forked from bradleypallen/keras-quora-question-pairs
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathkeras-quora-question-pairs.py
169 lines (144 loc) · 6.74 KB
/
keras-quora-question-pairs.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
from __future__ import print_function
import numpy as np
import csv, datetime, time, json
from zipfile import ZipFile
from os.path import expanduser, exists
from keras.preprocessing.text import Tokenizer
from keras.preprocessing.sequence import pad_sequences
from keras.models import Sequential
from keras.layers import Embedding, Dense, Dropout, Reshape, Merge, BatchNormalization, TimeDistributed, Lambda
from keras.regularizers import l2
from keras.callbacks import Callback, ModelCheckpoint
from keras.utils.data_utils import get_file
from keras import backend as K
from sklearn.model_selection import train_test_split
KERAS_DATASETS_DIR = expanduser('~/.keras/datasets/')
QUESTION_PAIRS_FILE_URL = 'http://qim.ec.quoracdn.net/quora_duplicate_questions.tsv'
QUESTION_PAIRS_FILE = 'quora_duplicate_questions.tsv'
GLOVE_ZIP_FILE_URL = 'http://nlp.stanford.edu/data/glove.840B.300d.zip'
GLOVE_ZIP_FILE = 'glove.840B.300d.zip'
GLOVE_FILE = 'glove.840B.300d.txt'
Q1_TRAINING_DATA_FILE = 'q1_train.npy'
Q2_TRAINING_DATA_FILE = 'q2_train.npy'
LABEL_TRAINING_DATA_FILE = 'label_train.npy'
WORD_EMBEDDING_MATRIX_FILE = 'word_embedding_matrix.npy'
NB_WORDS_DATA_FILE = 'nb_words.json'
MAX_NB_WORDS = 200000
MAX_SEQUENCE_LENGTH = 25
EMBEDDING_DIM = 300
MODEL_WEIGHTS_FILE = 'question_pairs_weights.h5'
VALIDATION_SPLIT = 0.1
TEST_SPLIT = 0.1
RNG_SEED = 13371447
NB_EPOCHS = 25
if exists(Q1_TRAINING_DATA_FILE) and exists(Q2_TRAINING_DATA_FILE) and exists(LABEL_TRAINING_DATA_FILE) and exists(NB_WORDS_DATA_FILE) and exists(WORD_EMBEDDING_MATRIX_FILE):
q1_data = np.load(open(Q1_TRAINING_DATA_FILE, 'rb'))
q2_data = np.load(open(Q2_TRAINING_DATA_FILE, 'rb'))
labels = np.load(open(LABEL_TRAINING_DATA_FILE, 'rb'))
word_embedding_matrix = np.load(open(WORD_EMBEDDING_MATRIX_FILE, 'rb'))
with open(NB_WORDS_DATA_FILE, 'r') as f:
nb_words = json.load(f)['nb_words']
else:
if not exists(KERAS_DATASETS_DIR + QUESTION_PAIRS_FILE):
get_file(QUESTION_PAIRS_FILE, QUESTION_PAIRS_FILE_URL)
print("Processing", QUESTION_PAIRS_FILE)
question1 = []
question2 = []
is_duplicate = []
with open(KERAS_DATASETS_DIR + QUESTION_PAIRS_FILE, encoding='utf-8') as csvfile:
reader = csv.DictReader(csvfile, delimiter='\t')
for row in reader:
question1.append(row['text1'])
question2.append(row['text2'])
is_duplicate.append(row['duplicate'])
print('Question pairs: %d' % len(question1))
questions = question1 + question2
tokenizer = Tokenizer(nb_words=MAX_NB_WORDS)
tokenizer.fit_on_texts(questions)
question1_word_sequences = tokenizer.texts_to_sequences(question1)
question2_word_sequences = tokenizer.texts_to_sequences(question2)
word_index = tokenizer.word_index
print("Words in index: %d" % len(word_index))
if not exists(KERAS_DATASETS_DIR + GLOVE_ZIP_FILE):
zipfile = ZipFile(get_file(GLOVE_ZIP_FILE, GLOVE_ZIP_FILE_URL))
zipfile.extract(GLOVE_FILE, path=KERAS_DATASETS_DIR)
print("Processing", GLOVE_FILE)
embeddings_index = {}
with open(KERAS_DATASETS_DIR + GLOVE_FILE, encoding='utf-8') as f:
for line in f:
values = line.split(' ')
word = values[0]
