GSoC Project - 2023

[ricardoV94]

@shreyas3156 Sorry for the delay in answering. All good questions, let me see if I can reply:

@shreyas3156

The pdf in the interval should be 0 and the pdf at the cutoff points should be the difference in the CDFs at that point and the previous one?

Yes, that’s my understanding

@shreyas3156

Are there more ways other than using pm.math.switch() to perform interval censoring? (I came across IfElse too but it does not execute both the branches of the clause so I don’t think it can be used)

I think switch is the most natural, with the last condition being the “didn’t match anything → logp=-np.inf”

@shreyas3156

Is it acceptable to define a new ScalarOp in PyTensor or is it something we generally want to avoid?

We generally want to avoid it but can be considered if it seems like a clear winner. Are you thinking of a Switch with multiple branches?

@shreyas3156

Since we use MeasurableVariables in the IR graphs and preserve the RV mappings, is the autodifferentiation carried out on the IR graph (through its logp) or the original graph? I ask this in light of the idea of defining a new MeasurableElemwise to replace the Elemwise{Switch} Op in the IR graph but I’m unsure if it will break posterior sampling.

Differentiation is carried out in the logp graph, so the IR doesn’t matter for that.