zjcerwin / hmm

First commit

hmm

An implementation of hidden Markov model in Python.
The Baum-welch and Viterbi algorithms for discrete emissions are implemented so far. It is a generic implementation, so one may need to write some wrapper to apply some real data.

Usage

Configure three parameters, which are transition, emission and initial probabilities. It is assumed that these are instances of numpy.array class. All emissions should be represented as integer, which are the indices of emission probability matrix.

import numpy as np
import hmm

t = np.array([[0.55, 0.45], [0.4, 0.6]])
e = np.array([[0.6, 0.3, 0.1], [0.1, 0.4, 0.5]])
i = np.array([0.4, 0.6])
h = hmm.HMM(t, e, i)

observations = np.array([[2, 2, 2, 2, 1, 0, 0, 1, 1, 1, 0, 2],
        [2, 2, 2, 1, 2, 1, 0, 0, 0, 1, 1, 2, 1, 2],
        [1, 0, 2, 2, 0, 0, 1, 1, 1, 0, 2, 0, 2, 0]])
h.baum_welch(observations, iter_limit=1000, 
        threshold=1e-5, pseudocounts=[0, 1e-4, 0])

path, l = h.viterbi([2, 2, 1, 0, 0, 2, 1, 1, 1, 0, 0])

Using multiprocessing module, you can speed up the calculation.

import numpy as np
import hmm_mp

... # Configure parameters t, e, i

h = hmm_mp.MultiProcessHMM(t, e, i, worker_num=4)

h.baum_welch(observations)

This module also offers importing an XML file of GHMM.

import hmm

(t, e, i) = hmm.load_ghmmxml('filename.xml')
h = hmm.MultiProcessHMM(t, e, i, 4)

Methods

• baum_welch : A basic Baum-Welch algorithm. It requires a list of observations, where each observation is represented as a list of integers.