`initmarginals` broadcasting not working for `PointMass`

[bartvanerp]

It is possible to start of the inference procedure like

inference(
    model = model(nr_samples),
    initmarginals = (z= NormalMeanVariance(0,1), ),
    data = (y = y, ),
)

which initializes all variables z with the NormalMeanVariance(0,1) distribution. However, it throws an iteration error for PointMass objects, which need to be explicitly repeated as

inference(
    model = model(nr_samples),
    initmarginals = (z= repeat([PointMass(10)], nr_samples), ),
    data = (y = y, ),
)

[bvdmitri]

This task has been added to the milestone for tracking and prioritization.

[bvdmitri]

To solve this you simply need to add

Base.broadcastable(pm::PointMass) = Ref(pm)

[bvdmitri]

@Chengfeng-Jia This task has been added to the milestone for tracking and prioritization.

[Chengfeng-Jia]

I tried running the following code,

inference(
    model = model(nr_samples),
    initmarginals = (z= PointMass(1.0), ),
    data = (y = y, ),
)

No error reported, should it have been fixed? Is there a mistake in my understanding

[bartvanerp]

@Chengfeng-Jia was z an array of random variables in your model? In this case the issue seems to have fixed itself.

[Chengfeng-Jia]

I misunderstood before, now I changed z to an array of random variables. Indeed, this code doesn't work. initmarginals = (z= PointMass(0),)

However, if I change the code like this, it runs well and doesn't need to repeat initmarginals = (z= [PointMass(0)],)

Do you think it's fine, or still need the broadcast part

[bartvanerp]

I would prefer initmarginals = (z= PointMass(0),) to work, similarly as we can do initmarginals = (z= Normal(0,1),)

[Chengfeng-Jia]

Uh, I see.

[bvdmitri]

However, if I change the code like this, it runs well and doesn't need to repeat initmarginals = (z= [PointMass(0)],)

It's a "bug", it silently initializes only one (the first) variable

[bartvanerp]

Will be closed by https://github.com/biaslab/ReactiveMP.jl/pull/360