-
Notifications
You must be signed in to change notification settings - Fork 2
/
2CNNs+bgrnn.py
145 lines (118 loc) · 6.24 KB
/
2CNNs+bgrnn.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
#! /usr/bin/env python
# -*- coding: utf-8 -*-
"""
@author: xiwen zhao
@created: 2016.12.2
"""
from optparse import OptionParser
from src.trainer import BaseTrainer
from src.common import data_manager
from keras.models import Sequential
from keras.layers import Dense, Dropout, Activation, Embedding, Merge
from keras.layers import LSTM, SimpleRNN, GRU
from keras.layers import Embedding
from keras.layers import Convolution1D, GlobalMaxPooling1D, MaxPooling1D
from keras.layers.wrappers import Bidirectional
from keras.optimizers import RMSprop, SGD
class Trainer(BaseTrainer):
def get_model_name(self):
return __file__.split('/')[-1].split('.')[0]
def post_prepare_X(self, x):
return [x for i in range(3)]
def set_model_config(self, options):
self.config = dict(
nb_filter_1 = options.nb_filter_1,
nb_filter_2 = options.nb_filter_2,
filter_length_1 = options.filter_length_1,
filter_length_2 = options.filter_length_2,
hidden_dims = options.hidden_dims,
dropout_W = options.dropout_W,
dropout_U = options.dropout_U,
optimizer = options.optimizer,
rnn_output_dims = options.rnn_output_dims
)
def get_optimizer(self, key_optimizer):
if key_optimizer == 'rmsprop':
return RMSprop(lr=0.001, rho=0.9, epsilon=1e-08)
else: # 'sgd'
return SGD(lr=0.01, decay=1e-6, momentum=0.9, nesterov=False)
def build_model(self, config, weights):
bgrnn_model = Sequential()
bgrnn_model.add(Embedding(config['max_features'],
config['embedding_dims'],
input_length = config['input_length'],
weights = [weights['Wemb']] if 'Wemb' in weights else None))
bgrnn_model.add(Bidirectional(GRU(config['rnn_output_dims'],
dropout_W=config['dropout_W'], dropout_U=config['dropout_U'])))
cnn_i_model = Sequential()
cnn_i_model.add(Embedding(config['max_features'],
config['embedding_dims'],
input_length = config['input_length'],
weights = [weights['Wemb']] if 'Wemb' in weights else None))
#dropout = 0.2))
cnn_i_model.add(Convolution1D(nb_filter = config['nb_filter_1'],
filter_length = config['filter_length_1'],
border_mode = 'valid',
activation = 'relu',
subsample_length = 1))
'''
cnn_i_model.add(MaxPooling1D(pool_length=6, stride=2, border_mode='valid'))
cnn_i_model.add(Convolution1D(nb_filter = config['nb_filter_2'],
filter_length = config['filter_length_2'],
border_mode = 'valid',
activation = 'relu',
subsample_length = 1))
'''
cnn_i_model.add(GlobalMaxPooling1D())
cnn_ii_model = Sequential()
cnn_ii_model.add(Embedding(config['max_features'],
config['embedding_dims'],
input_length = config['input_length'],
weights = [weights['Wemb']] if 'Wemb' in weights else None))
#dropout = 0.2))
cnn_ii_model.add(Convolution1D(nb_filter = config['nb_filter_2'],
filter_length = config['filter_length_2'],
border_mode = 'valid',
activation = 'relu',
subsample_length = 1))
cnn_ii_model.add(GlobalMaxPooling1D())
# merged model
merged_model = Sequential()
# merged_model.add(Merge([bgrnn_model, cnn_i_model, cnn_ii_model], mode='concat', concat_axis=1))
merged_model.add(Merge([bgrnn_model, cnn_i_model, cnn_ii_model], mode='concat', concat_axis=1))
merged_model.add(Dropout(0.25))
if config['nb_classes'] > 2:
merged_model.add(Dense(config['nb_classes'], activation='softmax'))
loss_type = 'categorical_crossentropy'
else:
merged_model.add(Dense(1, activation='sigmoid'))
loss_type = 'binary_crossentropy'
merged_model.compile(loss=loss_type,
optimizer=self.get_optimizer(config['optimizer']),
metrics=['accuracy'])
return merged_model
def main():
optparser = OptionParser()
optparser.add_option("-t", "--task", dest="key_subtask", default="D")
optparser.add_option("-p", "--nb_epoch", dest="nb_epoch", type="int", default=50)
optparser.add_option("-e", "--embedding", dest="fname_Wemb", default="glove.twitter.27B.25d.txt")
optparser.add_option("-d", "--hidden_dims", dest="hidden_dims", type="int", default=250)
optparser.add_option("-f", "--nb_filter_1", dest="nb_filter_1", type="int", default=200)
optparser.add_option("-F", "--nb_filter_2", dest="nb_filter_2", type="int", default=200)
optparser.add_option("-r", "--rnn_output_dims", dest="rnn_output_dims", type="int", default=100)
optparser.add_option("-l", "--filter_length_1", dest="filter_length_1", type="int", default=6)
optparser.add_option("-L", "--filter_length_2", dest="filter_length_2", type="int", default=3)
optparser.add_option("-w", "--dropout_W", dest="dropout_W", type="float", default=0.25)
optparser.add_option("-u", "--dropout_U", dest="dropout_U", type="float", default=0.25)
optparser.add_option("-o", "--optimizer", dest="optimizer", default="rmsprop")
opts, args = optparser.parse_args()
trainer = Trainer(opts)
trainer.train()
# test = data_manager.read_texts_labels(opts.key_subtask, 'devtest')
score = trainer.evaluate('test_new', verbose=1)
print "Evaluation score: %.3f" % score
trainer.load_model_weight()
score = trainer.evaluate('test_new', verbose=1)
print "Evaluation score: %.3f" % score
if __name__ == '__main__':
main()