-
Notifications
You must be signed in to change notification settings - Fork 0
/
my_model_selectors.py
198 lines (168 loc) · 8.71 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
"""
best_bic = float('Inf') # to start the selector
best_model = None # just in case we don't find any
# for each num_component to test
for num_components in range(self.min_n_components, self.max_n_components+1):
try:
# we train the model and get its log-likelihood
current_model = self.base_model(num_components)
logL = current_model.score(self.X, self.lengths)
# number of parameters according to https://discussions.udacity.com/t/number-of-parameters-bic-calculation/233235/15
bic = -2 * logL + (num_components ** 2 + 2 * num_components * current_model.n_features - 1 ) * np.log(len(self.sequences))
# shall it be better than the best_bic yet, it becomes the best_bic and we select this model as the best_model
if bic < best_bic:
best_bic = bic
best_model = current_model
except:
# copied from the function above (base_model)
if self.verbose:
print("failure on {} with {} states, continuing".format(self.this_word, num_components))
pass
# we return the best_model
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):
best_dic = float('-Inf') # to start the selector
best_model = None # just in case we find no model
other_words = [value for key, value in self.hwords.items() if key != self.this_word] #
for num_components in range(self.min_n_components, self.max_n_components+1):
try:
# we train for the current word and we get the logL
current_model = self.base_model(num_components)
Xi_logL = current_model.score(self.X, self.lengths)
sum_other_Xi_logL = float(0)
# we score every other class, calculating logL for competing words
for word in other_words:
sum_other_Xi_logL += current_model.score(word[0], word[1])
dic = Xi_logL - (1/len(other_words))*sum_other_Xi_logL
# according to the paper, if the model presents a greater criterion value, it's a better model
# shall it be better than the best_dic yet, it becomes the best_dic and we select this model as the best_model
if dic > best_dic:
best_dic = dic
best_model = current_model
except:
# copied from the function above (base_model)
if self.verbose:
print("failure on {} with {} states, continuing".format(self.this_word, num_components))
continue
# we return the best_model
return best_model
class SelectorCV(ModelSelector):
''' select best model based on average log Likelihood of cross-validation folds
'''
def select(self):
best_logLavg = float('-Inf') # to start the selector
best_model = None # just in case we don't find a model
best_num_components = None # just in case we don't find a model
def cv_loop(num_components):
""" CV loop helper function """
logLs = []
# I thought I needed to do something like this (as it was failing for FISH) but I confirmed it using the forums: https://discussions.udacity.com/t/selectorcv-fails-to-train-fish/338796
split_method = KFold(n_splits=min(3,len(self.sequences)))
# for each fold
for cv_train_idx, cv_test_idx in split_method.split(self.sequences):
try:
# we get X and lengths for both train and test set
X_train, lengths_train = combine_sequences(cv_train_idx, self.sequences)
X_test, lengths_test = combine_sequences(cv_test_idx, self.sequences)
# we train the model
current_model = GaussianHMM(n_components=num_components, covariance_type="diag", n_iter=1000,
random_state=self.random_state, verbose=False).fit(X_train, lengths_train)
# and we append the logL to our list
logLs.append(current_model.score(X_test, lengths_test))
except:
# copied from the function above (base_model)
if self.verbose:
print("failure on {} with {} states, continuing".format(self.this_word, num_components))
continue
# if we found at least one logL we return the average
if len(logLs) > 0:
return (sum(logLs)/len(logLs))
else:
return float('-Inf')
for num_components in range(self.min_n_components, self.max_n_components+1):
if len(self.sequences) > 1:
# in case CV is possible (>1 sequences) we do the cv loop
logLavg = cv_loop(num_components)
else:
# if <1 sequences, we train using all the data (no cv possible)
logLavg = float('-Inf')
try:
current_model = self.base_model(num_components)
logLavg = current_model.score(self.X, self.lengths)
except:
pass
# we compare the current logLavg with the best one yet, if it's better, we assign it as the best logLavg and set
# the number of components as the best yet
if logLavg > best_logLavg:
best_logLavg = logLavg
best_num_components = num_components
# if we found the best number of components, we create a new model
if best_num_components is not None:
best_model = self.base_model(best_num_components)
return best_model