[Question] Support for emission probabilites as a PDF?

[ishouldbedany]

Firstly, sorry if this is not the correct place to submit questions such as this one.

I have been reviewing all the Python package for implementing discrete-time HMMs and came across your work. First of all, thanks. Very neat library (plot_hmm is great!).

I am wondering if it's possible to have emission probabilities defined as a PDF (say, multinomial), instead of the current matrix form. In my use case, emissions take continuous values. Is this possible or is it out of the library's scope? Thanks!

[iamsurya]

I haven't used this toolkit yet, but I am considering using it to compare to another HMM toolkit. I wrote my own C code to calculate the probability from continuous Gaussian (1 Gaussian per feature, no multigaussian support, which I will compare against this). There is a MATLAB / Octave toolkit that does support continuous emissions: https://www.cs.ubc.ca/~murphyk/Software/HMM/hmm.html

I'll wait for the project author to write back, and if they don't in a week, try to chime in with help.

[lopatovsky]

Hi ishouldbedany, thank you for your positive review!
Sadly current version doesn't support emissions with continuous values.
However it is proposed to be a feature in library second version. (The time of release is a deep unknown)

[iamsurya]

ishouldbedany, do you already have the hmm parameters (transition, prior matrices and the means, covars for the emission matrix)? If you already do, I could add some code to this toolkit to use pdf's. How many observables (features) do you have?

[ishouldbedany]

@iamsurya I can estimate the prior means and covariances for the emission matrix of each state easily (currently, I have 12 observable features, but I am still playing around with them). I'd prefer for the transition matrix to be estimated automatically, but I can write code to do it.

As for the priors, for me it's OK if they are randomly initialized. Thanks!

[shahab-anjum]

transition matrix? im also stuck in that