use Flux.zeros32

[DhairyaLGandhi]

Fixes

julia> FastDense(1,12).initial_params()
13-element Vector{Union{Nothing, Float32}}:
  0.6627236f0
  0.676602f0
 -0.59016824f0
 -0.53156716f0
 -0.58506316f0
  0.00657402f0
  ⋮
 -0.58002657f0
 -0.092621975f0
  0.30483305f0
  0.30343133f0
   nothing

This still needs a fix in Flux.

[ChrisRackauckas]

@DhairyaLGandhi I tracked it down to:

using DiffEqFlux, Flux, OrdinaryDiffEq, DiffEqSensitivity
n = 1  # number of ODEs
x = rand(n, 5)
y = rand(n, 5)

tspan = (0.0, 1.0)
NN = Dense(n, n, tanh)
nODE = NeuralODE(NN, tspan, ROCK4(), reltol=1e-4, saveat=[tspan[end]], sensealg=InterpolatingAdjoint())
function loss_function(θ)
    z = nODE(x, θ)
    Flux.mse(y, z)
end
size(nODE(x, nODE.p)[:,1]) == size(y)
yy,back = Zygote.pullback(loss_function,nODE.p)
back(1)

I thought I fixed it with:

SciML/RecursiveArrayTools.jl#164

For some reason, the new getindex fallbacks were causing things to fail. But now, this overload catches it but has an issue. What was the old dispatch it was hitting?

[DhairyaLGandhi]

https://github.com/FluxML/Zygote.jl/blob/2d5edf44ad7191fcd5d5816d8ab161dff8e14765/src/tools/builtins.jl#L15

Is this the one you're thinking of?

[ChrisRackauckas]

No, the getindex methods on vector must've recently changed.

[DhairyaLGandhi]

Should we be pulling out the last index here? Seems like something that shouldn't need manual handling. ref FluxML/Flux.jl#1636