Hidden Markov Models in Python, with scikit-learn like API

I am trying to create audio vocabulary from MFCC features by applying HMM. Since I have 10 speakers in the MFCC features. I need 50 states per speaker. So I used N = 500 states and it throws Memory error, but it works fine with N =100 states.

Memory Error is because of computational in efficiency of a machine or due to in proper initialization?

Here is my code

import numpy as np
from hmmlearn import hmm
import librosa
import matplotlib.pyplot as plt

def getMFCC(episode):

    filename = getPathToGroundtruth(episode)

    y, sr = librosa.load(filename)  # Y gives 

    data = librosa.feature.mfcc(y=y, sr=sr)

    return data

def hmm_init(n,data):  #n = states d = no of feautures

    states =[]

    model = hmm.GaussianHMM(n_components=N, covariance_type="full")

    model.transmat_ = np.ones((N, N)) / N

    model.startprob_ = np.ones(N) / N

    fit = model.fit(data.T)

    z=fit.decode(data.T,algorithm='viterbi')[1]

    states.append(z)

    return states

data_m = getMFCC(1)  # Provides MFCC features of numpy array [20 X 56829]

N = 500

D= len(data)

states = hmm_init(N,data)

In [23]: run Final_hmm.py
---------------------------------------------------------------------------
MemoryError                               Traceback (most recent call last)
/home/elancheliyan/Final_hmm.py in <module>()
     73 D= len(data)
     74 
---> 75 states = hmm_init(N,data)
     76 states.dump("states")
     77 

/home/elancheliyan/Final_hmm.py in hmm_init(n, data)
     57     model.startprob_ = np.ones(N) / N
     58 
---> 59     fit = model.fit(data.T)
     60 
     61     z=fit.decode(data.T,algorithm='viterbi')[1]

/cal/homes/elancheliyan/.local/lib/python3.5/site-packages/hmmlearn-0.2.1-py3.5-linux-x86_64.egg/hmmlearn/base.py in fit(self, X, lengths)
    434                 self._accumulate_sufficient_statistics(
    435                     stats, X[i:j], framelogprob, posteriors, fwdlattice,
--> 436                     bwdlattice)
    437 
    438             # XXX must be before convergence check, because otherwise

/cal/homes/elancheliyan/.local/lib/python3.5/site-packages/hmmlearn-0.2.1-py3.5-linux-x86_64.egg/hmmlearn/hmm.py in _accumulate_sufficient_statistics(self, stats, obs, framelogprob, posteriors, fwdlattice, bwdlattice)
    221                                           posteriors, fwdlattice, bwdlattice):
    222         super(GaussianHMM, self)._accumulate_sufficient_statistics(
--> 223             stats, obs, framelogprob, posteriors, fwdlattice, bwdlattice)
    224 
    225         if 'm' in self.params or 'c' in self.params:

/cal/homes/elancheliyan/.local/lib/python3.5/site-packages/hmmlearn-0.2.1-py3.5-linux-x86_64.egg/hmmlearn/base.py in _accumulate_sufficient_statistics(self, stats, X, framelogprob, posteriors, fwdlattice, bwdlattice)
    620                 return
    621 
--> 622             lneta = np.zeros((n_samples - 1, n_components, n_components))
    623             _hmmc._compute_lneta(n_samples, n_components, fwdlattice,
    624                                  log_mask_zero(self.transmat_),

MemoryError:

In sklearn GMM was replaced by GaussianMixture. See https://github.com/scikit-learn/scikit-learn/blob/master/sklearn/mixture/gmm.py:

class GMM(_GMMBase): """ Legacy Gaussian Mixture Model .. deprecated:: 0.18 This class will be removed in 0.20. Use :class:sklearn.mixture.GaussianMixture instead. """

However, hmmlearn still uses the old version. A pull request is needed to upgrade hmmlearn to work with the newer API.

@anntzer Leaving an early draft of incorporating Variational Inference training of HMMs so I may receive feedback before I keep going.

Some Notes:

• I derive from BaseHMM, and am able to reuse most of it, with a few exceptions.
• VariationalGaussianHMM is still incomplete - only Full Covariance is supported.
• Tests are lacking.

Up Next:

• Finish the different covariance types for Gaussian
• Add Mixture of Gaussian Emissions

Hi, today I learned about your package, started to use it, faced the memory problem, and came up with a PR that fixes it.

