1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
|
# benchmark_hmm.py
# Contact: Jacob Schreiber ( jmschreiber91@gmail.com )
"""
Benchmark the HMM module, including multithreading.
"""
from pomegranate import *
import numpy as np
import time
import random
np.random.seed(0)
random.seed(0)
def global_alignment( match_distributions, insert_distribution ):
"""Create a global alignment HMM from match distributions.
"""
model = HiddenMarkovModel()
i0 = State( insert_distribution, name="i0" )
model.add_state(i0)
model.add_transition( i0, i0, 0.3 )
model.add_transition( model.start, i0, 0.3 )
last_match, last_insert, last_delete = model.start, i0, None
for i, distribution in enumerate( match_distributions ):
match = State( distribution, name="m{}".format(i+1) )
insert = State( insert_distribution, name="i{}".format(i+1) )
delete = State( None, name="d{}".format(i+1) )
model.add_states([match, insert, delete])
model.add_transition( last_match, match, 0.5 )
model.add_transition( last_match, delete, 0.1 )
model.add_transition( last_insert, match, 0.5 )
model.add_transition( last_insert, delete, 0.2 )
if last_delete is not None:
model.add_transition( last_delete, match, 0.7 )
model.add_transition( last_delete, delete, 0.2 )
model.add_transition( insert, insert, 0.3 )
model.add_transition( match, insert, 0.1 )
model.add_transition( delete, insert, 0.1 )
last_match, last_insert, last_delete = match, insert, delete
model.add_transition( last_match, model.end, 0.6 )
model.add_transition( last_insert, model.end, 0.7 )
model.add_transition( last_delete, model.end, 0.9 )
model.bake()
return model
def benchmark_forward( model, sample ):
tic = time.time()
for i in range(25000):
logp = model.forward( sample )[-1, model.end_index]
print("{:16}: time: {:5.5}, logp: {:5.5}".format( "FORWARD", time.time() - tic, logp ))
def benchmark_backward( model, sample ):
tic = time.time()
for i in range(25000):
logp = model.backward( sample )[0, model.start_index]
print("{:16}: time: {:5.5}, logp: {:5.5}".format( "BACKWARD", time.time() - tic, logp ))
def benchmark_forward_backward( model, sample ):
tic = time.time()
for i in range(25000):
model.forward_backward( sample )
print("{:16}: time: {:5.5}".format( "FORWARD-BACKWARD", time.time() - tic ))
def benchmark_viterbi( model, sample ):
tic = time.time()
for i in range(25000):
logp, path = model.viterbi( sample )
print("{:16}: time: {:5.5}, logp: {:5.5}".format( "VITERBI", time.time() - tic, logp ))
def benchmark_training( model, samples, n_jobs ):
tic = time.time()
improvement = model.fit( samples, max_iterations=10, verbose=False, n_jobs=n_jobs , return_history=True)
print("{:16}: time: {:5.5}, improvement: {:5.5} ({} jobs)".format( "BW TRAINING", time.time() - tic, improvement[1].total_improvement[0], n_jobs ))
def main():
n = 15
print("HIDDEN MARKOV MODEL BENCHMARKS")
print("gaussian emissions")
sigma = 3
means = np.random.randn(n)*20
gaussian_dists = [ NormalDistribution(mean, 1) for mean in means ]
gaussian_model = global_alignment( gaussian_dists, NormalDistribution(0, 10) )
gaussian_sample = np.random.randn(n)*sigma + means
gaussian_batch = np.random.randn(50, n)*sigma + means
benchmark_forward( gaussian_model, gaussian_sample )
benchmark_backward( gaussian_model, gaussian_sample )
benchmark_viterbi( gaussian_model, gaussian_sample )
benchmark_forward_backward( gaussian_model, gaussian_sample )
benchmark_training( gaussian_model, gaussian_batch, 1 )
# Reset the model
gaussian_dists = [ NormalDistribution(mean, 1) for mean in means ]
gaussian_model = global_alignment( gaussian_dists, NormalDistribution(0, 10) )
benchmark_training( gaussian_model, gaussian_batch, 4 )
print
print("multivariate gaussian emissions")
m = 10
means = np.random.randn(n, m) * 5 + np.arange(m) * 3
mgd = MultivariateGaussianDistribution
multivariate_gaussian_dists = [ mgd( mean, np.eye(m) ) for mean in means ]
multivariate_gaussian_model = global_alignment( multivariate_gaussian_dists, mgd( np.zeros(m), np.eye(m)*3 ) )
multivariate_gaussian_sample = np.random.randn(n, m)*sigma + means
multivariate_gaussian_batch = np.random.randn(500, n, m)*sigma + means
benchmark_forward( multivariate_gaussian_model, multivariate_gaussian_sample )
benchmark_backward( multivariate_gaussian_model, multivariate_gaussian_sample )
benchmark_viterbi( multivariate_gaussian_model, multivariate_gaussian_sample )
benchmark_forward_backward( multivariate_gaussian_model, multivariate_gaussian_sample )
benchmark_training( multivariate_gaussian_model, multivariate_gaussian_batch, 1 )
# Reset the model
multivariate_gaussian_dists = [ mgd( mean, np.eye(m) ) for mean in means ]
multivariate_gaussian_model = global_alignment( multivariate_gaussian_dists, mgd( np.zeros(m), np.eye(m)*3 ) )
benchmark_training( multivariate_gaussian_model, multivariate_gaussian_batch, 4 )
print
print("discrete distribution")
probs = np.abs( np.random.randn(n, 4) * 20 ).T
probs = (probs / probs.sum(axis=0)).T
discrete_dists = [ DiscreteDistribution({ char: prob for char, prob in zip('ACGT', row)}) for row in probs ]
discrete_model = global_alignment( discrete_dists, DiscreteDistribution({'A': 0.25, 'C': 0.25, 'G': 0.25, 'T': 0.25}))
discrete_sample = list('ACGTAGCTACGACATCAGAC')
discrete_batch = np.array([ np.random.choice(list('ACGT'), size=200, p=row) for row in probs ]).T
benchmark_forward( discrete_model, discrete_sample )
benchmark_backward( discrete_model, discrete_sample )
benchmark_viterbi( discrete_model, discrete_sample )
benchmark_forward_backward( discrete_model, discrete_sample )
benchmark_training( discrete_model, discrete_batch, 1 )
# Reset the model
discrete_dists = [ DiscreteDistribution({ char: prob for char, prob in zip('ACGT', row)}) for row in probs ]
discrete_model = global_alignment( discrete_dists, DiscreteDistribution({'A': 0.25, 'C': 0.25, 'G': 0.25, 'T': 0.25}))
benchmark_training( discrete_model, discrete_batch, 4 )
if __name__ == '__main__':
main()
|
