Cryptic error messages

[ThijsvdLaar]

In order to improve the usability of ForneyLab it is important to return informative error messages. Often the search for the true origin of an error requires so much in-depth knowledge of ForneyLab internals that it becomes impossible to decrypt. Let's collect such errors here, together with a short description of how they arised. Separate issues or pull requests can be opened for improvement proposals.

[ivan-bocharov]

When a factor graph contains a dangling edge, constructing a recognition factorization results in a following error message:

LoadError: type Nothing has no field node
Stacktrace:
 [1] getproperty(::Any, ::Symbol) at ./sysimg.jl:18
 [2] nodes(::Set{Edge}) at ~/.julia/dev/ForneyLab/src/factor_graph.jl:108
 [3] (::getfield(ForneyLab, Symbol("##RecognitionFactor#137#139")))(::RecognitionFactorization, ::Symbol, ::Type, ::Set{Variable}) at ~/.julia/dev/ForneyLab/src/algorithms/variational_bayes/recognition_factorization.jl:36

...

An example to reproduce the error:

using ForneyLab

g = FactorGraph()

@RV a ~ GaussianMeanVariance(0.0,1.0)
@RV b ~ GaussianMeanVariance(a,1.0)
@RV c ~ GaussianMeanVariance(b,1.0)
@RV d ~ GaussianMeanVariance(c,1.0)

q = RecognitionFactorization(a,b,c)

[ismailsenoz]

When marginal rule is not specified while constructing a structured algorithm, I get the following error message:

, Edges:
Edge belonging to variable x_t: ( gpc_1.i[out] )----( gaussianmeanvariance_3.i[m] ).
Edge belonging to variable x_t_min: ( gaussianmeanvariance_2.i[out] )----( gpc_1.i[m] ).
) with inbound types Type[Message{GaussianMeanVariance,var_type} where var_type<:VariateType, Message{GaussianMeanVariance,var_type} where var_type<:VariateType, ProbabilityDistribution]```

[wmkouw]

I'm building a 2D Gaussian mixture model demo. This is the current setup:

Specify generative and recognition models

@RV _pi ~ Dirichlet([1.0, 1.0])
@RV m_1 ~ GaussianMeanVariance([0.0, 0.0], 100*[1. 0.; 0. 1.])
@RV w_1 ~ Wishart([1. 0.; 0. 1.], 2.)
@RV m_2 ~ GaussianMeanVariance([0.0, 0.0], 100*[1. 0.; 0. 1.])
@RV w_2 ~ Wishart([1. 0.; 0. 1.], 2.)

z = Vector{Variable}(undef, N)
y = Vector{Variable}(undef, N)
for i = 1:N
    @RV z[i] ~ Categorical(_pi)
    @RV y[i] ~ GaussianMixture(z[i], m_1, w_1, m_2, w_2)
    
    placeholder(y[i], :y, dims=(2,), index=i)
end

q = RecognitionFactorization(_pi, m_1, w_1, m_2, w_2, z, ids=[:PI, :M1, :W1, :M2, :W2, :Z])
algo = variationalAlgorithm(q)
algo_F = freeEnergyAlgorithm(q);

Execute inference algorithm

data = Dict(:y => X)

marginals = Dict(:_pi => ProbabilityDistribution(Dirichlet, a=[1.0, 1.0]),
                 :m_1 => ProbabilityDistribution(Multivariate, GaussianMeanVariance, m=[-1.0, -1.0], v=1e4*[1. 0.;0. 1.]),
                 :w_1 => ProbabilityDistribution(MatrixVariate, Wishart, v=[1. 0.;0. 1.], nu=2.),
                 :m_2 => ProbabilityDistribution(Multivariate, GaussianMeanVariance, m=[1.0, 1.0], v=1e4*[1. 0.;0. 1.]),
                 :w_2 => ProbabilityDistribution(MatrixVariate, Wishart, v=[1. 0.;0. 1.], nu=2.))
for i = 1:N
    marginals[:z_*i] = ProbabilityDistribution(Categorical, p=[1/2., 1/2.])
end

# Execute algorithm
num_iterations = 10
F = Float64[]
for i = 1:num_iterations
    stepZ!(data, marginals)
    stepPI!(data, marginals)
    stepM1!(data, marginals)
    stepW1!(data, marginals)
    stepM2!(data, marginals)
    stepW2!(data, marginals)
    
    push!(F, freeEnergy(data, marginals))
end

The error message I get is:

TypeError: in keyword argument m, expected Array{T,1} where T, got Float64

Stacktrace:
 [1] (::getfield(Core, Symbol("#kw#Type")))(::NamedTuple{(:m,),Tuple{Float64}}, ::Type{ProbabilityDistribution}, ::Type{Multivariate}, ::Type{PointMass}) at ./none:0
 [2] stepZ!(::Dict{Symbol,LinearAlgebra.Adjoint{Float64,Array{Float64,2}}}, ::Dict{Symbol,ProbabilityDistribution}, ::Array{Message,1}) at ./none:6
 [3] stepZ!(::Dict{Symbol,LinearAlgebra.Adjoint{Float64,Array{Float64,2}}}, ::Dict{Symbol,ProbabilityDistribution}) at ./none:5
 [4] top-level scope at ./In[54]:17

If you open up algo, you'll see:

begin

function stepZ!(data::Dict, marginals::Dict=Dict(), messages::Vector{Message}=Array{Message}(undef, 20))

messages[1] = ruleVBCategoricalOut(nothing, marginals[:_pi])
messages[2] = ruleVBGaussianMixtureZBer(ProbabilityDistribution(Multivariate, PointMass, m=data[:y][9]), nothing, marginals[:m_1], marginals[:w_1], marginals[:m_2], marginals[:w_2])
messages[3] = ruleVBCategoricalOut(nothing, marginals[:_pi])
...

So, GaussianMixture still thinks z is Bernoulli distributed (it's calling ruleVBGaussianMixtureZBer), even though it was specified to be Categorical. This is not obvious from the error message.