [2] top-level scope @ REPL[5]:1 ```

Wrong Fill'ed gradient

1

```julia julia> Zygote.gradient(x -> sum(x .* randn.()), randn(5)) (Fill(-0.30782754826253156, 5),) ``` This is wrong. The result should be the random vector of coefficients.

Mutating arrays error from `hvcat` with `UniformScaling`

3

```julia using Zygote, LinearAlgebra gradient((A,B) -> sum([I A; B I]), randn(2,3), randn(3,2)) ``` gives an error.  Expanding `I` seems to involve calling `copyto!`, which is forbidden. However since no...

When following the Blue style guide on functions with lots of arguments, you are adviced to place the closing parenthesis in it's own line, unindented: ```julia function my_large_function( argument1, argument2,...

Conv with circular padding

6

PyTorch convolution layers admit different modes of padding (https://pytorch.org/docs/stable/generated/torch.nn.Conv1d.html#torch.nn.Conv1d): circular (i.e. periodic padding), reflect, replicate and zeros. From what I understand it seems that Flux (and `NNlib.conv`) only support zero...

DataLoader type instability

2

Here is an example: ``` using Flux, MLDatasets train_x, train_y = MNIST.traindata() train_loader = Flux.Data.DataLoader(train_x_reduced, batchsize=16, shuffle=true, partial=false) eltype(train_loader) # Any ``` Since the `eltype` is `Any`, iterating over the...

This method: https://github.com/FluxML/Flux.jl/blob/d0a5b776308a33085f876f09a40781b0528d716c/src/optimise/train.jl#L9 is never called. It's also not tested. Only the three-arg method is used. Do we need this two-arg method of this function here?

Use Tuple instead of Vector{Any} in the definition of Optimiser

4

Why is `Optimiser` defined like this: https://github.com/FluxML/Flux.jl/blob/7f375aa46fbdece7ae501f920c12003e9fdf2682/src/optimise/optimisers.jl#L553 This seems to gratuitously violate some of the receommendations in the manual: https://docs.julialang.org/en/v1/manual/performance-tips/#Avoid-fields-with-abstract-type https://docs.julialang.org/en/v1/manual/performance-tips/#man-performance-abstract-container Why not use a `Tuple` here instead? Something like...