LearnSPN about sumproductnetworks.jl HOT 6 CLOSED

Thanks for the interest in the package.

I'm currently working on functionality related to AD support and missing values but can clean up my existing structure learning codes (several variants of learnSPN and random structures) to provide some functions asap. I plan to write a library on structure learning for SPNs containing all (or at least most of them) existing approaches in the near future.

from sumproductnetworks.jl.

When you say 'missing values', do you mean not supplying a full vector of data and allowing the SPN to provide a probability of the values not supplied? In the examples in the readme, this would be like doing something like logpdf with a X vector of length 1 instead of length 2, yes?

from sumproductnetworks.jl.

Yes, almost.

I aim to provide a syntax like logpdf([1.5, NaN]) where the missing value is in dimension 2 and set to NaN. Further, I like to have convenience functions to calculate the approximative MAP for the missing value.

I had part of the functionality some time a go and want to put it in place again. I believe it’s relevant in real world scenarios like robotics.

from sumproductnetworks.jl.

Very relevant, and one of the use cases I'm hoping SPNs can help with. Thanks for taking the time to implement this in Julia; I'm happy to help in whatever way I can! I have some experience with the Reverse/ForwardDiff.jl packages. Zygote.jl looks interesting, but may not be applicable here without major changes.

from sumproductnetworks.jl.

Great to hear that. Feel free to create a PR if you like to contribute something.

I want to have a generic approach so that any AD can be used. As I'm also working on the Turing.jl project my aim is to provide example code using ForwardDiff and Flux.Tracker so that SPNs can be used for probabilistic programming within Turing and for deep learning within Flux.

I'll push some parameter learning (log likelihood maximisation) examples asap.

from sumproductnetworks.jl.

Thanks to @kouariga there will soon be a clean implementation of learnSPN. PR: #12

from sumproductnetworks.jl.