使用DiffEqFlux的sciml_train训练NeuralODE时触发MethodError报错

安全阿岳

2026-9-1

问题背景

运行DiffEqFlux官方神经ODE示例代码时抛出方法匹配错误，此前使用近似代码可正常运行，当前运行环境为Julia v1.7.1，所用代码取自DiffEqFlux v1.13.0官方教程。

复现代码

using DiffEqFlux, OrdinaryDiffEq, Flux, Optim, Plots


u0 = Float32[2.0; 0.0]
datasize = 30
tspan = (0.0f0, 1.5f0)
tsteps = range(tspan[1], tspan[2], length = datasize)

function trueODEfunc(du, u, p, t)
    true_A = [-0.1 2.0; -2.0 -0.1]
    du .= ((u.^3)'true_A)'
end

prob_trueode = ODEProblem(trueODEfunc, u0, tspan)
ode_data = Array(solve(prob_trueode, Tsit5(), saveat = tsteps))

dudt2 = FastChain((x, p) -> x.^3,
                  FastDense(2, 50, tanh),
                  FastDense(50, 2))

prob_neuralode = NeuralODE(dudt2, tspan, Tsit5(), saveat = tsteps)

function predict_neuralode(p)
  Array(prob_neuralode(u0, p))
end

function loss_neuralode(p)
    pred = predict_neuralode(p)
    loss = sum(abs2, ode_data .- pred)
    return loss, pred
end

# 训练过程回调函数
list_plots = []
iter = 0
callback = function ( l, pred; doplot = false)
  global list_plots, iter

  if iter == 0
    list_plots = []
  end
  iter += 1

  display(l)

  # 绘制当前预测值与真实数据对比图
  plt = scatter(tsteps, ode_data[1,:], label = "data")
  scatter!(plt, tsteps, pred[1,:], label = "prediction")
  push!(list_plots, plt)
  if doplot
    display(plot(plt))
  end

  return false
end

result_neuralode = DiffEqFlux.sciml_train(loss_neuralode, prob_neuralode.p,
                                          ADAM(0.05), cb = callback,
                                          maxiters = 300)

报错信息

MethodError: no method matching (OptimizationFunction{false, GalacticOptim.AutoZygote, OptimizationFunction{true, GalacticOptim.AutoZygote, DiffEqFlux.var"#84#89"{typeof(loss_neuralode)}, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing}, GalacticOptim.var"#268#278"{GalacticOptim.var"#267#277"{OptimizationFunction{true, GalacticOptim.AutoZygote, DiffEqFlux.var"#84#89"{typeof(loss_neuralode)}, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing}, Nothing}}, GalacticOptim.var"#271#281"{GalacticOptim.var"#267#277"{OptimizationFunction{true, GalacticOptim.AutoZygote, DiffEqFlux.var"#84#89"{typeof(loss_neuralode)}, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing}, Nothing}}, GalacticOptim.var"#276#286", Nothing, Nothing, Nothing})(::OptimizationFunction{true, GalacticOptim.AutoZygote, DiffEqFlux.var"#84#89"{typeof(loss_neuralode)}, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing}, ::GalacticOptim.AutoZygote, ::GalacticOptim.var"#268#278"{GalacticOptim.var"#267#277"{OptimizationFunction{true, GalacticOptim.AutoZygote, DiffEqFlux.var"#84#89"{typeof(loss_neuralode)}, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing}, Nothing}}, ::GalacticOptim.var"#271#281"{GalacticOptim.var"#267#277"{OptimizationFunction{true, GalacticOptim.AutoZygote, DiffEqFlux.var"#84#89"{typeof(loss_neuralode)}, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing, Nothing}, Nothing}}, ::GalacticOptim.var"#276#286", ::Nothing, ::Nothing, ::Nothing)

Stacktrace:
 [1] instantiate_function(f::Function, x::Vector{Float32}, ::GalacticOptim.AutoZygote, p::Nothing, num_cons::Int64)
   @ GalacticOptim C:\Users\User 1\.julia\packages\GalacticOptim\fow0r\src\function\zygote.jl:40
 [2] instantiate_function(f::Function, x::Vector{Float32}, ::GalacticOptim.AutoZygote, p::Nothing)
   @ GalacticOptim C:\Users\User 1\.julia\packages\GalacticOptim\fow0r\src\function\zygote.jl:4
 [3] sciml_train(::typeof(loss_neuralode), ::Vector{Float32}, ::ADAM, ::Nothing; lower_bounds::Nothing, upper_bounds::Nothing, maxiters::Int64, kwargs::Base.Pairs{Symbol, var"#43#45", Tuple{Symbol}, NamedTuple{(:cb,), Tuple{var"#43#45"}}})
   @ DiffEqFlux C:\Users\User 1\.julia\packages\DiffEqFlux\gH716\src\train.jl:87

排查思路

先排除Julia主版本影响：Julia 1.7.1在DiffEqFlux v1.13.0的官方兼容版本范围内，主版本不兼容的可能性极低
从报错栈定位根因：错误触发点在GalacticOptim包构造OptimizationFunction的环节，属于典型的依赖包版本接口不匹配问题
核对SciML生态迭代记录：对应版本迭代周期内SciML对优化层做过架构调整，原GalacticOptim后续重命名为Optimization.jl，若本地安装的GalacticOptim版本与DiffEqFlux v1.13.0要求的依赖版本不一致，就会出现构造函数参数列表不匹配的MethodError
提前排查次级适配点：部分版本迭代中调整过训练回调的入参顺序，修复版本问题后如果触发回调相关报错，可对应调整入参顺序。

解决方案

固定兼容的依赖版本，在Julia REPL中执行以下命令安装匹配的包版本：

using Pkg
Pkg.add(name="DiffEqFlux", version="1.13.0")
Pkg.add(name="GalacticOptim", version="2.0.2")
Pkg.add(name="OptimizationFlux", version="0.1.0")
Pkg.resolve()

重启Julia会话后重新运行代码即可正常执行。
如果不想固定旧版本，也可以直接升级DiffEqFlux到最新稳定版，按照最新版规范调整sciml_train为新的Optimization.solve接口写法，适配新版Optimization生态的调用规范。

注意：升级到最新版时，回调函数的参数顺序会变为callback(p, l, pred)，需要对应调整回调函数入参顺序，否则会出现参数不匹配的报错。

内容的提问来源于stack exchange，提问作者Adam Davis