-
Notifications
You must be signed in to change notification settings - Fork 7
/
Copy path4_train.py
179 lines (156 loc) · 5.36 KB
/
4_train.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
import csv
import random
random.seed(961)
import argparse
from os.path import join
from keras.models import Input, Model, load_model
from keras.layers import Lambda, Subtract, Multiply, Concatenate, Embedding, Dropout
from keras.layers import Dense, GRU, CuDNNGRU, LSTM, CuDNNLSTM, Bidirectional
from keras.optimizers import Adam, Adamax, Adadelta
from keras.callbacks import ModelCheckpoint, ReduceLROnPlateau, Callback
from keras_self_attention import SeqWeightedAttention
from keras_ordered_neurons import ONLSTM
from helpers import load_embeddings_dict
from helpers import map_sentence, f1
from data_generator import DataGenerator
def build_model(embeddings_size):
# Inputs
q1_embeddings_input = Input(shape=(None, embeddings_size,), name='q1_word_embeddings')
q2_embeddings_input = Input(shape=(None, embeddings_size,), name='q2_word_embeddings')
# RNN
word_lstm1 = Bidirectional(
ONLSTM(
units=256,
chunk_size=8,
dropout=args.dropout_rate,
return_sequences=True,
kernel_initializer='glorot_normal'
)
)
q1_word_lstm1 = word_lstm1(q1_embeddings_input)
q2_word_lstm1 = word_lstm1(q2_embeddings_input)
word_lstm2 = Bidirectional(
ONLSTM(
units=256,
chunk_size=8,
dropout=args.dropout_rate,
return_sequences=True,
kernel_initializer='glorot_normal'
)
)
q1_word_lstm2 = word_lstm2(q1_word_lstm1)
q2_word_lstm2 = word_lstm2(q2_word_lstm1)
word_attention = SeqWeightedAttention()
q1_word_attention = word_attention(q1_word_lstm2)
q2_word_attention = word_attention(q2_word_lstm2)
# Concatenate
subtract = Subtract()([q1_word_attention, q2_word_attention])
multiply_subtract = Multiply()([subtract, subtract])
# Fully Connected
dense1 = Dropout(args.dropout_rate)(
Dense(units=1024, activation='relu', kernel_initializer='glorot_normal')(multiply_subtract)
)
dense2 = Dropout(args.dropout_rate)(
Dense(units=512, activation='relu', kernel_initializer='glorot_normal')(dense1)
)
dense3 = Dropout(args.dropout_rate)(
Dense(units=256, activation='relu', kernel_initializer='glorot_normal')(dense2)
)
dense4 = Dropout(args.dropout_rate)(
Dense(units=128, activation='relu', kernel_initializer='glorot_normal')(dense3)
)
# Predict
output = Dense(units=1, activation='sigmoid', kernel_initializer='glorot_normal')(dense4)
model = Model([q1_embeddings_input, q2_embeddings_input], output)
model.compile(optimizer=Adam(lr=0.001), loss='binary_crossentropy', metrics=['accuracy', f1])
model.summary()
return model
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('--data-dir', default='data_dir')
parser.add_argument('--embeddings-type', default='elmo', choices=['elmo', 'bert'])
parser.add_argument('--dropout-rate', default=0.2, type=float)
parser.add_argument('--epochs', default=100, type=int)
parser.add_argument('--batch-size', default=256, type=int)
parser.add_argument('--initial-epoch', default=0, type=int)
parser.add_argument('--dev-split', default=2000, type=int)
args = parser.parse_args()
embeddings_dict = load_embeddings_dict(join(args.data_dir, '%s_dict.pkl' % args.embeddings_type))
data = list()
sentences = set()
cnt = 0
with open(join(args.data_dir, 'train_processed_enlarged.csv'), 'r') as file:
reader = csv.reader(file)
for row in reader:
cnt += 2; print('Prepare Data: %s' % (cnt), end='\r')
data.append((
map_sentence(row[0], embeddings_dict),
map_sentence(row[1], embeddings_dict),
int(row[2])
))
data.append((
map_sentence(row[1], embeddings_dict),
map_sentence(row[0], embeddings_dict),
int(row[2])
))
sentences.add(row[0])
sentences.add(row[1])
for sentence in sentences:
cnt += 1; print('Prepare Data: %s' % (cnt), end='\r')
data.append((
map_sentence(sentence, embeddings_dict),
map_sentence(sentence, embeddings_dict),
1
))
print('Prepare %d examples: Done ' % (cnt))
random.shuffle(data)
train = data[args.dev_split:]
dev = data[:args.dev_split]
train_q1, train_q2, train_label = zip(*train)
if len(dev) != 0:
dev_q1, dev_q2, dev_label = zip(*dev)
if args.initial_epoch == 0:
model = build_model(
len(embeddings_dict[list(embeddings_dict)[0]][0])
)
else:
model = load_model(
filepath='checkpoints/epoch%s.h5' % args.initial_epoch,
custom_objects={
'f1': f1,
'SeqWeightedAttention': SeqWeightedAttention,
'ONLSTM': ONLSTM
}
)
train_gen = DataGenerator(
train_q1,
train_q2,
train_label,
args.batch_size
)
if len(dev) != 0:
dev_gen = DataGenerator(
dev_q1,
dev_q2,
dev_label,
args.batch_size
)
checkpoint_cb = ModelCheckpoint(
filepath='checkpoints/epoch{epoch:02d}.h5',
monitor='val_f1',
verbose=1,
save_best_only=False,
mode='max',
period=10
)
if len(dev) != 0:
model.fit_generator(generator=train_gen,
validation_data=dev_gen,
epochs=args.epochs,
callbacks=[checkpoint_cb],
initial_epoch=args.initial_epoch)
else:
model.fit_generator(generator=train_gen,
epochs=args.epochs,
callbacks=[checkpoint_cb],
initial_epoch=args.initial_epoch)