Class HiddenMarkovModel
HiddenMarkovModel implements hidden Markov models with Gaussian mixtures as distributions on top of TensorFlow 2.0
Installation
pip install --upgrade git+https://gitlab.com/kesmarag/hmm-gmm-tf2
HiddenMarkovModel(p0, tp, em_w, em_mu, em_var)
Args:
p0: 1D numpy array
Determines the probability of the first hidden variable
in the Markov chain for each hidden state.
e.g. np.array([0.5, 0.25, 0.25]) (3 hidden states)
tp: 2D numpy array
Determines the transition probabilities for moving from one hidden state to each
other. The (i,j) element of the matrix denotes the probability of
transiting from i-th state to the j-th state.
e.g. np.array([[0.80, 0.15, 0.05],
[0.20, 0.55, 0.25],
[0.15, 0.15, 0.70]])
(3 hidden states)
em_w: 2D numpy array
Contains the weights of the Gaussian mixtures.
Each line correspond to a hidden state.
e.g. np.array([[0.8, 0.2],
[0.5, 0.5],
[0.1, 0.9]])
(3 hidden states, 2 Gaussian mixtures)
em_mu: 3D numpy array
Determines the mean value vector for each component
of the emission distributions.
The first dimension refers to the hidden states whereas the
second one refer to the mixtures.
e.g. np.array([[[2.2, 1.3], [1.2, 0.2]], 1st hidden state
[[1.3, 5.0], [4.3, -2.3]], 2nd hidden state
[[0.0, 1.2], [0.4, -2.0]]]) 3rd hidden state
(3 hidden states, 2 Gaussian mixtures)
em_var: 3D numpy array
Determines the variance vector for each component of the
emission distributions.
e.g. np.array([[[2.2, 1.3], [1.2, 0.2]], 1st hidden state
[[1.3, 5.0], [4.3, -2.3]], 2nd hidden state
[[0.0, 1.2], [0.4, -2.0]]]) 3rd hidden state
(3 hidden states, 2 Gaussian mixtures)
log_posterior
HiddenMarkovModel.log_posterior(self, data)
Log probability density function.
Args:
data: 3D numpy array
The first dimension refers to each component of the batch.
The second dimension refers to each specific time interval.
The third dimension refers to the values of the observed data.
Returns:
1D numpy array with the values of the log-probability function with respect to the observations.
viterbi_algorithm
HiddenMarkovModel.viterbi_algorithm(self, data)
Performs the viterbi algorithm for calculating the most probable
hidden state path of some batch data.
Args:
data: 3D numpy array
The first dimension refers to each component of the batch.
The second dimension refers to each specific time interval.
The third dimension refers to the values of the observed data.
Returns:
2D numpy array with the most probable hidden state paths.
The first dimension refers to each component of the batch.
The second dimension the order of the hidden states.
(0, 1, ..., K-1), where K is the total number of hidden states.
fit
HiddenMarkovModel.fit(self, data, max_iter=100, min_var=0.01, verbose=False)
This method re-adapts the model parameters with respect to a batch of
observations, using the Expectation-Maximization (E-M) algorithm.
Args:
data: 3D numpy array
The first dimension refers to each component of the batch.
The second dimension refers to each specific time step.
The third dimension refers to the values of the observed data.
max_iter: positive integer number
The maximum number of iterations.
min_var: non-negative real value
The minimum acceptance variance. We use this restriction
in order to prevent overfitting of the model.
Returns:
1D numpy array with the log-posterior probability densities for each training iteration.
generate
HiddenMarkovModel.generate(self, length, num_series=1, p=0.2)
Generates a batch of time series using an importance sampling like approach.
Args:
length: positive integer
The length of each time series.
num_series: positive integer (default 1)
The number of the time series.
p: real value between 0.0 and 1.0 (default 0.2)
The importance sampling parameter.
At each iteration:
k[A] Draw X and calculate p(X)
if p(X) > p(X_{q-1}) then
accept X as X_q
else
draw r from [0,1] using the uniform distribution.
if r > p then
accept the best of the rejected ones.
else
go to [A]
Returns:
3D numpy array with the drawn time series.
2D numpy array with the corresponding hidden states.
kl_divergence
HiddenMarkovModel.kl_divergence(self, other, data)
Estimates the value of the Kullback-Leibler divergence (KLD)
between the model and another model with respect to some data.
Example
import numpy as np
from kesmarag.hmm import HiddenMarkovModel, new_left_to_right_hmm, store_hmm, restore_hmm, toy_example
dataset = toy_example()
This helper function creates a test dataset with a single two dimensional time series with 700 samples.
