Extension to HSMMs

[dan-sprague]

Hi, I'm curious what your take is on the feasibility of extending this package for HSMM (Hidden Semi Markov Models). Geometric state duration distributions can be undesirable when very short durations are implausible. This involves adding a distribution D, such as a NegBin, into the forward/backward/viterbi, and the BW if compatible with EM (I think it is?)

I'm curious if you think this is something that would fit into this package, or if you think this should be a separate project.

Thanks!

[gdalle]

Hi @dan-sprague!
Thank you for the suggestion. I think the extension to HSMMs is feasible in theory, but it might be a lot of work. Is it something you'd be interested in trying?

[rsenne]

I started working on a PR for this. I'm using PyHSMM, a now deprecated package, as somewhat of a reference. Right now I think i have forward/backward working, but am working out some bugs in baum-welch. Just wanted to check in that you'd still be interested in supporting this?

[gdalle]

I guess it will depend how different the interface must become, I'd have to see the PR to make a call. Worst-case scenario this can always become another package?

[rsenne]

Fair point. So far I'm trying to mimic the original interface as best I can. So far the biggest difference is a user also has to pass a vector of sojourn time distributions. Here is the analogous HSMM struct:

struct HSMM{
    V<:AbstractVector,
    M<:AbstractMatrix, 
    VD<:AbstractVector,
    VDur<:AbstractVector,
    Vl<:AbstractVector,
    Ml<:AbstractMatrix,
} <: AbstractHSMM
    "initial state probabilities"
    init::V
    "state transition probabilities (excluding self-transitions)"
    trans::M
    "observation distributions"
    dists::VD
    "state duration distributions"
    durations::VDur
    "logarithms of initial state probabilities"
    loginit::Vl
    "logarithms of state transition probabilities"
    logtrans::Ml

    function HSMM(init::AbstractVector, trans::AbstractMatrix, 
                  dists::AbstractVector, durations::AbstractVector)
        # Remove self-transitions from trans matrix for HSMMs
        trans_no_self = copy(trans)
        for i in 1:size(trans, 1)
            trans_no_self[i, i] = 0.0
        end
        # Renormalize rows
        foreach(sum_to_one!, eachrow(trans_no_self))
        
        log_init = elementwise_log(init)
        log_trans = elementwise_log(trans_no_self)
        
        hsmm = new{
            typeof(init),typeof(trans_no_self),typeof(dists),
            typeof(durations),typeof(log_init),typeof(log_trans)
        }(
            init, trans_no_self, dists, durations, log_init, log_trans
        )
        @argcheck valid_hsmm(hsmm)
        return hsmm
    end
end

I'll continue to work and let you know when I'm closer to PR'ing.