I've exploited the lengths option and added another meaning to it. Currently, for the GMMHMM only. Curious users will find a way to extend my implementation to other models as well.

This also partially addresses the comment left in https://github.com/hmmlearn/hmmlearn/commit/08dee6640483cda232f7d2fcc7935d4008f4d368:

https://github.com/hmmlearn/hmmlearn/blob/0562ca65756ffb60da836eeeb1845e61767c705b/lib/hmmlearn/hmm.py#L918-L922

I got rid of the unnecessary 'centered' arrays in the stats dict. If you don't want to store the post_comp_mix matrices in the stats, the logic of computing intermediate variables - c_n and c_d for the covariance - should be moved from the _do_mstep to _accumulate_sufficient_statistics function. Since this is my first PR, I've decided not to rummage through your code a lot. In either case, this should be considered in a separate PR, if you will.

Best, Danylo

Hi there!

Using the latest master of hmmlearn, I tried running a simple MultinomialHMM example (code below) that results in the following error:

File "build/bdist.macosx-10.5-x86_64/egg/hmmlearn/base.py", line 307, in decode ValueError: could not broadcast input array from shape (6) into shape (1)

Could you please tell me what i am doing wrong? My expectation is that applying Viterbi should give me the most probable hidden sequence. However passing a list of observation doesn't work unlike passing a single value which does.

Thanks!

Vlad

from __future__ import division
import numpy as np
from hmmlearn import hmm

states = ["Rainy", "Sunny"]
n_states = len(states)

observations = ["walk", "shop", "clean"]
n_observations = len(observations)

start_probability = np.array([0.6, 0.4])

transition_probability = np.array([
  [0.7, 0.3],
  [0.4, 0.6]
])

emission_probability = np.array([
  [0.1, 0.4, 0.5],
  [0.6, 0.3, 0.1]
])

model = hmm.MultinomialHMM(n_components=n_states)
model.startprob=start_probability
model.transmat=transition_probability
model.emissionprob=emission_probability

# predict a sequence of hidden states based on visible states
bob_says = [0, 2, 1, 1, 2, 0]
model = model.fit(bob_says)
logprob, alice_hears = model.decode(bob_says, algorithm="viterbi")
print "Bob says:", ", ".join(map(lambda x: observations[x], bob_says))
print "Alice hears:", ", ".join(map(lambda x: states[x], alice_hears))

Hi,

I used the hmm module from sklearn and tried to replace it by the hmmlearn module. Unfortunately I could not import it to my notebook.

from hmmlearn import hmm --------------------------------------------------------------------------- ImportError Traceback (most recent call last) <ipython-input-7-8b8c029fb053> in <module>() ----> 1 from hmmlearn import hmm

ImportError: cannot import name hmm

I tried first pip-3.3 install git+https://github.com/hmmlearn/hmmlearn.git

As this didn't work I cloned the project and run the setup.py (with python 3.3) but I still get an import error.

If I try to import

import hmmlearn.hmm

I get another error

ImportError Traceback (most recent call last) <ipython-input-8-8dbb2cfe75b2> in <module>() ----> 1 import hmmlearn.hmm

/home/ipython/python/lib/python3.3/site-packages/hmmlearn/hmm.py in <module>() 22 from sklearn import cluster 23 ---> 24 from .utils.fixes import log_multivariate_normal_density 25 26 from . import _hmmc

ImportError: No module named 'hmmlearn.utils'

What did I do wrong?

Cheers, Evelyn

I get hmmlearn/_hmmc.c:239:28: fatal error: numpy/npy_math.h: No such file or directory yet the installation seems to finish successfully.

requirements.txt file:

click==6.7
cython==0.25.2
joblib==0.11
numpy==1.12.1
pandas==0.19.2
python-speech-features==0.5
scikit-learn==0.18.1
scipy==0.19.0
hmmlearn==0.2.0

