MethodError: no method matching zero(::Tuple{Float64, Float64})

[troyrock]

I have the following code which is very similar to Chris Rackauckas' code from a tutorial where I have changed from Float32 to Float64. The function loss4() works with train!() but functions loss() and loss3() do not work with train!() (they work stand-alone). It's baffling to me. I have tried to do an @enter trace but I quickly got lost with all of the callbacks.

using Flux
using Statistics
# using DifferentialEquations

NNode = Chain(x -> [x],
              Dense(1, 32, tanh),
              Dense(32, 1),
              first)
Flux.f64(NNode)
g(t) = t*NNode(t) + 3.0
ϵ = sqrt(eps(Float64))

loss() = mean(abs2(((g(t+ϵ)-g(t))/ϵ) - cos(2π*t)) for t in 0:(rand(Float64)/50):1.0)

loss4 = function()
  t = rand(Float64)
  a1 = abs2(((g(t+ϵ)-g(t))/ϵ)-cos(2π*t))

  for _ in 1:10
      t = rand(Float64)
      a1 += abs2(((g(t+ϵ)-g(t))/ϵ)-cos(2π*t))
  end
  return(a1/11)
end

loss3 = function()
  sum = 0
  count = 0
  for t in 0:rand(Float64)/50:1.0  # changing this to 0:0.1:1.0 doesn't work either.
      sum += abs2(((g(t+ϵ)-g(t))/ϵ)-cos(2π*t))
      count += 1
  end
  return(sum/count)  
end

Flux.train!(loss3, Flux.params(NNode), Iterators.repeated((), 5000), Flux.Descent(0.01))

[mcabbott]

Same as #1247 I think, runs if you add https://github.com/FluxML/Zygote.jl/issues/1247#issuecomment-1175332642