Can hmmlearn (HMMGMM) be used for supervised learning?

[andrey-vinokurov-1991]

I am interested in HMM/GMM. I would like to use it for supervised learning, i.e. I have a sequence and labels for the sequence. Seqlearn allows us to use just MultinomialHMM. How can I implement hmmgmm for supervised learning using hmmlearn and seqlearn?
I want to solve this task http://cslu.ohsu.edu/~bedricks/courses/cs655/hw/hw4/hw4.html

[superbobry]

Hi, no hmmlearn does not currently target supervised learning. Have a look at seqlearn instead. It does not implement a GMMHMM, but I think you get around it by combining sklearn.mixture.GMM with sequlearn.hmm.MultinomialHMM.

I've skimmed through the task you've linked and noticed some minor inaccuracies regarding the hmmlearn API. Specifically hmmlearn does not have an implementation of GMM, it re-uses the one from sckit-learn.

[luigivieira]

Hi there @superbobry.

I found this discussion because I would like to use a GMMHMM in a supervised manner. I understand that I can not use seqlearn because my observations are not discrete. Your suggestion of using sklearn.mixture.GMM seems very interesting, but I don't even know how to start: I am new to HMM and GMM, and the documentation of scikit's GMM seems to indicate that the GMM is fit pretty much in the same way that it is done with hmmlearn: in a non-supervised fashion (i.e. without providing the labels for the samples).

Any help you could offer would be very appreciated. :)

[pls331]

@luigivieira have you get any idea of doing that? I was doing similar thing (continuous observation as well), using GMM of sklearn and HMM (and multivariate Gaussian distribution for modeling observation) in pomogranate. (But this gmm-hmm model does not work well for my case.)
Did you find any other possible alternative solution to do this?

[luigivieira]

@RoshanPAN Nope, I didn't find any alternative solution to do this. :(

[chananshgong]

Hi @AWin9 , I did something similar by feeding the model pretrained gmm per state the transition matrix and the start probabilities. It would look something like:

        model = GMMHMM(n_mix=self.n_mix,
                       n_components=n_states,
                       init_params='',
                       params='')
        # starting probabilities
        model.startprob_ = start_prob
        # transition matrix
        model.transmat_ = prob_trans_mat
        # gmm parameters
        model.gmms_ = gmm_per_state