-
Notifications
You must be signed in to change notification settings - Fork 2
/
data.py
216 lines (174 loc) · 7 KB
/
data.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
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
from __future__ import absolute_import, division, unicode_literals
import io
import numpy as np
import logging
import cmath
import random
import math
import os
from keras.preprocessing.sequence import pad_sequences
def load_complex_embedding(embedding_dir):
word2id = np.load(os.path.join(embedding_dir,'word2id.npy')).item()
phase_embedding = np.load(os.path.join(embedding_dir,'phase_embedding.npy'))
amplitude_embedding = np.load(os.path.join(embedding_dir,'amplitude_embedding.npy'))
complex_embedding_params = {'word2id':word2id, 'phase_embedding': phase_embedding, 'amplitude_embedding': amplitude_embedding}
print(phase_embedding.shape)
print(amplitude_embedding.shape)
print(len(word2id))
return complex_embedding_params
# Create dictionary
def create_dictionary(sentences, threshold=0):
words = {}
for s in sentences:
for word in s:
words[word] = words.get(word, 0) + 1
if threshold > 0:
newwords = {}
for word in words:
if words[word] >= threshold:
newwords[word] = words[word]
words = newwords
# words['<s>'] = 1e9 + 4
# words['</s>'] = 1e9 + 3
# words['<p>'] = 1e9 + 2
sorted_words = sorted(words.items(), key=lambda x: -x[1]) # inverse sort
id2word = []
word2id = {}
for i, (w, _) in enumerate(sorted_words):
id2word.append(w)
word2id[w] = i+1
return id2word, word2id
def form_matrix(file_name):
word_list = []
ll = []
with io.open(file_name, 'r',encoding='utf-8') as f:
for line in f:
word, vec = line.split(' ', 1)
ll.append(np.fromstring(vec, sep=' '))
word_list.append(word)
matrix = np.asarray(ll)
return matrix, word_list
def orthonormalized_word_embeddings(word_embeddings_file):
matrix, word_list = form_matrix(word_embeddings_file)
print('Initial matrix constructed!')
matrix_norm = np.zeros((matrix.shape[0], matrix.shape[1]))
matrix = matrix.astype(np.float)
matrix_sum = np.sqrt(np.sum(np.square(matrix), axis=1))
for i in range(np.shape(matrix)[0]):
matrix_norm[i] = matrix[i]/matrix_sum[i]
print('Matrix normalized')
##q - basis vectors(num_words x dimension).
##r - coefficients of each word in the basis(dimension x num_words)
q, r = np.linalg.qr(np.transpose(matrix_norm), mode = 'complete')
print('qr factorization completed. Matrix orthogonalized!')
## Dot product of king and prince vectors same as in the original embeddings (0.76823)
# king = word_list.index('king')
# prince = word_list.index('prince')
# print (np.dot(r[:, king], r[:, prince]))
return r, word_list
# Get word vectors from vocabulary (glove, word2vec, fasttext ..)
def get_wordvec(path_to_vec, word2id=None, orthonormalized=True):
if orthonormalized:
coefficients_matrix, word_list = orthonormalized_word_embeddings(path_to_vec)
else:
matrix, word_list = form_matrix(path_to_vec)
coefficients_matrix = np.transpose(matrix)
word_vec = {}
with io.open(path_to_vec, 'r', encoding='utf-8') as f:
# if word2vec or fasttext file : skip first line "next(f)"
if word2id == None:
print('program goes here!')
for word in word_list:
word_vec[word] = coefficients_matrix[:, word_list.index(word)]
else:
for line in f:
word, vec = line.split(' ', 1)
if word in word2id:
word_vec[word] = coefficients_matrix[:, word_list.index(word)]
logging.info('Found {0} words with word vectors, out of \
{1} words'.format(len(word_vec), len(word2id)))
return word_vec
# Set word phases
# Currently only using random phase
def set_wordphase(word2id):
word2phase = {}
for word in word2id.keys():
word2phase[word] = random.random()*2*math.pi
return word2phase
def get_index_batch(embedding_params, batch):
batch = [sent if sent != [] else ['.'] for sent in batch]
embeddings = []
word2id = embedding_params['word2id']
for sent in batch:
sentvec = []
for word in sent:
if word in word2id:
assert word2id[word] > 0
sentvec.append(word2id[word])
if not sentvec:
vec = np.zeros(embedding_params['wvec_dim'])
sentvec.append(vec)
# word_count = len(sentvec)
# sentvec = np.mean(sentvec, 0)*math.sqrt(word_count)
embeddings.append(sentvec)
# embeddings = np.vstack(embeddings)
return embeddings
def get_vector_batch(embedding_params, batch):
batch = [sent if sent != [] else ['.'] for sent in batch]
embeddings = []
for sent in batch:
sentvec = []
for word in sent:
if word in embedding_params['word_vec']:
wordvec = embedding_params['word_vec'][word]
if word in embedding_params['word_complex_phase']:
complex_phase = embedding_params['word_complex_phase'][word]
wordvec = [x * cmath.exp(1j*complex_phase) for x in wordvec]
sentvec.append(wordvec)
if not sentvec:
vec = np.zeros(embedding_params['wvec_dim'])
sentvec.append(vec)
word_count = len(sentvec)
sentvec = np.mean(sentvec, 0)*math.sqrt(word_count)
embeddings.append(sentvec)
embeddings = np.vstack(embeddings)
return embeddings
def get_lookup_table(embedding_params):
id2word = embedding_params['id2word']
word_vec = embedding_params['word_vec']
lookup_table = []
# Index 0 corresponds to nothing
lookup_table.append([0]* embedding_params['wvec_dim'])
for i in range(0, len(id2word)):
word = id2word[i]
wvec = [0]* embedding_params['wvec_dim']
if word in word_vec:
wvec = word_vec[word]
# print(wvec)
lookup_table.append(wvec)
lookup_table = np.asarray(lookup_table)
return(lookup_table)
def batch_gen(data, max_sequence_length):
sentences = data['X']
labels = data['y']
# print(labels)
for batch, label in zip(sentences, labels):
padded_batch = pad_sequences(batch, maxlen=max_sequence_length, dtype='int32',
padding='post', truncating='post', value=0.)
yield np.asarray(padded_batch), np.asarray(label)
def data_gen(data, max_sequence_length):
sentences = data['X']
labels = data['y']
# print(labels)
# batch_list = []
# for batch, label in zip(sentences, labels):
# padded_batch = pad_sequences(batch, maxlen=max_sequence_length, dtype='int32',
# padding='post', truncating='post', value=0.)
# batch_list=padded_batch
padded_sentences = pad_sequences(sentences[0], maxlen=max_sequence_length, dtype='int32',padding='post', truncating='post', value=0.)
return np.asarray(padded_sentences), np.transpose(np.asarray(labels))
def main():
complex_embedding_dir = 'eval/eval_CR/embedding'
load_complex_embedding(complex_embedding_dir)
if __name__ == '__main__':
main()