embedding = np.asarray(values[1:], dtype='float32')
embeddings_index[word] = embedding
print('Word embeddings: %d' % len(embeddings_index))
nb_words = min(MAX_NB_WORDS, len(word_index))
word_embedding_matrix = np.zeros((nb_words + 1, EMBEDDING_DIM))
for word, i in word_index.items():
if i > MAX_NB_WORDS:
continue
embedding_vector = embeddings_index.get(word)
if embedding_vector is not None:
word_embedding_matrix[i] = embedding_vector
print('Null word embeddings: %d' % np.sum(np.sum(word_embedding_matrix, axis=1) == 0))
q1_data = pad_sequences(question1_word_sequences, maxlen=MAX_SEQUENCE_LENGTH)
q2_data = pad_sequences(question2_word_sequences, maxlen=MAX_SEQUENCE_LENGTH)
labels = np.array(is_duplicate, dtype=int)
print('Shape of question1 data tensor:', q1_data.shape)
print('Shape of question2 data tensor:', q2_data.shape)
print('Shape of label tensor:', labels.shape)
np.save(open(Q1_TRAINING_DATA_FILE, 'wb'), q1_data)
np.save(open(Q2_TRAINING_DATA_FILE, 'wb'), q2_data)
np.save(open(LABEL_TRAINING_DATA_FILE, 'wb'), labels)
np.save(open(WORD_EMBEDDING_MATRIX_FILE, 'wb'), word_embedding_matrix)
with open(NB_WORDS_DATA_FILE, 'w') as f:
json.dump({'nb_words': nb_words}, f)
X = np.stack((q1_data, q2_data), axis=1)
y = labels
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=TEST_SPLIT, random_state=RNG_SEED)
Q1_train = X_train[:,0]
Q2_train = X_train[:,1]
Q1_test = X_test[:,0]
Q2_test = X_test[:,1]
Q1 = Sequential()
Q1.add(Embedding(nb_words + 1, EMBEDDING_DIM, weights=[word_embedding_matrix], input_length=MAX_SEQUENCE_LENGTH, trainable=False))
Q1.add(TimeDistributed(Dense(EMBEDDING_DIM, activation='relu')))
Q1.add(Lambda(lambda x: K.max(x, axis=1), output_shape=(EMBEDDING_DIM, )))
Q2 = Sequential()
Q2.add(Embedding(nb_words + 1, EMBEDDING_DIM, weights=[word_embedding_matrix], input_length=MAX_SEQUENCE_LENGTH, trainable=False))
Q2.add(TimeDistributed(Dense(EMBEDDING_DIM, activation='relu')))
Q2.add(Lambda(lambda x: K.max(x, axis=1), output_shape=(EMBEDDING_DIM, )))
model = Sequential()
model.add(Merge([Q1, Q2], mode='concat'))
model.add(BatchNormalization())
model.add(Dense(200, activation='relu'))
model.add(BatchNormalization())
model.add(Dense(200, activation='relu'))
model.add(BatchNormalization())
model.add(Dense(200, activation='relu'))
model.add(BatchNormalization())
model.add(Dense(200, activation='relu'))
model.add(BatchNormalization())
model.add(Dense(1, activation='sigmoid'))
model.compile(loss='binary_crossentropy',
optimizer='adam',
metrics=['accuracy', 'precision', 'recall', 'fbeta_score'])
callbacks = [ModelCheckpoint(MODEL_WEIGHTS_FILE, monitor='val_acc', save_best_only=True)]
print("Starting training at", datetime.datetime.now())
t0 = time.time()
history = model.fit([Q1_train, Q2_train],
y_train,
nb_epoch=NB_EPOCHS,
validation_split=VALIDATION_SPLIT,
verbose=1,
callbacks=callbacks)
t1 = time.time()
print("Training ended at", datetime.datetime.now())
print("Minutes elapsed: %f" % ((t1 - t0) / 60.))
model.load_weights(MODEL_WEIGHTS_FILE)
loss, accuracy, precision, recall, fbeta_score = model.evaluate([Q1_test, Q2_test], y_test)
print('')
print('loss = {0:.4f}'.format(loss))
print('accuracy = {0:.4f}'.format(accuracy))
print('precision = {0:.4f}'.format(precision))
print('recall = {0:.4f}'.format(recall))
print('F = {0:.4f}'.format(fbeta_score))