-
Notifications
You must be signed in to change notification settings - Fork 0
/
maximum_entropy.py
170 lines (118 loc) · 5.96 KB
/
maximum_entropy.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
""" Maximum entropy model for Assignment 1: Starter code.
You can change this code however you like. This is just for inspiration.
"""
import propername
import util
import sys
import csv
import numpy as np
from collections import defaultdict
from scipy.optimize import fmin_l_bfgs_b
class MaximumEntropyModel():
""" Maximum entropy model for classification.
Attributes:
"""
def __init__(self, batch_size=None, pgtol=0.005, w_scale=0.1):
self.batch_size = batch_size #mini-batch size for gradient descent
self.pgtol = pgtol #early stopping criteria for training
self.w_scale = w_scale #factor to scale initial random normal w values
def objective_function(self, w):
dot_products = []
for i in range(self.num_labels):
w_subset = w[i*self.num_features:(i+1)*self.num_features]
dot_products.append((w_subset*self.train_features).sum(axis=1))
dot_products = np.array(dot_products).T
numerators = np.exp(dot_products[np.arange(self.n), self.train_labels])
denominators = np.sum(np.exp(dot_products), axis=1)
correct_probs = numerators / denominators
print(-np.sum(np.log(correct_probs)))
return -np.sum(np.log(correct_probs))
def gradient_function(self, w):
temp_train_features = self.train_features
temp_phi = self.phi
temp_n = self.n
# adjust the data we're working with if mini-batch size is set
if self.batch_size:
random_rows = np.random.choice(self.n, self.batch_size, replace=False)
temp_train_features = temp_train_features[random_rows, :]
temp_phi = temp_phi[random_rows, :]
temp_n = self.batch_size
#first, compute the probabilities
dot_products = []
for i in range(self.num_labels):
w_subset = w[i*self.num_features:(i+1)*self.num_features]
dot_products.append((w_subset*temp_train_features).sum(axis=1))
dot_products = np.array(dot_products).T
numerators = np.exp(dot_products)
denominators = np.sum(np.exp(dot_products), axis=1)
probs = np.divide(numerators, denominators.reshape(len(denominators), 1))
probs_repeated = np.repeat(probs, self.num_features, axis=1)
expected_counts = np.multiply(probs_repeated,
np.tile(temp_train_features, self.num_labels))
return -(temp_phi - expected_counts).sum(axis=0) / temp_n
def train(self, train_features, train_labels):
self.train_features = train_features
self.train_labels = train_labels
self.n = self.train_features.shape[0]
self.num_features = self.train_features.shape[1]
self.num_labels = len(set(self.train_labels))
self.w = np.random.randn(self.num_features*self.num_labels) * self.w_scale
self.phi = np.zeros((self.n, self.num_features*self.num_labels))
for i in range(self.n):
self.phi[i, (self.train_labels[i]*self.num_features):((self.train_labels[i]+1)*self.num_features)] = self.train_features[i, :]
self.w, final_value, _ = fmin_l_bfgs_b(self.objective_function, self.w, self.gradient_function,
pgtol=self.pgtol)
print(final_value)
def predict(self, X):
dot_products = []
for i in range(self.num_labels):
w_subset = self.w[i*self.num_features:(i+1)*self.num_features]
dot_products.append((w_subset*X).sum(axis=1))
dot_products = np.array(dot_products).T
numerators = np.exp(dot_products)
denominators = np.sum(np.exp(dot_products), axis=1)
probs = np.divide(numerators, denominators.reshape(len(denominators), 1))
return np.argmax(probs, axis=1)
if __name__ == "__main__":
dataset = sys.argv[1]
output_filename = ""
dev_output_filename = ""
output_header = []
if dataset == 'propernames':
train_features, train_labels, dev_features, dev_labels, test_features = util.load_propernames()
label_encodings = util.load_labels_from_array(train_labels)
label_encodings_reverse = {label_encodings[k]:k for k in label_encodings}
train_labels = [label_encodings[x] for x in train_labels]
dev_labels = [label_encodings[x] for x in dev_labels]
output_filename = "results/maxent_propername_test_predictions.csv"
dev_output_filename = "data/propernames/dev/dev_pred_maxent.csv"
output_header = ['id', 'type']
elif dataset == 'newsgroups':
train_features, train_labels, dev_features, dev_labels, test_features = util.load_newsgroups()
label_encodings = util.load_labels_from_array(train_labels)
label_encodings_reverse = {label_encodings[k]:k for k in label_encodings}
train_labels = [label_encodings[x] for x in train_labels]
dev_labels = [label_encodings[x] for x in dev_labels]
output_filename = "results/maxent_newsgroup_test_predictions.csv"
dev_output_filename = "data/newsgroups/dev/dev_pred_maxent.csv"
output_header = ['id', 'newsgroup']
else:
print("Argument must be propernames or newsgroups")
exit()
print(train_labels)
model = MaximumEntropyModel()
model.train(train_features, train_labels)
predictions = model.predict(train_features)
predictions_dev = model.predict(dev_features)
predictions_test = model.predict(test_features)
print(np.mean(predictions_dev == dev_labels))
with open(dev_output_filename, 'w', newline='') as f:
writer = csv.writer(f)
writer.writerow(output_header)
for i in range(predictions_dev.shape[0]):
writer.writerow([i, label_encodings_reverse[predictions_dev[i]]])
with open(output_filename, 'w', newline='') as f:
writer = csv.writer(f)
writer.writerow(output_header)
for i in range(predictions_test.shape[0]):
writer.writerow([i, label_encodings_reverse[predictions_test[i]]])