-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathDataGenerator.py
257 lines (221 loc) · 11.9 KB
/
DataGenerator.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
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
from tensorflow.keras.utils import Sequence
import numpy as np
import pickle
import re
from tensorflow.keras.preprocessing import sequence
from resource_loading import load_NRC, load_LIWC, load_vocabulary, load_stopwords
from feature_encoders import encode_emotions, encode_pronouns, encode_stopwords, encode_liwc_categories
class DataGenerator(Sequence):
'Generates data for Keras'
def __init__(self, user_level_data, subjects_split, set_type,
hyperparams_features,
batch_size, seq_len,
compute_liwc=False,
post_groups_per_user=None, posts_per_group=10, post_offset = 0,
max_posts_per_user=None,
pronouns=["i", "me", "my", "mine", "myself"],
shuffle=True,
keep_last_batch=True, return_subjects=False, chunk_level_datapoints=True,
keep_first_batches=False,
ablate_emotions=False, ablate_liwc=False, logger=None):
'Initialization'
self.seq_len = seq_len
# Instantiate tokenizer
self.subjects_split = subjects_split
self.set = set_type
self.batch_size = batch_size
self.data = user_level_data
self.pronouns = pronouns
self.compute_liwc = compute_liwc
self.keep_last_batch = keep_last_batch
self.shuffle = shuffle
self.max_posts_per_user = max_posts_per_user
self.post_groups_per_user = post_groups_per_user
self.post_offset = post_offset
self.posts_per_group = posts_per_group
self.generated_labels = []
self.padding = "pre"
self.pad_value = 0
self.keep_first_batches=keep_first_batches # in the rolling window case, whether it will keep
self.chunk_level_datapoints = chunk_level_datapoints
self.logger = logger
self.return_subjects = return_subjects
self.vocabulary = load_vocabulary(hyperparams_features['vocabulary_path'])
self.voc_size = hyperparams_features['max_features']
if ablate_emotions:
self.emotions = []
else:
self.emotion_lexicon = load_NRC(hyperparams_features['nrc_lexicon_path'])
self.emotions = list(self.emotion_lexicon.keys())
self.liwc_dict = load_LIWC(hyperparams_features['liwc_path'])
self.liwc_words_for_categories = pickle.load(open(hyperparams_features["liwc_words_cached"], "rb"))
if ablate_liwc:
self.liwc_categories = []
else:
self.liwc_categories = set(self.liwc_dict.keys())
self.stopwords_list = load_stopwords(hyperparams_features['stopwords_path'])
self._post_indexes_per_user()
self.on_epoch_end()
def _post_indexes_per_user(self):
self.indexes_per_user = {u: [] for u in range(len(self.subjects_split[self.set]))}
self.indexes_with_user = []
self.item_weights = []
for u in range(len(self.subjects_split[self.set])):
if self.subjects_split[self.set][u] not in self.data:
if self.logger:
self.logger.warning("User %s has no posts in %s set. Ignoring.\n" % (
self.subjects_split[self.set][u], self.set))
continue
user_posts = self.data[self.subjects_split[self.set][u]]['texts']
if self.max_posts_per_user:
user_posts = user_posts[:self.max_posts_per_user]
if self.chunk_level_datapoints:
# Non-overlapping chunks
nr_post_groups = int(np.ceil(len(user_posts) / self.posts_per_group))
if self.post_groups_per_user:
nr_post_groups = min(self.post_groups_per_user, nr_post_groups)
for i in range(nr_post_groups):
# Generate random ordered samples of the posts
self.indexes_per_user[u].append(range(i*self.posts_per_group + self.post_offset,
min((i+1)*self.posts_per_group + self.post_offset, len(user_posts))))
self.indexes_with_user.append((u, range(i*self.posts_per_group ,
min((i+1)*self.posts_per_group + self.post_offset, len(user_posts)))))
else:
# Rolling window of datapoints: chunks with overlapping posts
nr_post_groups = len(user_posts)
if self.post_groups_per_user:
nr_post_groups = min(self.post_groups_per_user, nr_post_groups)
if self.keep_first_batches:
# Generate datapoints for first posts, before a complete chunk
for i in range(self.posts_per_group):
self.indexes_per_user[u].append(range(self.post_offset, i + self.post_offset,
))
self.indexes_with_user.append((u, range(self.post_offset, i + self.post_offset,
)))
for i in range(nr_post_groups):
# Stop at the last complete chunk
if i + self.posts_per_group + self.post_offset > len(user_posts):
break
self.indexes_per_user[u].append(range(i + self.post_offset,
min(i + self.posts_per_group + self.post_offset,
len(user_posts))))
self.indexes_with_user.append((u, range(i,
min(i+self.posts_per_group + self.post_offset,
