Error when running infer due to splatting

[albertpod]

I know the issue with splatting was addressed here, but I am having a trouble with it, when running the infer function with the provided PCA model (see below). The following error is encountered:

ERROR: MethodError: no method matching iterate(::GraphPPL.VariableRef{…})

Closest candidates are:
  iterate(::Revise.LineSkippingIterator)
   @ Revise ~/.julia/packages/Revise/bAgL0/src/relocatable_exprs.jl:70
  iterate(::Revise.LineSkippingIterator, ::Any)
   @ Revise ~/.julia/packages/Revise/bAgL0/src/relocatable_exprs.jl:70
  iterate(::Base.MethodSpecializations)
   @ Base reflection.jl:1148
  ...

Stacktrace:
  [1] macro expansion
    @ ~/.julia/dev/GraphPPL/src/model_macro.jl:543 [inlined]
  [2] macro expansion

The error occurs at the line:

GraphPPL.jl/src/model_macro.jl

Line 542 in b3ae2a6

generate_lhs_proxylabel(var, index::Nothing) = quote

The minimum code example to reproduce the error:

using RxInfer

PCA_block(x, w...) = hcat(w...)*x
# PCA_block(x, w1, w2) = [w1 w2]*x

@model function pca_mode(y, components, obs_dim, lat_dim)
    local w
    for j in 1:components
        w[j] ~ MvNormal(μ=zeros(obs_dim), Λ=diageye(obs_dim))
    end

    for i in eachindex(y)
        x[i] ~ MvNormalMeanPrecision(ones(lat_dim), diageye(lat_dim))
        y[i] ~ MvNormal(μ=PCA_block(x[i], w...), Λ=diageye(obs_dim))
        # y[i] ~ MvNormal(μ=PCA_block(x[i], w[1], w[2]), Λ=diageye(obs_dim))
    end
end

components = 2
n_samples = 100
obs_dim = 4
lat_dim = 2

w1 = [2.0, -1.0, 0.5, -0.2] 
w2 = [0.8, 1.5, -0.3, 0.1]

latent_x = [rand(MvNormal(zeros(lat_dim), diageye(lat_dim))) for i in 1:n_samples]

y = [PCA_block(latent_x[i], w1, w2) + rand(MvNormal(zeros(obs_dim), 0.1diageye(obs_dim))) for i in 1:n_samples]

pca_meta = @meta begin
    PCA_block() -> Linearization()
end

initialization = @initialization begin
    μ(w) = MvNormalMeanPrecision(zeros(obs_dim), diageye(obs_dim))
end

result = infer(model=pca_mode(components=components, obs_dim=obs_dim, lat_dim=lat_dim), initialization=initialization, data=(y=y, ), meta=pca_meta, free_energy=true, iterations=5, showprogress=true, returnvars=KeepLast())

The issue appears to be related to the splatting of the w array in the PCA_block function.

[bvdmitri]

Can you try and see if it works as a workaround?

μ=PCA_block(in = [ x[i], w... ])

EDIT: ah, sorry, it won't work probably either

[bvdmitri]

I think its a real issue, but should be fixable. We need to define iterate, which probably should reuse some code from Base.broadcastable(ref::VariableRef) (it may even call broadcastable?). We may had some justification for not including iterate on VariableRef, but I cannot recall it.

[wouterwln]

@bvdmitri the problem when I try to define iterate for VariableRef is that under the hood proxylabel(name, v..., nothing, False()) will be called, which splats out the variable we want to splat inside the proxylabel function. Do you have an idea how to fix this? I was thinking we maybe fix the argument order of proxylabel to include a dispatch option on Vararg, but it is an ugly solution I do not like. What do you think? An alternative viable strategy?

[bvdmitri]

Can we transform v... into something like proxylabel(:v, v, Splat()), in the same way as we transform v[i] into proxylabel(:v, v, (i, )) and then later on dispatch on it? Only a rough idea..