The first 200 samples corresponds to a Gaussian mixture with
w1 = 0.6, w2=0.4
mu1 = [0.5, 1], mu2 = [2, 1]
var1 = [1, 1], var2=[1.2, 1]
the next 300 corresponds to a Gaussian mixture with
w1 = 0.6, w2=0.4
mu1 = [2, 5], mu2 = [4, 5]
var1 = [0.8, 1], var2=[0.8, 1]
and the last 200 corresponds to a Gaussian mixture with
w1 = 0.6, w2=0.4
mu1 = [4, 1], mu2 = [6, 5]
var1 = [1, 1], var2=[0.8, 1.2]
print(dataset.shape)
(1, 700, 2)
model = new_left_to_right_hmm(states=3, mixtures=2, data=dataset)
model.fit(dataset, verbose=True)
epoch: 0 , ln[p(X|λ)] = -3094.3748904062295
epoch: 1 , ln[p(X|λ)] = -2391.3602228316568
epoch: 2 , ln[p(X|λ)] = -2320.1563724302564
epoch: 3 , ln[p(X|λ)] = -2284.996645965759
epoch: 4 , ln[p(X|λ)] = -2269.0055909790053
epoch: 5 , ln[p(X|λ)] = -2266.1395773469876
epoch: 6 , ln[p(X|λ)] = -2264.4267494952455
epoch: 7 , ln[p(X|λ)] = -2263.156612481979
epoch: 8 , ln[p(X|λ)] = -2262.2725752851293
epoch: 9 , ln[p(X|λ)] = -2261.612564557431
epoch: 10 , ln[p(X|λ)] = -2261.102826808333
epoch: 11 , ln[p(X|λ)] = -2260.7189908960695
epoch: 12 , ln[p(X|λ)] = -2260.437608729253
epoch: 13 , ln[p(X|λ)] = -2260.231860238426
epoch: 14 , ln[p(X|λ)] = -2260.0784163526014
epoch: 15 , ln[p(X|λ)] = -2259.960659542152
epoch: 16 , ln[p(X|λ)] = -2259.8679640963023
epoch: 17 , ln[p(X|λ)] = -2259.793721328861
epoch: 18 , ln[p(X|λ)] = -2259.733658260372
epoch: 19 , ln[p(X|λ)] = -2259.684791553708
epoch: 20 , ln[p(X|λ)] = -2259.6448728507144
epoch: 21 , ln[p(X|λ)] = -2259.6121181368353
epoch: 22 , ln[p(X|λ)] = -2259.5850765029527
[-3094.3748904062295,
-2391.3602228316568,
-2320.1563724302564,
-2284.996645965759,
-2269.0055909790053,
-2266.1395773469876,
-2264.4267494952455,
-2263.156612481979,
-2262.2725752851293,
-2261.612564557431,
-2261.102826808333,
-2260.7189908960695,
-2260.437608729253,
-2260.231860238426,
-2260.0784163526014,
-2259.960659542152,
-2259.8679640963023,
-2259.793721328861,
-2259.733658260372,
-2259.684791553708,
-2259.6448728507144,
-2259.6121181368353,
-2259.5850765029527]
print(model)
### [kesmarag.hmm.HiddenMarkovModel] ###
=== Prior probabilities ================
[1. 0. 0.]
=== Transition probabilities ===========
[[0.995 0.005 0. ]
[0. 0.996666 0.003334]
[0. 0. 1. ]]
=== Emission distributions =============
*** Hidden state #1 ***
--- Mixture #1 ---
weight : 0.779990073797613
mean_values : [0.553266 1.155844]
variances : [1.000249 0.967666]
--- Mixture #2 ---
weight : 0.22000992620238702
mean_values : [2.598735 0.633391]
variances : [1.234133 0.916872]
*** Hidden state #2 ***
--- Mixture #1 ---
weight : 0.5188217626642593
mean_values : [2.514082 5.076246]
variances : [1.211327 0.903328]
--- Mixture #2 ---
weight : 0.4811782373357407
mean_values : [3.080913 5.039015]
variances : [1.327171 1.152902]
*** Hidden state #3 ***
--- Mixture #1 ---
weight : 0.5700082256217439
mean_values : [4.03977 1.118112]
variances : [0.97422 1.00621]
--- Mixture #2 ---
weight : 0.429991774378256
mean_values : [6.162698 5.064422]
variances : [0.753987 1.278449]
store_hmm(model, 'test_model.npz')
load_model = restore_hmm('test_model.npz')
gen_data = model.generate(700, 10, 0.05)
0 -2129.992044055025
1 -2316.443344656749
2 -2252.206072731434
3 -2219.667047368621
4 -2206.6760352374367
5 -2190.952289092368
6 -2180.0268345326112
7 -2353.7153702977475
8 -2327.955163192414
9 -2227.4471755146196
print(gen_data)
(array([[[-0.158655, 0.117973],
[ 4.638243, 0.249049],
[ 0.160007, 1.079808],
...,
[ 4.671152, 4.18109 ],
[ 2.121958, 3.747366],
[ 2.572435, 6.352445]],
[[-0.158655, 0.117973],
[-1.379849, 0.998761],
[-0.209945, 0.947926],
...,
[ 3.93909 , 1.383347],
[ 5.356786, 1.57808 ],
[ 5.0488 , 5.586755]],
[[-0.158655, 0.117973],
[ 1.334 , 0.979797],
[ 3.708721, 1.321735],
...,
[ 3.819756, 0.78794 ],
[ 6.53362 , 4.177215],
[ 7.410012, 6.30113 ]],
...,
[[-0.158655, 0.117973],
[-0.152573, 0.612675],
[-0.917723, -0.632936],
...,
[ 4.110186, -0.027864],
[ 2.82694 , 0.65438 ],
[ 6.825696, 5.27543 ]],
[[-0.158655, 0.117973],
[ 3.141896, 0.560984],
[ 2.552211, -0.223568],
...,
[ 4.41791 , -0.430231],
[ 2.525892, -0.64211 ],
[ 5.52568 , 6.313566]],
[[-0.158655, 0.117973],
[ 0.845694, 2.436781],
[ 1.564802, -0.652546],
...,
[ 2.33009 , 0.932121],
[ 7.095326, 6.339674],
[ 3.748988, 2.25159 ]]]), array([[0., 0., 0., ..., 1., 1., 1.],
[0., 0., 0., ..., 2., 2., 2.],
[0., 0., 0., ..., 2., 2., 2.],
...,
[0., 0., 0., ..., 2., 2., 2.],
[0., 0., 0., ..., 2., 2., 2.],
[0., 0., 0., ..., 2., 2., 2.]]))