Running setup.py bdist_wheel for hmmlearn: started
  Running setup.py bdist_wheel for hmmlearn: finished with status 'error'
  Complete output from command /opt/conda/bin/python -u -c "import setuptools, tokenize;__file__='/tmp/pip-build-8l6nu2n1/hmmlearn/setup.py';f=getattr(tokenize, 'open', open)(__file__);code=f.read().replace('\r\n', '\n');f.close();exec(compile(code, __file__, 'exec'))" bdist_wheel -d /tmp/tmpi_45qjtvpip-wheel- --python-tag cp36:
  running bdist_wheel
  running build
  running build_py
  creating build
  creating build/lib.linux-x86_64-3.6
  creating build/lib.linux-x86_64-3.6/hmmlearn
  copying hmmlearn/hmm.py -> build/lib.linux-x86_64-3.6/hmmlearn
  copying hmmlearn/utils.py -> build/lib.linux-x86_64-3.6/hmmlearn
  copying hmmlearn/base.py -> build/lib.linux-x86_64-3.6/hmmlearn
  copying hmmlearn/__init__.py -> build/lib.linux-x86_64-3.6/hmmlearn
  creating build/lib.linux-x86_64-3.6/hmmlearn/tests
  copying hmmlearn/tests/test_utils.py -> build/lib.linux-x86_64-3.6/hmmlearn/tests
  copying hmmlearn/tests/test_gaussian_hmm.py -> build/lib.linux-x86_64-3.6/hmmlearn/tests
  copying hmmlearn/tests/test_gmm_hmm.py -> build/lib.linux-x86_64-3.6/hmmlearn/tests
  copying hmmlearn/tests/test_multinomial_hmm.py -> build/lib.linux-x86_64-3.6/hmmlearn/tests
  copying hmmlearn/tests/test_base.py -> build/lib.linux-x86_64-3.6/hmmlearn/tests
  copying hmmlearn/tests/__init__.py -> build/lib.linux-x86_64-3.6/hmmlearn/tests
  running build_ext
  building 'hmmlearn._hmmc' extension
  creating build/temp.linux-x86_64-3.6
  creating build/temp.linux-x86_64-3.6/hmmlearn
  gcc -pthread -Wsign-compare -DNDEBUG -g -fwrapv -O3 -Wall -Wstrict-prototypes -fPIC -I/opt/conda/include/python3.6m -c hmmlearn/_hmmc.c -o build/temp.linux-x86_64-3.6/hmmlearn/_hmmc.o -O3
  hmmlearn/_hmmc.c:239:28: fatal error: numpy/npy_math.h: No such file or directory
   #include "numpy/npy_math.h"
                              ^
  compilation terminated.
  error: command 'gcc' failed with exit status 1
  
  ----------------------------------------
  Failed building wheel for hmmlearn
  Running setup.py clean for hmmlearn
Successfully built python-speech-features
Failed to build hmmlearn
Installing collected packages: click, cython, joblib, numpy, pytz, python-dateutil, pandas, python-speech-features, scikit-learn, scipy, hmmlearn
  Running setup.py install for hmmlearn: started
    Running setup.py install for hmmlearn: finished with status 'done'
**Successfully installed** click-6.7 cython-0.25.2 **hmmlearn-0.2.0** joblib-0.11 numpy-1.12.1 pandas-0.19.2 python-dateutil-2.6.0 python-speech-features-0.5 pytz-2017.2 scikit-learn-0.18.1 scipy-0.19.0

When I used GaussianHMM().fit() to train HMM, there is a RuntimeWarning: divide by zero encountered in log. Then I found that the start probability would approach to 0 after several EM iterations. My question is how to avoid probability approaching to 0 ?

my OS is win7 x64 . visual studio 2015, also visual studio 2013, and python 3.5 x64(by anaconda) are set up. hmmlearn is set up successfully. and validated by the code: >>>import hmmlearn >>> hmmlearn.version and the output is '0.2.0' which is last version of hmmlearn. but, if i put the code like the following, >>>from hmmlearn import hmm i get the error as the following,

