-
Notifications
You must be signed in to change notification settings - Fork 0
/
my_model_selectors.py
198 lines (165 loc) · 7.96 KB
/
my_model_selectors.py
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
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
import math
import statistics
import warnings
import numpy as np
from hmmlearn.hmm import GaussianHMM
from sklearn.model_selection import KFold
from asl_utils import combine_sequences
class ModelSelector(object):
'''
base class for model selection (strategy design pattern)
'''
def __init__(self, all_word_sequences: dict, all_word_Xlengths: dict, this_word: str,
n_constant=3,
min_n_components=2, max_n_components=10,
random_state=14, verbose=False):
self.words = all_word_sequences
self.hwords = all_word_Xlengths
self.sequences = all_word_sequences[this_word]
self.X, self.lengths = all_word_Xlengths[this_word]
self.this_word = this_word
self.n_constant = n_constant
self.min_n_components = min_n_components
self.max_n_components = max_n_components
self.random_state = random_state
self.verbose = verbose
def select(self):
raise NotImplementedError
def base_model(self, num_states):
# with warnings.catch_warnings():
warnings.filterwarnings("ignore", category=DeprecationWarning)
warnings.filterwarnings("ignore", category=RuntimeWarning)
try:
hmm_model = GaussianHMM(n_components=num_states, covariance_type="diag", n_iter=1000,
random_state=self.random_state, verbose=False).fit(self.X, self.lengths)
if self.verbose:
print("model created for {} with {} states".format(self.this_word, num_states))
return hmm_model
except:
if self.verbose:
print("failure on {} with {} states".format(self.this_word, num_states))
return None
class SelectorConstant(ModelSelector):
""" select the model with value self.n_constant
"""
def select(self):
""" select based on n_constant value
:return: GaussianHMM object
"""
best_num_components = self.n_constant
return self.base_model(best_num_components)
class SelectorBIC(ModelSelector):
""" select the model with the lowest Bayesian Information Criterion(BIC) score
http://www2.imm.dtu.dk/courses/02433/doc/ch6_slides.pdf
Bayesian information criteria: BIC = -2 * logL + p * logN
"""
def select(self):
""" select the best model for self.this_word based on
BIC score for n between self.min_n_components and self.max_n_components
:return: GaussianHMM object
"""
warnings.filterwarnings("ignore", category=DeprecationWarning)
warnings.filterwarnings("ignore", category=RuntimeWarning)
minBIC = np.inf
best_model = None
for i in range(self.min_n_components, self.max_n_components + 1):
try:
test_model = GaussianHMM(n_components=i, covariance_type="diag", n_iter=1000,
random_state=self.random_state, verbose=False).fit(self.X, self.lengths)
logL = test_model.score(self.X, self.lengths)
logN = math.log(len(self.X))
# Really no idea of how to calculate parameters, got help from forums:
# https://discussions.udacity.com/t/verifing-bic-calculation/246165
# but it doesn't mean I understand it... it just works.
p = i * i + 2 * i * len(self.X[0]) - 1
# Calculate BIC = -2 * logL + p * logN
# where L is the likelihood of the fitted model, p is the number of parameters, and N
# is the number of data points.
test_bic = -2 * logL + p * logN
if test_bic < minBIC:
minBIC = test_bic
best_model = test_model
except:
pass
return best_model
class SelectorDIC(ModelSelector):
''' select best model based on Discriminative Information Criterion
Biem, Alain. "A model selection criterion for classification: Application to hmm topology optimization."
Document Analysis and Recognition, 2003. Proceedings. Seventh International Conference on. IEEE, 2003.
http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.58.6208&rep=rep1&type=pdf
DIC = log(P(X(i)) - 1/(M-1)SUM(log(P(X(all but i))
'''
def select(self):
warnings.filterwarnings("ignore", category=DeprecationWarning)
warnings.filterwarnings("ignore", category=RuntimeWarning)
# Reimplemented following the comments from the reviewer
# This is the pseudocode:
# best_dic <- float("-Inf") # Initialize best_model and best_dic
# best_model <- None
# for each value of n_components:
# model <- self.base_model(n)
# logL <- model.score(self.X, self.lengths)
# penalty <- np.mean( [ model.score(self.hwords[word]) for word in self.words if word != self.this_word ] )
# dic <- logL - penalty
# if dic > best_dic:
# best_model <- model
# best_dic <- dic
# return best_model
maxDIC = -np.inf
best_model = None
for i in range(self.min_n_components, self.max_n_components + 1):
try:
penalty = []
model = self.base_model(i)
logL = model.score(self.X,self.lengths)
for word in self.words:
if word != self.this_word:
penalty.append(model.score(self.hwords[word]))
penaltyAvg = np.mean(penalty)
dic = logL - penaltyAvg
if dic > maxDIC:
maxDIC = dic
best_model = model
except:
pass
return best_model
class SelectorCV(ModelSelector):
''' select best model based on average log Likelihood of cross-validation folds
'''
def select(self):
warnings.filterwarnings("ignore", category=DeprecationWarning)
warnings.filterwarnings("ignore", category=RuntimeWarning)
# FISH was returning wrong values, number of splits seems cannot be more than 2, so got help here:
# https://discussions.udacity.com/t/issue-with-selectorcv/299868
# but this solution doesn't work on part 3.
# n_splits = min(len(self.lengths), 3)
n_splits = 2
split_method = KFold(n_splits)
final_score = -np.inf
best_model = None
# For each number of n_components...
for i in range(self.min_n_components, self.max_n_components + 1):
try:
# We do a cross validation
sum_score = 0
denominator = 0
average = 0
for cv_train_idx, cv_test_idx in split_method.split(self.sequences):
# Train and test dataset split
self.train_X, self.train_lengths = combine_sequences(cv_train_idx,self.sequences)
self.test_X, self.test_lengths = combine_sequences(cv_test_idx,self.sequences)
test_model = GaussianHMM(n_components=i, covariance_type="diag", n_iter=1000,
random_state=self.random_state, verbose=False).fit(self.train_X, self.train_lengths)
# Get the score and calculate the average
denominator += 1
sum_score = sum_score + test_model.score(self.test_X, self.test_lengths)
average = sum_score / denominator
# If average is better, this model is also better than the previous ones
if average > final_score:
final_score = average
# Return a model trained with all the data
best_model = GaussianHMM(n_components=i, covariance_type="diag", n_iter=1000,
random_state=self.random_state, verbose=False).fit(self.X, self.lengths)
except:
pass
return best_model