Problem with deterministic computation

[organic-chemistry]

Hello, I have a 'deterministic function' (called efficiency) than compute some differences between datavar and
then some product with a random variable and it gives me this error

MethodError: no method matching -(::DataVariable{PointMass{Float64}, Rocket.RecentSubjectInstance{Message{PointMass{Float64}}, Subject{Message{PointMass{Float64}}, AsapScheduler, AsapScheduler}}}, ::DataVariable{PointMass{Float64}, Rocket.RecentSubjectInstance{Message{PointMass{Float64}}, Subject{Message{PointMass{Float64}}, AsapScheduler, AsapScheduler}}})
Closest candidates are:
  -(::VectorizationBase.CartesianVIndex, ::Any) at ~/.julia/packages/VectorizationBase/G9NJI/src/cartesianvindex.jl:80
  -(::ChainRulesCore.AbstractThunk, ::Any) at ~/.julia/packages/ChainRulesCore/C73ay/src/tangent_types/thunks.jl:34
  -(::ChainRulesCore.NoTangent, ::Any) at ~/.julia/packages/ChainRulesCore/C73ay/src/tangent_arithmetic.jl:61
  ...

Stacktrace:
  [1] (::var"#2#5"{Vector{DataVariable{PointMass{Float64}, Rocket.RecentSubjectInstance{Message{PointMass{Float64}}, Subject{Message{PointMass{Float64}}, AsapScheduler, AsapScheduler}}}}, Vector{RandomVariable}, Int64})(j::Int64)
    @ Main ./array.jl:0
  [2] iterate
    @ ./generator.jl:47 [inlined]
  [3] collect
    @ ./array.jl:787 [inlined]
  [4] efficiency(tis::Vector{DataVariable{PointMass{Float64}, Rocket.RecentSubjectInstance{Message{PointMass{Float64}}, Subject{Message{PointMass{Float64}}, AsapScheduler, AsapScheduler}}}}, kis::Vector{RandomVariable})
    @ Main ./In[1]:10
  [5] macro expansion
    @ ./In[1]:38 [inlined]
  [6] macro expansion
    @ ~/.julia/packages/GraphPPL/n5QGe/src/model.jl:445 [inlined]
  [7] var"##estimate_kis#315"(model#312::FactorGraphModel{UnspecifiedConstraints, UnspecifiedMeta, RxInfer.ModelInferenceOptions{Nothing, Nothing, Nothing}}, n::Int64)
    @ Main ~/.julia/packages/RxInfer/kKYpd/src/graphppl.jl:33
  [8] ModelGenerator
    @ ~/.julia/packages/RxInfer/kKYpd/src/model.jl:248 [inlined]
  [9] #create_model#98
    @ ~/.julia/packages/RxInfer/kKYpd/src/model.jl:265 [inlined]
 [10] inference(; model::RxInfer.ModelGenerator{var"###estimate_kis#315", Tuple{Int64}, NamedTuple{(), Tuple{}}}, data::NamedTuple{(:E, :tis), Tuple{Vector{Float64}, Vector{Vector{Float64}}}}, initmarginals::Nothing, initmessages::Nothing, constraints::Nothing, meta::Nothing, options::Nothing, returnvars::Nothing, iterations::Nothing, free_energy::Bool, free_energy_diagnostics::Tuple{BetheFreeEnergyCheckNaNs, BetheFreeEnergyCheckInfs}, showprogress::Bool, callbacks::Nothing, addons::Nothing, postprocess::DefaultPostprocess, warn::Bool, catch_exception::Bool)
    @ RxInfer ~/.julia/packages/RxInfer/kKYpd/src/inference.jl:550
 [11] top-level scope
    @ In[1]:44
 [12] eval
    @ ./boot.jl:368 [inlined]
 [13] include_string(mapexpr::typeof(REPL.softscope), mod::Module, code::String, filename::String)
    @ Base ./loading.jl:1428

here is my code:

using RxInfer

function efficiency(tis,kis)
    tot = 0
    for i in 1:length(tis)-1
        denominator = sum([kis[j] for j in 1:i])
        first_sum = sum([kis[j]*(tis[i]-tis[j]) for j in 1:i])
        second_sum = sum([kis[j]*(tis[i+1]-tis[j]) for j in 1:i])
        e=2.718281828459
        term = 1/denominator * (e^(-first_sum)-e^(-second_sum))
        tot += term
    end
    return kis[1] * tot
end

@model function estimate_kis(n)
    
    E = datavar(Float64, n)
    tis = datavar(Float64,(2,3))

    kiss=[[2,1],[1,2]] 
    
    kis = randomvar(n)
    
    for i in 1:n
        kis[i] ~ Gamma(1.,1.)
    end
  
    for i = 1:n
        eff = efficiency(tis[i,:],[kis[j] for j in kiss[i]])
        E[i] ~ NormalMeanVariance(eff,0.1)
    end
end

E_exp=[0.1,0.1]

result = inference(
    model = estimate_kis(length(E_exp)),
    data  = (E = E_exp,tis=[[0,320,Inf],[0,320,Inf]])
)

I am not sure how to modify my function efficiency so that it works,
Any help would be appreciated!
Best

[albertpod]

Hi @organic-chemistry
Few comments about your problem.

• random vars and data vars cannot be used outside of the ~ context. Only with~.

eff ~ efficiency(tis[i,:],[kis[j] for j in kiss[i]])

• functions on the right hand side cannot accept arrays of random variables or data variables (implementation limitation, which will be fixed in the future) and ... operator is not supported.
• random vars and datavars cannot be mixed as inputs for deterministic relation
• custom deterministic functions require custom approximation methods and do not work out of the box (yet)

Generally speaking, if you want to use RxInfer.jl for your problem, you'd need to redesign the estimate_kis function. Have a look at the examples of delta nodes in our documentation (1, 2, 3)

Alternatively, you can run this model in Turing.jl.

[organic-chemistry]

Ok, thanks a lot for your fast reply.
I will check the examples,
Best