C:\Anaconda3_64\python.exe E:/pycharm/plot_hmm_stock_analysis/hmm_stock_analysis.py Traceback (most recent call last): File "E:/pycharm/plot_hmm_stock_analysis/hmm_stock_analysis.py", line 17, in from hmmlearn import hmm File "C:\Anaconda3_64\lib\site-packages\hmmlearn-0.2.0-py3.5-win-amd64.egg\hmmlearn\hmm.py", line 14, in from sklearn import cluster File "C:\Anaconda3_64\lib\site-packages\sklearn__init__.py", line 57, in from .base import clone File "C:\Anaconda3_64\lib\site-packages\sklearn\base.py", line 11, in from .utils.fixes import signature File "C:\Anaconda3_64\lib\site-packages\sklearn\utils__init__.py", line 11, in from .validation import (as_float_array, File "C:\Anaconda3_64\lib\site-packages\sklearn\utils\validation.py", line 16, in from ..utils.fixes import signature File "C:\Anaconda3_64\lib\site-packages\sklearn\utils\fixes.py", line 324, in from scipy.sparse.linalg import lsqr as sparse_lsqr File "C:\Anaconda3_64\lib\site-packages\scipy\sparse\linalg__init__.py", line 109, in from .isolve import * File "C:\Anaconda3_64\lib\site-packages\scipy\sparse\linalg\isolve__init__.py", line 6, in from .iterative import * File "C:\Anaconda3_64\lib\site-packages\scipy\sparse\linalg\isolve\iterative.py", line 7, in from . import _iterative ImportError: DLL load failed: The specified module could not be found.

why? and how to fix it!?

by the way, if in cmd, using "pip freeze" commond, it shows hmmlearn and the version of it is 0.2.0. BUT, if using "conda list", no hmmlearn shows!!

Hi all, I am having a problem when trying to fit multiple GMMHMM models to solve a classification problem of emotions recognition from speech samples. Basically, the models often don't converge: even if the monitor reports 'True' if printed, I can see in the history that the likelihood is not strictly increasing. Actually, it decreases at some point and the training procedure stops.

Here, I report only the procedure for training one of the models (I should have seven, each one trained with a different training set). The data loaded are attached: data_training.npy.zip

from hmmlearn import hmm
import numpy as np 
data = np.load('data_training.npy', allow_pickle=True)


hmm = hmm.GMMHMM(n_components=2, n_mix=2,n_iter=1000, covariance_type="diag", verbose=True ) 

X_sequence_concat = np.concatenate(data) 
lengths = []
for el in data:
    lengths.append(len(el))

hmm.fit(X_sequence_concat, np.array(lengths))   
print("Is the HMM training converged? " + str(hmm.monitor_.converged))

In my actual implementation I have to do this for seven different models and sometimes I get this problem and sometimes I don't, as you can see from the results reported below:

Can you please help me? I'm really struggling with this and I can't find a possible cause of the problem.

Thanks in advance!

Adds a PoissonHMM with an example.

I think this is somewhat close so if you have time, a review would be great @anntzer @blckmaxima.

I'm not sure if there is a standard we could compare to like the Wikipedia one for the MultiNomialHMM or if that's necessary.

I am requesting a new feature

Everywhere in the code, Kmeans clusters init uses Kmeans default params (except n_clusters) :

    n_init=10,
    max_iter=300,
    tol=1e-4,
    verbose=0,
    random_state=None,
    copy_x=True,
    algorithm="lloyd"

...

main_kmeans = cluster.KMeans(n_clusters=nc, random_state=self.random_state) or kmeans = cluster.KMeans(n_clusters=self.n_components, random_state=self.random_state)

I got great improvments in my particular case (lot of very very noisy datasets) by modifying kmeans cluster initialization

kmeans = cluster.KMeans(n_clusters=self.n_components,n_init=100, random_state=self.random_state, tol=1e-6)

so will be great to allow to pass kmeans parameters when instanciating the model.

for instance: hmm = hmm.GaussianHMM(n_components, ..., kmeans_params={'n_init': xxx, 'max_iter': yyy, 'tol':zzz})

the n_init params for kmeans++ is quite important in some cases.

thx

Reference Issues/PRs

None

What does this implement/fix? Explain your changes.

• Adds the methods _n_parameters, bic, and aic to the GaussianHMM class.
• Essentially I copied the implementation from sklearn implementation for gaussian mixture model

Any other comments?

I haven't fully implemented the methods for the other classes, e.g. GMMHMM etc. but it's in the works.