len(user_posts)))))
self.item_weights = []
def __encode_text__(self, tokens, raw_text):
# Using voc_size-1 value for OOV token
encoded_tokens = [self.vocabulary.get(w, self.voc_size-1) for w in tokens]
encoded_emotions = encode_emotions(tokens, self.emotion_lexicon, self.emotions)
encoded_pronouns = encode_pronouns(tokens, self.pronouns)
encoded_stopwords = encode_stopwords(tokens, self.stopwords_list)
if not self.compute_liwc:
encoded_liwc = None
else:
encoded_liwc = encode_liwc_categories(tokens, self.liwc_categories, self.liwc_words_for_categories)
return (encoded_tokens, encoded_emotions, encoded_pronouns, encoded_stopwords, encoded_liwc,
)
def __len__(self):
'Denotes the number of batches per epoch'
if self.keep_last_batch:
return int(np.ceil(len(self.indexes) / self.batch_size)) # + 1 to not discard last batch
return int((len(self.indexes))/self.batch_size)
def __getitem__(self, index):
'Generate one batch of data'
# Reset generated labels
if index == 0:
self.generated_labels = []
# Generate indexes of the batch
indexes = self.indexes[index*self.batch_size:(index+1)*self.batch_size]
# Find users
user_indexes = [t[0] for t in indexes]
users = set([self.subjects_split[self.set][i] for i in user_indexes
if self.subjects_split[self.set][i] in self.data.keys()]) # TODO: maybe needs a warning that user is missing
post_indexes_per_user = {u: [] for u in users}
# Sample post ids
for u, post_indexes in indexes:
user = self.subjects_split[self.set][u]
# Note: was bug here - changed it into a list
post_indexes_per_user[user].append(post_indexes)
X, s, y = self.__data_generation_hierarchical__(users, post_indexes_per_user)
if self.return_subjects:
return X, s, y
else:
return X, y
def on_epoch_end(self):
'Updates indexes after each epoch'
self.indexes = self.indexes_with_user
# np.arange(len(self.subjects_split[self.set]))
if self.shuffle:
np.random.shuffle(self.indexes)
def __data_generation_hierarchical__(self, users, post_indexes):
'Generates data containing batch_size samples' # X : (n_samples, *dim, n_channels)
user_tokens = []
user_categ_data = []
user_sparse_data = []
labels = []
subjects = []
for subject in users:
all_words = []
all_raw_texts = []
liwc_scores = []
if 'label' in self.data[subject]:
label = self.data[subject]['label']
else:
label = None
for post_index_range in post_indexes[subject]:
# Sample
texts = [self.data[subject]['texts'][i] for i in post_index_range]
if 'liwc' in self.data[subject] and not self.compute_liwc:
liwc_selection = [self.data[subject]['liwc'][i] for i in post_index_range]
raw_texts = [self.data[subject]['raw'][i] for i in post_index_range]
all_words.append(texts)
if 'liwc' in self.data[subject] and not self.compute_liwc:
liwc_scores.append(liwc_selection)
all_raw_texts.append(raw_texts)
for i, words in enumerate(all_words):
tokens_data = []
categ_data = []
sparse_data = []
raw_text = all_raw_texts[i]
words = all_words[i]
for p, posting in enumerate(words):
encoded_tokens, encoded_emotions, encoded_pronouns, encoded_stopwords, encoded_liwc, \
= self.__encode_text__(words[p], raw_text[p])
if 'liwc' in self.data[subject] and not self.compute_liwc:
liwc = liwc_scores[i][p]
else:
liwc = encoded_liwc
try:
subject_id = int(re.findall('[0-9]+', subject)[0])
except IndexError:
subject_id = subject
tokens_data.append(encoded_tokens)
categ_data.append(encoded_emotions + [encoded_pronouns] + liwc)
sparse_data.append(encoded_stopwords)
# For each range
tokens_data_padded = np.array(sequence.pad_sequences(tokens_data, maxlen=self.seq_len,
padding=self.padding,
truncating=self.padding))
user_tokens.append(tokens_data_padded)
user_categ_data.append(categ_data)
user_sparse_data.append(sparse_data)
labels.append(label)
subjects.append(subject)
user_tokens = sequence.pad_sequences(user_tokens,
maxlen=self.posts_per_group,
value=self.pad_value)
user_tokens = np.rollaxis(np.dstack(user_tokens), -1)
user_categ_data = sequence.pad_sequences(user_categ_data,
maxlen=self.posts_per_group,
value=self.pad_value, dtype='float32')
user_categ_data = np.rollaxis(np.dstack(user_categ_data), -1)
user_sparse_data = sequence.pad_sequences(user_sparse_data,
maxlen=self.posts_per_group,
value=self.pad_value)
user_sparse_data = np.rollaxis(np.dstack(user_sparse_data), -1)
self.generated_labels.extend(labels)
labels = np.array(labels, dtype=np.float32)
return ((user_tokens, user_categ_data, user_sparse_data),
np.array(subjects),
labels)