# Flux.train error: no method matching back!(::Float64)
**URL:** https://discourse.julialang.org/t/flux-train-error-no-method-matching-back-float64/31956
**Category:** Machine Learning
**Tags:** diffeq, flux
**Created:** [December 7, 2019, 2:33am UTC](https://discourse.julialang.org/t/flux-train-error-no-method-matching-back-float64/31956 "2019-12-07T02:33:07Z")
**Posts on this page:** 5
**Page:** 1
### Author:  [@janak](https://discourse.julialang.org/u/janak)
#### Post date: [December 7, 2019, 2:33am UTC](https://discourse.julialang.org/t/flux-train-error-no-method-matching-back-float64/31956/1 "2019-12-07T02:33:08Z")
I am working with Julia 1.2.0 on Jupyter lab on Windows 10 machine.
I am using the following package versions:
DiffEqFlux : v0.8.1
DifferentialEquations: v6.8.0
Flux: v0.9.0
I am trying to train a neural network using a differential equation optimizer using the DiffEqFlux package.
Here is my code:
```julia
using Flux
using DifferentialEquations
using DiffEqFlux
using LinearAlgebra
using JLD2
using Plots
using Flux: @epochs
cr = 32
Nt = 512
T_cs = zeros(Nt, cr)
T₀ = T_cs[:, 1]
ann = Chain(Dense(cr, 2cr, tanh),
Dense(2cr, cr))
p1 = Flux.data(DiffEqFlux.destructure(ann))
ps = Flux.params(ann)
foretold(u,p,t) = DiffEqFlux.restructure(ann,p[1:4192])(u)
prob = ODEProblem(foretold,T₀,tspan_train,p1)
Flux.Tracker.collect(diffeq_adjoint(p1,prob,Tsit5(),u0=T₀,abstol=1e-8,reltol=1e-6))
function predict_adjoint()
diffeq_adjoint(p1,prob,Tsit5(),u0=T₀,saveat=t_train,reltol=1e-6, abstol=1e-8)
end
opt = ADAM(1)
data = [(T₀, T_cs[:, 1:n_train])]
loss_function(T₀, T_data) = sum((predict_adjoint() .- T_data).^2)
cb = function ()
loss = loss_function(T₀, T_cs[:, 1:n_train]) # Not very generalizable...
println("loss = $loss")
end
for _ in 1:100
Flux.train!(loss_function, ps, data, opt, cb = cb)
end
```
I get the following error when i run this:
```julia
MethodError: no method matching back!(::Float64)
Closest candidates are:
back!(::Any, !Matched::Any; once) at C:\Users\daddyj\.julia\packages\Tracker\JhqMQ\src\back.jl:75
back!(!Matched::Tracker.TrackedReal; once) at C:\Users\daddyj\.julia\packages\Tracker\JhqMQ\src\lib\real.jl:14
back!(!Matched::TrackedArray) at C:\Users\daddyj\.julia\packages\Tracker\JhqMQ\src\lib\array.jl:68
Stacktrace:
[1] gradient_(::getfield(Flux.Optimise, Symbol("##15#21")){typeof(loss_function),Tuple{Array{Float64,1},Array{Float64,2}}}, ::Tracker.Params) at C:\Users\daddyj\.julia\packages\Tracker\JhqMQ\src\back.jl:4
[2] #gradient#24(::Bool, ::typeof(Tracker.gradient), ::Function, ::Tracker.Params) at C:\Users\daddyj\.julia\packages\Tracker\JhqMQ\src\back.jl:164
[3] gradient at C:\Users\daddyj\.julia\packages\Tracker\JhqMQ\src\back.jl:164 [inlined]
[4] macro expansion at C:\Users\daddyj\.julia\packages\Flux\dkJUV\src\optimise\train.jl:71 [inlined]
[5] macro expansion at C:\Users\daddyj\.julia\packages\Juno\oLB1d\src\progress.jl:134 [inlined]
[6] #train!#12(::getfield(Main, Symbol("##27#28")), ::typeof(Flux.Optimise.train!), ::Function, ::Tracker.Params, ::Array{Tuple{Array{Float64,1},Array{Float64,2}},1}, ::ADAM) at C:\Users\daddyj\.julia\packages\Flux\dkJUV\src\optimise\train.jl:69
[7] (::getfield(Flux.Optimise, Symbol("#kw##train!")))(::NamedTuple{(:cb,),Tuple{getfield(Main, Symbol("##27#28"))}}, ::typeof(Flux.Optimise.train!), ::Function, ::Tracker.Params, ::Array{Tuple{Array{Float64,1},Array{Float64,2}},1}, ::ADAM) at .\none:0
[8] top-level scope at .\In[91]:2
```
I would appreciate any suggestions as to why I am getting this error.