# make our generative model with two components, a fair die and a
# loaded die
gen_model = hmm.MultinomialHMM(n_components=2, random_state=99)

# the first state is the fair die so let's start there so no one
# catches on right away
gen_model.startprob_ = np.array([1.0, 0.0])

# now let's say that we sneak the loaded die in:
# here, we have a 95% chance to continue using the fair die and a 5%
# chance to switch to the loaded die
# when we enter the loaded die state, we have a 90% chance of staying
# in that state and a 10% chance of leaving
gen_model.transmat_ = np.array([[0.95, 0.05],
                                [0.1, 0.9]])

# now let's set the emission means:
# the first state is a fair die with equal probabilities and the
# second is loaded by being biased toward rolling a six
gen_model.emissionprob_ = \
    np.array([[1 / 6, 1 / 6, 1 / 6, 1 / 6, 1 / 6, 1 / 6],
              [1 / 10, 1 / 10, 1 / 10, 1 / 10, 1 / 10, 1 / 2]])

# simulate the loaded dice rolls
rolls, gen_states = gen_model.sample(30000)

# plot states over time, let's just look at the first rolls for clarity
fig, ax = plt.subplots()
ax.plot(gen_states[:500])
ax.set_title('States over time')
ax.set_xlabel('Time (# of rolls)')
ax.set_ylabel('State')
fig.show()

# plot rolls for the fair and loaded states
fig, ax = plt.subplots()
ax.hist(rolls[gen_states == 0], label='fair', alpha=0.5,
        bins=np.arange(7) - 0.5, density=True)
ax.hist(rolls[gen_states == 1], label='loaded', alpha=0.5,
        bins=np.arange(7) - 0.5, density=True)
ax.set_title('Roll probabilities by state')
ax.set_xlabel('Count')
ax.set_ylabel('Roll')
ax.legend()
fig.show()

MultinomialHMM has undergone major changes. The previous version was implementing CategoricalHMM (a special case of MultinomialHMM). This new implementation follows the standard definition for a Multinomial distribution, e.g. as in https://en.wikipedia.org/wiki/Multinomial_distributionSee these issues for details: https://github.com/hmmlearn/hmmlearn/issues/335 https://github.com/hmmlearn/hmmlearn/issues/340

TypeError                                 Traceback (most recent call last)
[<ipython-input-9-4c6e1e68a7c1>](https://localhost:8080/#) in <module>()
     22 
     23 # simulate the loaded dice rolls
---> 24 rolls, gen_states = gen_model.sample(30000)
     25 
     26 # plot states over time, let's just look at the first rolls for clarity

3 frames
[/root/.local/lib/python3.7/site-packages/hmmlearn/base.py](https://localhost:8080/#) in sample(self, n_samples, random_state, currstate)
    461         state_sequence = [currstate]
    462         X = [self._generate_sample_from_state(
--> 463             currstate, random_state=random_state)]
    464 
    465         for t in range(n_samples - 1):

[/root/.local/lib/python3.7/site-packages/hmmlearn/hmm.py](https://localhost:8080/#) in _generate_sample_from_state(self, state, random_state)
    481         sample = multinomial.rvs(
    482             n=self.n_trials, p=self.emissionprob_[state, :],
--> 483             size=1, random_state=self.random_state)
    484         return sample.squeeze(0)  # shape (1, nf) -> (nf,)
    485 

[/usr/local/lib/python3.7/dist-packages/scipy/stats/_multivariate.py](https://localhost:8080/#) in rvs(self, n, p, size, random_state)
   3216         %(_doc_callparams_note)s
   3217         """
-> 3218         n, p, npcond = self._process_parameters(n, p)
   3219         random_state = self._get_random_state(random_state)
   3220         return random_state.multinomial(n, p, size)

[/usr/local/lib/python3.7/dist-packages/scipy/stats/_multivariate.py](https://localhost:8080/#) in _process_parameters(self, n, p)
   3016         pcond |= np.any(p > 1, axis=-1)
   3017 
-> 3018         n = np.array(n, dtype=np.int, copy=True)
   3019 
   3020         # true for bad n

TypeError: int() argument must be a string, a bytes-like object or a number, not 'NoneType'

I am requesting a new feature

This is a feature request to track the question: When computing log of probabilities in _src/hmmc.cpp, what floating point errors do we want to clear for?

History:

1. In lib/hmmlearn/utils.py:log_mask_zero, we ignore errors related to divided by zero.
2. A pull request to move this to C++ copied this behavior.

However, we may want to copy this code from numpy and clear more errors.