# Why does this Flux code fail with Float64, but run for Float32?

**URL:** <https://discourse.julialang.org/t/why-does-this-flux-code-fail-with-float64-but-run-for-float32/71077>\
**Category:** New to Julia\
**Tags:** flux\
**Created:** [November 7, 2021, 3:41am UTC](https://discourse.julialang.org/t/why-does-this-flux-code-fail-with-float64-but-run-for-float32/71077 "2021-11-07T03:41:04Z")\
**Posts on this page:** 1\
**Page:** 1

<div class="post-metadata">

**Author:** ![Ruvi\_Lecamwasam](https://sea2.discourse-cdn.com/julialang/user_avatar/discourse.julialang.org/ruvi_lecamwasam/32/28374_2.png) [@Ruvi\_Lecamwasam](https://discourse.julialang.org/u/Ruvi_Lecamwasam)\
**Post date:** [November 7, 2021, 3:41am UTC](https://discourse.julialang.org/t/why-does-this-flux-code-fail-with-float64-but-run-for-float32/71077/1 "2021-11-07T03:41:04Z")

</div>

I’m following along with [this tutorial by Chris Rackauckas](https://mitmath.github.io/18337/lecture3/sciml.html). This is the code provided for a simple neural network:

```julia
using Flux
NNODE = Chain(x -> [x], # Take in a scalar and transform it into an array
           Dense(1,32,tanh),
           Dense(32,1),
           first) # Take first value, i.e. return a scalar
NNODE(1.0)

g(t) = t*NNODE(t) + 1f0

using Statistics
ϵ = sqrt(eps(Float32))
loss() = mean(abs2(((g(t+ϵ)-g(t))/ϵ) - cos(2π*t)) for t in 0:1f-2:1f0)

opt = Flux.Descent(0.01)
data = Iterators.repeated((), 1000)
iter = 0
cb = function () #callback function to observe training
  global iter += 1
  if iter % 500 == 0
    display(loss())
  end
end
display(loss())
Flux.train!(loss, Flux.params(NNODE), data, opt; cb=cb)

```

This code runs just fine. I noticed however that the numbers are defined as Float32 rather than Float64, expressions such as _1f0_ or _eps(Float32)_. If I change all of these to Float64, by replacing all of my _1f0_ with _1e0_, I get this code, which no longer runs:

```julia
using Flux
NNODE = Chain(x -> [x], # Take in a scalar and transform it into an array
           Dense(1,32,tanh),
           Dense(32,1),
           first) # Take first value, i.e. return a scalar
NNODE(1.0)

g(t) = t*NNODE(t) + 1e0

using Statistics
ϵ = sqrt(eps(Float64))
loss() = mean(abs2(((g(t+ϵ)-g(t))/ϵ) - cos(2π*t)) for t in 0:1e-2:1e0)

opt = Flux.Descent(0.01)
data = Iterators.repeated((), 1000)
iter = 0
cb = function () #callback function to observe training
  global iter += 1
  if iter % 500 == 0
    display(loss())
  end
end
display(loss())
Flux.train!(loss, Flux.params(NNODE), data, opt; cb=cb)

```

> 0.6015101698151941  
> MethodError: no method matching zero(::Tuple{Float64, Float64})  
> Closest candidates are:  
> zero(::Union{Type{P}, P}) where P\<:Dates.Period at C:\buildbot\worker\package\_win64\build\usr\share\julia\stdlib\v1.6\Dates\src\periods.jl:53  
> zero(::StatsBase.Histogram{T, N, E}) where {T, N, E} at C:\Users\ruvil.julia\packages\StatsBase\Q76Ni\src\hist.jl:538  
> zero(::SparseArrays.AbstractSparseArray) at C:\buildbot\worker\package\_win64\build\usr\share\julia\stdlib\v1.6\SparseArrays\src\SparseArrays.jl:55  
> …

I have read that with Flux it is recommend to use Float32, because the extra precision isn’t needed and you halve the memory usage. However I still would like to understand what the source of this error is.