Models a sequence of states coming from an Autoregressive Hidden Markov Model [2] again as an HMM. Multiple runs can be run in parallel to compute using different numbers of hidden states of the HMM (meta-states).
Assuming independence of the different experimental trials fed to the algorithm, it is also possible to use cross-validation to determine the optimal (in the sense of maximal likelihood of the sequence of states) choice for the number of hidden states. See [3] for a discussion of two alternative methods in the case that one only has a single run for training, in which case removing samples for cross-validation would break the Markov property.
- Python 2.7, numpy, matplotlib.
- hmmlearn: unsupervised learning and
inference of Hidden Markov Models.
Install in conda with
conda install -c https://conda.anaconda.org/bcbio hmmlearnThis should also install numpy, scipy, scikit-learn and other dependencies. - concurrent.futures: Backport
from Python 3.2. Install in conda with
conda install futures
Usage: python metastates.py -i <input-file> -o <output-file>
Other options:
-h, --help This help
-i, --input-file Specifies the input file
-o, --output-file Specifies the output file
[default=input-file.out]
-f, --shift Apply shift to data in input file [default=-1]
-t, --trials Set number of trials for input file [default=1]
-s, --states Number of states to use [default=[2]]
(interpreted as a maximum when cross-validating)
-a, --auto Parse input filename for number of trials and
number of states (overrides -s, -t) [default=False]
-c, --crossval Perform n-fold cross-validation [default n=None]
-j, --jobs Number of concurrent jobs to launch [default=1]
-v, --verbose Display progress and time information
This software was developed with specific datasets in mind in the context of a short stay at the chair of Computational Neuroscience of the department of Neurobiology in the Ludwig-Maximilians-Universität and is not a general purpose implementation.
[1] L. R. Rabiner, “A tutorial on hidden Markov models and selected applications in speech recognition,” in Proceedings of the IEEE, 1989, vol. 77, pp. 257–286.
[2] M. J. Cassidy and P. Brown, “Hidden Markov based autoregressive analysis of stationary and non-stationary electrophysiological signals for functional coupling studies,” Journal of Neuroscience Methods, vol. 116, no. 1, pp. 35–53, Apr. 2002.
[3] G. Celeux and J.-B. Durand, “Selecting hidden Markov model state number with cross-validated likelihood,” Comput Stat, vol. 23, no. 4, pp. 541–564, Dec. 2007.
This software falls under the GNU general public license version 3 or later. It comes without any warranty whatsoever. For details see http://www.gnu.org/licenses/gpl-3.0.html.