This is a Python library for approximate unsupervised sampling inference in Bayesian Hidden Markov Models (HMMs) and explicit-duration Hidden semi-Markov Models (HSMMs), focusing on the Bayesian Nonparametric extensions, the HDP-HMM and HDP-HSMM, via the weak-limit approximation.

There are also some plugins to extend the functionality:

• factorial models
• autoregressive models
• collapsed HDP sampling inference.

The inference can be run in parallel over multiple cores and/or multiple machines (even on EC2!) using ipython's excellent ipython.parallel module. Someday I might even document how to do it!

You can clone this library and its dependencies into your current working directory with

git clone --recursive git://github.com/mattjj/pyhsmm.git

For the Python interpreter to be able to import pyhsmm, you'll need it on your Python path. Since the current working directory is usually included in the Python path, you can probably run the examples from the same directory in which you run the git clone with commands like python pyhsmm/examples/hsmm.py. You might also want to add pyhsmm to your global Python path (e.g. by copying it to your site-packages directory).

To pull updates, you can do this:

cd /path/to/pyhsmm
git pull
git submodule update --init --recursive

The library depends on numpy, scipy, and, for visualization, matplotlib.

Disabling assertions may speed things up; to disable assertions in CPython, call it with the -O flag.

Here's how to draw from the HDP-HSMM posterior over HSMMs given a sequence of observations. (The same example, along with the code to generate the synthetic data loaded in this example, can be found in examples/basic.py.)

Let's say we have some 2D data in a data.txt file:

$ head -5 data.txt
-3.711962552600095444e-02 1.456401745267922598e-01
7.553818775915704942e-02 2.457422192223903679e-01
-2.465977987699214502e+00 5.537627981813508793e-01
-7.031638516485749779e-01 1.536468304146855757e-01
-9.224669847039665971e-01 3.680035337673161489e-01

In Python, we can plot the data in a 2D plot, collapsing out the time dimension:

import numpy as np
from matplotlib import pyplot as plt

data = np.loadtxt('data.txt')
plt.plot(data[:,0],data[:,1],'kx')

We can also make a plot of time versus the first principal component:

from pyhsmm.util.plot import pca_project_data
plt.plot(pca_project_data(data,1))

To learn an HSMM, we'll use pyhsmm to create an HSMM instance using some reasonable hyperparameters. We'll ask this model to infer the number of states as well (since an HDP-HSMM is instantiated by default), so we'll give it an Nmax parameter:

import pyhsmm
import pyhsmm.basic.distributions as distributions

obs_dim = 2
Nmax = 25

obs_hypparams = {'mu_0':np.zeros(obs_dim),
                'sigma_0':np.eye(obs_dim),
                'kappa_0':0.3,
                'nu_0':obs_dim+5}
dur_hypparams = {'alpha_0':2*30,
                 'beta_0':2}

obs_distns = [distributions.Gaussian(**obs_hypparams) for state in range(Nmax)]
dur_distns = [distributions.PoissonDuration(**dur_hypparams) for state in range(Nmax)]

posteriormodel = pyhsmm.models.HSMM(
        alpha=6.,gamma=6., # better to sample over these; see concentration-resampling.py
        init_state_concentration=6., # pretty inconsequential
        obs_distns=obs_distns,
        dur_distns=dur_distns)

(The first two arguments set the "new-table" proportionality constant for the meta-Chinese Restaurant Process and the other CRPs, respectively, in the HDP prior on transition matrices. For this example, they really don't matter at all, but on real data it's much better to infer these parameters, as in examples/concentration_resampling.py.)

Then, we add the data we want to condition on:

posteriormodel.add_data(data,trunc=60)

The trunc parameter is an optional argument that can speed up inference: it sets a truncation limit on the maximum duration for any state. If you don't pass in the trunc argument, no truncation is used and all possible state duration lengths are considered. (pyhsmm has fancier ways to speed up message passing over durations, but they aren't documented.)

If we had multiple observation sequences to learn from, we could add them to the model just by calling add_data() for each observation sequence.

Now we run a resampling loop. For each iteration of the loop, all the latent variables of the model will be resampled by Gibbs sampling steps, including the transition matrix, the observation means and covariances, the duration parameters, and the hidden state sequence. We'll also copy some samples so that we can plot them.

models = []
for idx in progprint_xrange(150):
    posteriormodel.resample_model()
    if (idx+1) % 10 == 0:
        models.append(copy.deepcopy(posteriormodel))

Now we can plot our saved samples:

fig = plt.figure()
for idx, model in enumerate(models):
    plt.clf()
    model.plot()
    plt.gcf().suptitle('HDP-HSMM sampled after %d iterations' % (10*(idx+1)))
    plt.savefig('iter_%.3d.png' % (10*(idx+1)))

I generated these data from an HSMM that looked like this:

So the posterior samples look pretty good!

A convenient shortcut to build a list of sampled models is to write

model_samples = [model.resample_and_copy() for itr in progprint_xrange(150)]

That will build a list of model objects (each of which can be inspected, plotted, pickled, etc, independently) in a way that won't duplicate data that isn't changed (like the observations or hyperparameter arrays) so that memory usage is minimized. It also minimizes file size if you save samples like

import cPickle
with open('sampled_models.pickle','w') as outfile:
    cPickle.dump(model_samples,outfile,protocol=-1)

HSMMs constitute a much more powerful model class than plain-old HMMs, and that enhanced power comes with a computational price: each sampling iteration for an HSMM is much slower than that of an HMM. But that price is often worthwhile if you want to place priors on state durations or have the model learn duration structure present in the data. (In the example, strong duration structure is what made the inference algorithm latch onto the correct explanation so easily.) In addition, the increased cost of each iteration often pays for itself, since HSMM samplers empirically seem to take fewer iterations to converge than comparable HMM samplers.

Using my nothing-special i7-920 desktop machine and a NumPy/SciPy built against Intel's MKL along with the Eigen-backed classes, here's how long the demo iterations took:

$ python examples/hsmm.py
.........................  [  25/100,    0.05sec avg,    3.95sec ETA ]
.........................  [  50/100,    0.05sec avg,    2.64sec ETA ]
.........................  [  75/100,    0.05sec avg,    1.34sec ETA ]
.........................  [ 100/100,    0.05sec avg,    0.05sec ETA ]
   0.05sec avg,    5.21sec total

To add your own observation or duration distributions, implement the interfaces defined in basic/abstractions.py. Also see the plugins. To get a flavor of the style, see pybasicbayes.

• Matthew J. Johnson and Alan S. Willsky. Bayesian Nonparametric Hidden Semi-Markov Models. Journal of Machine Learning Research (JMLR), 14:673–701, February 2013.

• Matthew J. Johnson and Alan S. Willsky, The Hierarchical Dirichlet Process Hidden Semi-Markov Model. 26th Conference on Uncertainty in Artificial Intelligence (UAI 2010), Avalon, California, July 2010.

@article{johnson2013hdphsmm,
    title={Bayesian Nonparametric Hidden Semi-Markov Models},
    author={Johnson, Matthew J. and Willsky, Alan S.},
    journal={Journal of Machine Learning Research},
    pages={673--701},
    volume={14},
    month={February},
    year={2013},
}