-
Notifications
You must be signed in to change notification settings - Fork 7
/
Copy pathutils.py
172 lines (152 loc) · 7.59 KB
/
utils.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
import pickle
import numpy as np
def convert_to_sequence(texts, word_to_index, padding=False, size_limit=10):
sequences = {}
for idx, tokens in texts.items():
if padding:
sequences[idx] = np.array([word_to_index[token] for i, token in enumerate(tokens)
if i < size_limit] + [0]*(max(0, size_limit-len(tokens))),
dtype=np.int32)
else:
sequences[idx] = np.array([word_to_index[token] for token in tokens],
dtype=np.int32)
return sequences
def generate_dataset(pp_name, lower_opt, version, max_seq_length=-1,
reverse_train_pairs=False, padding=True, autoneg=0):
if padding:
res = generate_dataset_with_padding(pp_name, lower_opt, version, max_seq_length,
reverse_train_pairs, autoneg)
else:
res = generate_dataset_without_padding(pp_name, lower_opt, version, max_seq_length,
reverse_train_pairs, autoneg)
return res
def generate_dataset_with_padding(pp_name, lower_opt, version, max_seq_length=-1,
reverse_train_pairs=False, autoneg=0):
parsed_fn = "msrpc_{}_{}_{}.pickle".format(pp_name, lower_opt, version)
# Loading pre-processed corpus
[parsed_texts,
index_to_word,
word_to_index,
pairs_train,
Y_train_list,
pairs_test,
Y_test_list] = pickle.load(open("./data/"+parsed_fn, 'rb'))
# Computing the max_seq_length if not provided
if max_seq_length < 0:
max_seq_length = np.max([len(tokens) for idx, tokens in parsed_texts.items()])
# Transforming list of tokens to sequence of indices
sequences = convert_to_sequence(parsed_texts, word_to_index,
padding=True, size_limit=max_seq_length)
# Training Data
if reverse_train_pairs:
X_train1 = np.zeros((len(pairs_train)*2+autoneg, max_seq_length), dtype=np.int32)
X_train2 = np.zeros((len(pairs_train)*2+autoneg, max_seq_length), dtype=np.int32)
for i, (x1, x2) in enumerate(pairs_train):
X_train1[i*2,:] = sequences[x1]
X_train1[i*2+1,:] = sequences[x2]
X_train2[i*2,:] = sequences[x2]
X_train2[i*2+1,:] = sequences[x1]
Y_train = np.array([Y_train_list[i//2] for i in range(len(Y_train_list)*2)]+[0]*autoneg, dtype=np.int32)
else:
X_train1 = np.zeros((len(pairs_train)+autoneg, max_seq_length), dtype=np.int32)
X_train2 = np.zeros((len(pairs_train)+autoneg, max_seq_length), dtype=np.int32)
for i, (x1, x2) in enumerate(pairs_train):
X_train1[i,:] = sequences[x1]
X_train2[i,:] = sequences[x2]
Y_train = np.array(Y_train_list+[0]*autoneg, dtype=np.int32)
# Adding automatically generated negative samples
# from sentences in positive samples
left, right = zip(*[tup for tup, _class in zip(pairs_train, Y_train_list) if _class==1])
pos_ids = np.array(list(set(left+right)), dtype=np.int32)
selected_pos_ids = np.random.choice(pos_ids, size=autoneg)
pairs_train_set = set(pairs_train)
pairs_test_set = set(pairs_test)
all_ids = np.array(list(parsed_texts.keys()), dtype=np.int32)
starting_i = len(pairs_train)*2 if reverse_train_pairs else len(pairs_train)
for i, pos_id in enumerate(selected_pos_ids, start=starting_i):
while True:
paired_id = np.random.choice(all_ids)
# Check it is not in test set too
if ((pos_id, paired_id) not in pairs_train_set and
(paired_id, pos_id) not in pairs_train_set and
(pos_id, paired_id) not in pairs_test_set and
(paired_id, pos_id) not in pairs_test_set):