---
### Author:  [@PolarizedPoutine](https://discourse.julialang.org/u/PolarizedPoutine)
#### Post date: [December 8, 2019, 1:03am UTC](https://discourse.julialang.org/t/flux-train-error-no-method-matching-back-float64/31956/2 "2019-12-08T01:03:14Z")
Hey @janak I had to modify your example a little bit to get it to run (syntax errors and some indexing issues) but ended up running into the same error.
I’m not super familiar with Flux but I noticed that the loss function does not return a tracked scalar so I’m guessing it can’t backpropagate as Flux can only calculate gradients of tracked values. But not exactly sure which array or which value should be tracked…
```julia
using Flux
using DifferentialEquations
using DiffEqFlux
using LinearAlgebra
using JLD2
using Plots
using Flux: @epochs
cr = 32
Nt = 512
T_cs = zeros(Nt, cr)
T₀ = T_cs[1, :]
ann = Chain(Dense(cr, 2cr, tanh),
Dense(2cr, cr))
p1 = Flux.data(DiffEqFlux.destructure(ann))
ps = Flux.params(ann)
n_train = 256
t_train = range(0.0, 1.0; length=n_train)
tspan_train = (0.0,1.0)
foretold(u,p,t) = DiffEqFlux.restructure(ann,p[1:4192])(u)
prob = ODEProblem(foretold,T₀,tspan_train,p1)
Flux.Tracker.collect(diffeq_adjoint(p1,prob,Tsit5(),u0=T₀,abstol=1e-8,reltol=1e-6))
function predict_adjoint()
diffeq_adjoint(p1,prob,Tsit5(),u0=T₀,saveat=t_train,reltol=1e-6, abstol=1e-8)
end
opt = ADAM(1)
data = [(T₀, T_cs[1:n_train, :])]
loss_function(T₀, T_data) = sum((hcat(predict_adjoint().u...)' .- T_data).^2)
cb = function ()
loss = loss_function(T₀, T_cs[1:n_train, :]) # Not very generalizable...
println("loss = $loss")
end
for _ in 1:100
Flux.train!(loss_function, ps, data, opt, cb = cb)
end
```
---
### Author:  [@ManvithaPonnapati](https://discourse.julialang.org/u/ManvithaPonnapati)
#### Post date: [December 8, 2019, 2:00am UTC](https://discourse.julialang.org/t/flux-train-error-no-method-matching-back-float64/31956/3 "2019-12-08T02:00:38Z")
@janak - I took your example and modified it a bit and it should work now without any errors. Not sure if I got the loss function to do what you intended you it to do but here is a working example.
```julia
using DiffEqFlux, Flux, OrdinaryDiffEq
cr = 2
Nt = 512
T_cs = zeros(Nt, cr)
T₀ = T_cs[1, :]
n_train = 251
u0 = param(Float32[0.8; 0.8])
tspan = (0.0f0,25.0f0)
ann = Chain(Dense(cr, 2cr, tanh),
Dense(2cr, cr))
p1 = Flux.data(DiffEqFlux.destructure(ann))
p2 = Float32[-2.0,1.1]
p3 = param([p1;p2])
ps = Flux.params(p3,u0)
function dudt_(du,u,p,t)
x, y = u
du[1] = DiffEqFlux.restructure(ann,p[1:length(p)-2])(u)[1]
du[2] = p[end-1]*y + p[end]*x
end
prob = ODEProblem(dudt_,u0,tspan,p3)
diffeq_adjoint(p3,prob,Tsit5(),u0=u0,abstol=1e-8,reltol=1e-6)
function predict_adjoint()
diffeq_adjoint(p3,prob,Tsit5(),u0=u0,saveat=0.0:0.1:25.0)
end
opt = ADAM(1)
evalcb = () -> @show(loss(X, Y))
loss_adjoint(T,T_data) = sum(hcat(predict_adjoint()' .- T_data).^2)
data = [(T₀, T_cs[1:n_train, :])]
evalcb = () -> @show(loss_adjoint(T₀, T_cs[1:n_train, :]))
for _ in 1:100
Flux.train!(loss_adjoint, ps, data, opt, cb = evalcb)
end
```
---
### Author:  [@CollinsPeriodLee](https://discourse.julialang.org/u/CollinsPeriodLee)
#### Post date: [October 27, 2021, 4:45am UTC](https://discourse.julialang.org/t/flux-train-error-no-method-matching-back-float64/31956/4 "2021-10-27T04:45:17Z")
Any updates for this topic? I have same issue for the example code of DiffEqFlux.jl package.
