Regression, new bug, related to useMsgLikelihoods

[dehann]

Looks like it has something to do with getSolverParams(fg).useMsgLikelihoods = true. MWE:

using IncrementalInference

fg = generateGraph_CaesarRing1D()
getSolverParams(fg).useMsgLikelihoods = true

solveGraph!(fg)

Code completes fine if useMsgLikelihoods=false

Originally posted by @dehann in #1572 (comment)

[dehann]

Was first seen in testDefaultDeconv.jl

[dehann]

This is the printout:

IncrementalInference.jl/src/Factors/LinearRelative.jl

Line 41 in 60134f9

!(z isa Vector{Float64}) && (@warn "H" z x1 x2)

@Affie, what does the "H" mean? Looks like the way to read this line of code is that z should be a Vector{Float64} -- so being a ::Matrix is bad?

┌ Warning: H
│ z = 1×1 Matrix{Float64}: …
│ x1 = 1-element Vector{Float64}: …
│ x2 = 1-element Vector{Float64}: …
└ @ IncrementalInference ~/.julia/dev/IncrementalInference/src/Factors/LinearRelative.jl:41

[dehann]

So here is one inconsistency, sampleFactor returns different types depending on whether the factor was created by the user or from a tree likelihood message.

This is good (returns Vector{Flpat64}):

IncrementalInference.jl/test/testUseMsgLikelihoods.jl

Line 16 in c249ff1

@test z isa Vector{<:Real}

But then just lower down this, which is bad (returns Matrix{Float64}):

IncrementalInference.jl/test/testUseMsgLikelihoods.jl

Line 63 in c249ff1

@test z isa Vector{<:Real}

[dehann]

Ah, that's because the differential factor from the tree likelihood factor is an MKD (not Normal like user factors) which uses a different getSample function. There is some inconsistency there.

[dehann]

Okay so sampleTangent on ::Normal returns a ::Vector:

IncrementalInference.jl/test/testUseMsgLikelihoods.jl

Lines 25 to 26 in b4e9c6f

	X = sampleTangent(M, fT.Z)
	@test X isa Vector{<:Real}

while sampleTangent on ::MKD returns a ::Matrix:

IncrementalInference.jl/test/testUseMsgLikelihoods.jl

Lines 83 to 84 in b4e9c6f

	X = sampleTangent(M, fT.Z)
	@test X isa Vector{<:Real}

[dehann]

So fixed with forced type change of sample return for MKD case. If matrix has just one column, then convert to a vector:
71de84e

[Affie]

KDE returns a matrix on sample and it is used as coordinates which should be a vector, a recent update in Manifolds.jl showed this error.