forked from vaibhavgrover/Opinion-Target-Extraction
-
Notifications
You must be signed in to change notification settings - Fork 0
/
main.py
executable file
·195 lines (157 loc) · 7.84 KB
/
main.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
import numpy
import time
import sys
import subprocess
import os
import random
import json
from rnn import elman, jordan, elman_M1, elman_M2,elman_M3,elman_GRU
from metrics.accuracy import conlleval
from utils.tools import shuffle, minibatch, contextwin
if __name__ == '__main__':
data_path = sys.argv[1]
rnn_type = sys.argv[2]
model_folder = sys.argv[3]
window = int(sys.argv[4])
nhidden = int(sys.argv[5])
dimension = int(sys.argv[6])
initialize = False
if (sys.argv[7].lower() == 'true'):
initialize = True
s = {'lr':0.01,
'verbose':1,
'decay':False, # decay on the learning rate if improvement stops
'win':window, # number of words in the context window
'bs':6, # number of backprop through time steps
'nhidden':nhidden, # number of hidden units
'seed':1234567890,
'emb_dimension':dimension, # dimension of word embedding
'nepochs':30}
# load the dataset
dataset = json.load(open(data_path))
develop_set, valid_set, test_set = dataset["development"], dataset["validate"], dataset["test"]
word2idx, label2idx = dataset["word2Idx"], dataset["label2Idx"]
idx2label = dict((k, v) for v, k in label2idx.iteritems())
idx2word = dict((k, v) for v, k in word2idx.iteritems())
# xIndexes=indexes corresponding to each, xFeatures= length 14 features corresponding to each word
train_lex, train_y, train_feat = develop_set["xIndexes"], develop_set["yLabels"], develop_set["xFeatures"]
valid_lex, valid_y, valid_feat = valid_set["xIndexes"], valid_set["yLabels"], valid_set["xFeatures"]
test_lex, test_y, test_feat = test_set["xIndexes"], test_set["yLabels"], test_set["xFeatures"]
assert len(train_feat) == len(train_lex) == len(train_y)
assert len(test_feat) == len(test_lex) == len(test_y)
assert len(valid_feat) == len(valid_lex) == len(valid_y)
voc_size = len(set(reduce(lambda x, y: list(x) + list(y), train_lex + valid_lex + test_lex)))
num_classes = len(set(reduce(lambda x, y: list(x) + list(y), train_y + valid_y + test_y)))
num_sentences = len(train_lex)
# instantiate the model
numpy.random.seed(s['seed'])
random.seed(s['seed'])
if rnn_type == "elman":
rnn = elman.model(nh=s['nhidden'],
nc=num_classes,
ne=voc_size,
de=s['emb_dimension'],
cs=s['win'],
em=dataset["embeddings"],
init=initialize,
featdim=14)
elif rnn_type == "jordan":
rnn = jordan.model(nh=s['nhidden'],
nc=num_classes,
ne=voc_size,
de=s['emb_dimension'],
cs=s['win'],
em=dataset["embeddings"],
init=initialize)
elif rnn_type == "elman_M1":
rnn = elman_M1.model(nh=s['nhidden'],
nc=num_classes,
ne=voc_size,
de=s['emb_dimension'],
cs=s['win'],
em=dataset["embeddings"],
init=initialize,
featdim=14)
elif rnn_type == "elman_M2":
rnn = elman_M2.model(nh=s['nhidden'],
nc=num_classes,
ne=voc_size,
de=s['emb_dimension'],
cs=s['win'],
em=dataset["embeddings"],
init=initialize,
featdim=14)
elif rnn_type == "elman_M3":
rnn = elman_M3.model(nh=s['nhidden'],
nc=num_classes,
ne=voc_size,
de=s['emb_dimension'],
cs=s['win'],
em=dataset["embeddings"],
init=initialize,
featdim=14)
elif rnn_type == "elman_GRU":
rnn = elman_GRU.model(nh=s['nhidden'],
nc=num_classes,
ne=voc_size,
de=s['emb_dimension'],
cs=s['win'],
em=dataset["embeddings"],
init=initialize,
featdim=14)
else:
print "Invalid RNN type: ", rnn_type
sys.exit(-1)
# create a folder for store the models
if not os.path.exists(model_folder): os.mkdir(model_folder)
# train with early stopping on validation set
best_f1_test, best_f1_test_val = -numpy.inf, -numpy.inf
s['clr'] = s['lr'] # learning rate
for e in xrange(s['nepochs']):
# shuffle
shuffle([train_lex, train_y, train_feat], s['seed'])
s['ce'] = e
tic = time.time()
for i in xrange(num_sentences):
context_words = contextwin(train_lex[i], s['win']) #list of list of indexes corresponding to context windows surrounding each word in the sentence
words = map(lambda x: numpy.asarray(x).astype('int32'), minibatch(context_words, s['bs']))
features = minibatch(train_feat[i], s['bs'])
labels = train_y[i]
for word_batch, feature_batch, label_last_word in zip(words, features, labels):
rnn.train(word_batch, feature_batch, label_last_word, s['clr'])
rnn.normalize()
if s['verbose']:
print '[learning] epoch %i >> %2.2f%%' % (e, (i+1)*100./num_sentences),\
'completed in %.2f (sec) <<\r' % (time.time()-tic),
sys.stdout.flush()
# evaluation // back into the real world : idx -> words
predictions_test = [map(lambda x: idx2label[x],
rnn.classify(numpy.asarray(contextwin(x, s['win'])).astype('int32'), f) ) for x, f in zip (test_lex, test_feat)]
ground_truth_test = [map(lambda x: idx2label[x], y) for y in test_y]
words_test = [ map(lambda x: idx2word[x], w) for w in test_lex]
predictions_valid = [map(lambda x: idx2label[x],
rnn.classify(numpy.asarray(contextwin(x, s['win'])).astype('int32'), f)) for x, f in zip (valid_lex, valid_feat)]
ground_truth_valid = [map(lambda x: idx2label[x], y) for y in valid_y]
words_valid = [map(lambda x: idx2word[x], w) for w in valid_lex]
# evaluation // compute the accuracy using conlleval.pl
res_test = conlleval(predictions_test, ground_truth_test, words_test, model_folder + '/current.test.txt')
res_valid = conlleval(predictions_valid, ground_truth_valid, words_valid, model_folder + '/current.valid.txt')
if res_test['f1'] > best_f1_test:
rnn.save(model_folder)
best_f1_test, best_f1_test_val = res_test['f1'], res_valid['f1']
if s['verbose']:
print 'NEW BEST: epoch', e, 'valid F1', res_valid['f1'], 'best test F1', res_test['f1'], ' '*20
s['vf1'], s['vp'], s['vr'] = res_valid['f1'], res_valid['p'], res_valid['r']
s['tf1'], s['tp'], s['tr'] = res_test['f1'], res_test['p'], res_test['r']
s['be'] = e
subprocess.call(['mv', model_folder + '/current.test.txt', model_folder + '/best.test.txt'])
subprocess.call(['mv', model_folder + '/current.valid.txt', model_folder + '/best.valid.txt'])
else:
print ''
# learning rate decay if no improvement in 10 epochs
if s['decay'] and abs(s['be']-s['ce']) >= 10:
s['clr'] *= 0.5
if s['clr'] < 1e-5:
break
print 'BEST RESULT: epoch', s['be'], 'valid F1:', best_f1_test_val, 'best test precision: ', \
s['tp'], 'best test recall: ', s['tr'], 'best test F1', best_f1_test, 'with the model', model_folder