The code is
```julia
using DifferentialEquations
function lotka_volterra(du,u,p,t)
x, y = u
α, β, δ, γ = p
du[1] = dx = α*x - β*x*y
du[2] = dy = -δ*y + γ*x*y
end
u0 = [1.0,1.0]
tspan = (0.0,10.0)
p = [1.5,1.0,3.0,1.0]
prob = ODEProblem(lotka_volterra,u0,tspan,p)
u0_f(p,t0) = [p[2],p[4]]
tspan_f(p) = (0.0,10*p[4])
p = [1.5,1.0,3.0,1.0]
prob = ODEProblem(lotka_volterra,u0_f,tspan_f,p)
p = [1.5,1.0,3.0,1.0]
prob = ODEProblem(lotka_volterra,u0,tspan,p)
sol = solve(prob,Tsit5(),saveat=0.1)
A = convert(Vector{Float32},sol[1,:]) # length 101 vector
using Flux, DiffEqFlux
p = Float32[2.2, 1.0, 2.0, 0.4] # Initial Parameter Vector
params = Flux.params(p)
function predict_rd() # Our 1-layer "neural network"
solve(prob,Tsit5(),p=p,saveat=0.1)[1,:] # override with new parameters
end
loss_rd() = sum(abs2,x-1 for x in predict_rd()) # loss function
data = Iterators.repeated((), 100)
opt = ADAM(0.1)
cb = function () #callback function to observe training
display(loss_rd())
# using `remake` to re-create our `prob` with current parameters `p`
# display(plot(solve(remake(prob,p=p),Tsit5(),saveat=0.1),ylim=(0,6)))
end
# Display the ODE with the initial parameter values.
cb()
Flux.train!(loss_rd, params, data, opt, cb = cb)
```
The errors are the same:
```julia
ERROR: LoadError: MethodError: no method matching back!(::Float64)
Closest candidates are:
back!(::Any, ::Any; once) at ~/.julia/packages/Tracker/YNNTM/src/back.jl:74
back!(::Tracker.TrackedReal; once) at ~/.julia/packages/Tracker/YNNTM/src/lib/real.jl:13
back!(::TrackedArray) at ~/.julia/packages/Tracker/YNNTM/src/lib/array.jl:68
Stacktrace:
[1] gradient_(f::Flux.Optimise.var"#15#21"{typeof(loss_rd), Tuple{}}, xs::Tracker.Params)
@ Tracker ~/.julia/packages/Tracker/YNNTM/src/back.jl:4
[2] gradient(f::Function, xs::Tracker.Params; nest::Bool)
@ Tracker ~/.julia/packages/Tracker/YNNTM/src/back.jl:164
[3] gradient
@ ~/.julia/packages/Tracker/YNNTM/src/back.jl:164 [inlined]
[4] macro expansion
@ ~/.julia/packages/Flux/qXNjB/src/optimise/train.jl:73 [inlined]
[5] macro expansion
@ ~/.julia/packages/Juno/oLB1d/src/progress.jl:134 [inlined]
[6] train!(loss::Function, ps::Tracker.Params, data::Base.Iterators.Take{Base.Iterators.Repeated{Tuple{}}}, opt::ADAM; cb::var"#3#4")
@ Flux.Optimise ~/.julia/packages/Flux/qXNjB/src/optimise/train.jl:71
```
---
### Author:  [@CollinsPeriodLee](https://discourse.julialang.org/u/CollinsPeriodLee)
#### Post date: [October 27, 2021, 7:41pm UTC](https://discourse.julialang.org/t/flux-train-error-no-method-matching-back-float64/31956/5 "2021-10-27T19:41:14Z")
Solved. Julia installed the tracker.jl that is never used since then. Solve this issue by updating Flux.