trappmartin / sumproductnetworks.jl Goto Github PK

Hey Martin, I might be missing something obvious, but is there not yet a way to compute the log marginalized likelihood of the SPN?

i.e. in this mixture of bivariate gaussians, I can't find a way to work with the Normal(-1, 1) + Normal(1,1) mixture.

root = FiniteSumNode{Float32}();

add!(root, FiniteProductNode(), Float32(log(0.5))); # Weight 0.5
add!(root, FiniteProductNode(), Float32(log(0.5))); # Weight 0.5

add!(children(root)[1], UnivariateNode(Normal(-1, 1), 1))
add!(children(root)[2], UnivariateNode(Normal(1, 1), 1))
add!(children(root)[1], UnivariateNode(Normal(-1, 10), 2))
add!(children(root)[2], UnivariateNode(Normal(1, 10), 2))

The tag name "v.0.1.1" is not of the appropriate SemVer form (vX.Y.Z).
cc: @trappmartin

Hey, I was trying to run the example code in the readme. First line calls FiniteSumNode(). It looks like it'll need FiniteSumNode{Float64}() or some other concrete float type. Is this unintended and a bug, or should the readme be updated to reflect that? Thanks!

Hi Martin,
I was delighted to see an SPN implementation in Julia just as I was starting to dig into them. As part of my exploration, I was wondering if you had a timeline for filling out the generate_spn function, either in this package or another. I'm curious about replacing other function approximators with SPNs in some reinforcement learning projects I have going on; with the right structure, SPNs offer a lot over neural nets for these problems.

Thanks for opening your work!

Kendall

Todo:

• Make sure the package works under julia 1.0
• Add more unit tests
• Add a more extensive README and documentation of the package

I should make types immutable by default and add additional mutable types, e.g. MutableSumNode.

Further, I should sub-type Distributions.Distribution instead of my own SPNNode type. This way the code would be compatible with Turing.

• Make types immutable
• Add additional mutable types
• Sub-type Distributions instead of SPNNode
• Add fit overload for EM fitting of an SPN.

Make SumProductNetworks ready for julia 1.0

I should add region graphs, similar to the implementation for BayesianSPNs, at some point to this package.

• Add regions and partitions
• Add atomic regions for univariate and multivariate distributions
• Add random structure generation
• Add bayesian structure learning (posterior bootstrap?)

Add the implementation of safe semi-supervised learning of SPNs