X_train1[i,:] = sequences[pos_id]
X_train2[i,:] = sequences[paired_id]
break
else:
print("Ignoring randomly generated sample that already exists")
# Test Data
X_test1 = np.zeros((len(pairs_test), max_seq_length), dtype=np.int32)
X_test2 = np.zeros((len(pairs_test), max_seq_length), dtype=np.int32)
for i, (x1, x2) in enumerate(pairs_test):
X_test1[i,:] = sequences[x1]
X_test2[i,:] = sequences[x2]
Y_test = np.array(Y_test_list, dtype=np.int32)
return index_to_word, word_to_index, X_train1, X_train2, Y_train, X_test1, X_test2, Y_test
def generate_dataset_without_padding(pp_name, lower_opt, version, max_seq_length=-1,
reverse_train_pairs=False, autoneg=0):
parsed_fn = "msrpc_{}_{}_{}.pickle".format(pp_name, lower_opt, version)
# Loading pre-processed corpus
[parsed_texts,
index_to_word,
word_to_index,
pairs_train,
Y_train_list,
pairs_test,
Y_test_list] = pickle.load(open("./data/"+parsed_fn, 'rb'))
# Computing the max_seq_length if not provided
if max_seq_length < 0:
max_seq_length = np.max([len(tokens) for idx, tokens in parsed_texts.items()])
# Transforming list of tokens to sequence of indices
sequences = convert_to_sequence(parsed_texts, word_to_index,
padding=False, size_limit=max_seq_length)
# Training Data
if reverse_train_pairs:
X_train1 = []
X_train2 = []
for i, (x1, x2) in enumerate(pairs_train):
X_train1.append(np.array(sequences[x1], dtype=np.int32))
X_train1.append(np.array(sequences[x2], dtype=np.int32))
X_train2.append(np.array(sequences[x2], dtype=np.int32))
X_train2.append(np.array(sequences[x1], dtype=np.int32))
Y_train = np.array([Y_train_list[i//2] for i in range(len(Y_train_list)*2)]+[0]*autoneg, dtype=np.int32)
else:
X_train1 = []
X_train2 = []
for i, (x1, x2) in enumerate(pairs_train):
X_train1.append(np.array(sequences[x1], dtype=np.int32))
X_train2.append(np.array(sequences[x2], dtype=np.int32))
Y_train = np.array(Y_train_list+[0]*autoneg, dtype=np.int32)
# Adding automatically generated negative samples
# from sentences in positive samples
left, right = zip(*[tup for tup, _class in zip(pairs_train, Y_train_list) if _class==1])
pos_ids = np.array(list(set(left+right)), dtype=np.int32)
selected_pos_ids = np.random.choice(pos_ids, size=autoneg)
pairs_train_set = set(pairs_train)
pairs_test_set = set(pairs_test)
all_ids = np.array(list(parsed_texts.keys()), dtype=np.int32)
starting_i = len(pairs_train)*2 if reverse_train_pairs else len(pairs_train)
for i, pos_id in enumerate(selected_pos_ids, start=starting_i):
while True:
paired_id = np.random.choice(all_ids)
# Check it is not in test set too
if ((pos_id, paired_id) not in pairs_train_set and
(paired_id, pos_id) not in pairs_train_set and
(pos_id, paired_id) not in pairs_test_set and
(paired_id, pos_id) not in pairs_test_set):
X_train1.append(np.array(sequences[pos_id], dtype=np.int32))
X_train2.append(np.array(sequences[paired_id], dtype=np.int32))
break
else:
print("Ignoring randomly generated sample that already exists")
# Test Data
X_test1 = []
X_test2 = []
for i, (x1, x2) in enumerate(pairs_test):
X_test1.append(np.array(sequences[x1], dtype=np.int32))
X_test2.append(np.array(sequences[x2], dtype=np.int32))
Y_test = np.array(Y_test_list, dtype=np.int32)
return (index_to_word, word_to_index,
np.array(X_train1), np.array(X_train2), Y_train,
np.array(X_test1), np.array(X_test2), Y_test)