Julia中使用ReverseDiff遇类型错误及矩阵乘法歧义问题求助

问题1：严格类型签名导致的MethodError #

场景重现 #

原始代码可正常执行：

error = prior_error(data,s_vals,ones(20)/20)

打包参数后调用ReverseDiff的gradient函数计算梯度时出错：

inputs = (data,s_vals,ones(20)/20)
test = gradient(prior_error,inputs)

prior_error的原始类型签名：

prior_error(data::Matrix,sample_vals::Vector,prior::Vector)

错误信息 #

MethodError: no method matching prior_error(::ReverseDiff.TrackedArray{Float64, Float64, 2, Matrix{Float64}, Matrix{Float64}}, ::ReverseDiff.TrackedArray{Float64, Float64, 1, Vector{Float64}, Vector{Float64}}, ::ReverseDiff.TrackedArray{Float64, Float64, 1, Vector{Float64}, Vector{Float64}})

ReverseDiff.GradientTape(::Function, ::Tuple{Matrix{Float64}, Vector{Float64}, Vector{Float64}}, ::ReverseDiff.GradientConfig{Tuple{ReverseDiff.TrackedArray{Float64, Float64, 2, Matrix{Float64}, Matrix{Float64}}, ReverseDiff.TrackedArray{Float64, Float64, 1, Vector{Float64}, Vector{Float64}}, ReverseDiff.TrackedArray{Float64, Float64, 1, Vector{Float64}, Vector{Float64}}}})@tape.jl:207
gradient(::Function, ::Tuple{Matrix{Float64}, Vector{Float64}, Vector{Float64}}, ::ReverseDiff.GradientConfig{Tuple{ReverseDiff.TrackedArray{Float64, Float64, 2, Matrix{Float64}, Matrix{Float64}}, ReverseDiff.TrackedArray{Float64, Float64, 1, Vector{Float64}, Vector{Float64}}, ReverseDiff.TrackedArray{Float64, Float64, 1, Vector{Float64}, Vector{Float64}}}})@gradients.jl:22
top-level scope@Local: 5

原因与解决方法 #

• 原因：ReverseDiff计算梯度时会将输入包装为TrackedArray类型，但原始函数签名限定参数必须是 concrete 类型Matrix/Vector，而TrackedArray并不直接继承这些类型，导致方法分派失败。
• 解决方法：将类型签名改为抽象类型，既保留类型约束又兼容ReverseDiff的跟踪数组：

  prior_error(data::AbstractMatrix, sample_vals::AbstractVector, prior::AbstractVector)
  

问题2：矩阵乘法方法歧义错误 #

场景重现 #

移除类型签名后自动微分可进行，但执行矩阵乘法时出现新错误：

错误信息 #

MethodError: *(::ReverseDiff.TrackedArray{Float64, Float64, 2, Matrix{Float64}, Matrix{Float64}}, ::LinearAlgebra.Diagonal{ReverseDiff.TrackedReal{Float64, Float64, ReverseDiff.TrackedArray{Float64, Float64, 1, Vector{Float64}, Vector{Float64}}}, ReverseDiff.TrackedArray{Float64, Float64, 1, Vector{Float64}, Vector{Float64}}}) is ambiguous. Candidates:

*(A::AbstractMatrix, D::LinearAlgebra.Diagonal) in LinearAlgebra at /home/peter/julia-1.7.3/share/julia/stdlib/v1.7/LinearAlgebra/src/diagonal.jl:222

*(x::ReverseDiff.TrackedArray{V, D, 2}, y::AbstractMatrix) where {V, D} in ReverseDiff at /home/peter/.julia/packages/ReverseDiff/GtPeW/src/derivatives/linalg/arithmetic.jl:213

*(x::ReverseDiff.TrackedArray{V, D, 2}, y::AbstractArray) where {V, D} in ReverseDiff at /home/peter/.julia/packages/ReverseDiff/GtPeW/src/derivatives/linalg/arithmetic.jl:213

*(x::ReverseDiff.TrackedArray{V, D}, y::AbstractMatrix) where {V, D} in ReverseDiff at /home/peter/.julia/packages/ReverseDiff/GtPeW/src/derivatives/linalg/arithmetic.jl:213

*(A::AbstractMatrix, B::AbstractMatrix) in LinearAlgebra at /home/peter/julia-1.7.3/share/julia/stdlib/v1.7/LinearAlgebra/src/matmul.jl:151

*(x::ReverseDiff.TrackedArray{V, D}, y::AbstractArray) where {V, D} in ReverseDiff at /home/peter/.julia/packages/ReverseDiff/GtPeW/src/derivatives/linalg/arithmetic.jl:213

Possible fix, define

*(::ReverseDiff.TrackedArray{V, D, 2}, ::LinearAlgebra.Diagonal) where {V, D}

原因与解决方法 #

• 原因：Julia的方法分派系统无法确定优先调用哪个乘法实现——LinearAlgebra提供了AbstractMatrix * Diagonal的实现，ReverseDiff提供了TrackedArray * AbstractArray/AbstractMatrix的实现，两者优先级相同导致歧义。
• 解决方法（任选其一）：
  1. 手动定义歧义方法：按照错误提示添加专门的方法，确保ReverseDiff跟踪信息不丢失：

     import ReverseDiff: TrackedArray
     import LinearAlgebra: Diagonal
     
     function *(x::TrackedArray{V, D, 2}, y::Diagonal) where {V, D}
         ReverseDiff.TrackedArray(x.track, x.data * y)
     end
     

  2. 转换Diagonal为普通矩阵：如果不需要保留对角矩阵的结构，直接转换后再计算：

     result = tracked_matrix * Matrix(diagonal_matrix)
     

  3. 升级ReverseDiff版本：检查是否有新版本已修复该方法歧义问题，更新包后可能自动解决。

内容的提问来源于stack exchange，提问作者Pablo