-
Notifications
You must be signed in to change notification settings - Fork 0
/
utils.py
340 lines (277 loc) · 11.9 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
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
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
import os
import torch
from torch.utils.data import DataLoader, TensorDataset
import re
import pickle
import torch.nn as nn
import sys
from torch.nn.utils.rnn import pack_padded_sequence, pad_packed_sequence
import pathlib
UNKNOWN_CHAR = '*'
PAD = "*PAD*"
CHAR_PAD = "*CHAR_PAD*"
UNIQUE_WORD = "UUUKKKK"
ARTIFACTS_PATH = "./artifacts"
MODEL_FILE_NAME = "saved_model.pt"
TRAIN_DATASET_NAME = "train_dataset.pickle"
DEV_DATASET_NAME = "dev_dataset.pickle"
pathlib.Path(ARTIFACTS_PATH).mkdir(parents=True, exist_ok=True)
DEV_500_ACC_PATH = os.path.join(ARTIFACTS_PATH, TRAIN_DATASET_NAME)
TRAIN_DATASET_DIR = os.path.join(ARTIFACTS_PATH, TRAIN_DATASET_NAME)
DEV_DATASET_DIR = os.path.join(ARTIFACTS_PATH, DEV_DATASET_NAME)
MODEL_DIR = os.path.join(ARTIFACTS_PATH, MODEL_FILE_NAME)
class BILSTMNet(nn.Module):
def __init__(self, vocab_size, embedding_len, lstm_out_dim, output_dim, dicts, char_embedding_len, batch_size,
choice="a"):
super(BILSTMNet, self).__init__()
self.char_embedding_len = char_embedding_len
self.batch_size = batch_size
self.word_embed_dim = embedding_len
self.hidden_dim = lstm_out_dim
self.choice = choice
self.dicts = dicts
# Representation options before model:
if choice in ['a', 'c', 'd']:
self.word_embed = nn.Embedding(vocab_size, embedding_len)
if choice in ['b', 'd']:
self.char_embed = nn.Embedding(len(dicts.C2I), self.char_embedding_len)
self.chars_lstm = nn.LSTM(input_size=self.char_embedding_len, hidden_size=embedding_len, batch_first=True)
if choice == 'c':
self.prefix_embed = nn.Embedding(len(dicts.P2I), embedding_len)
self.suffix_embed = nn.Embedding(len(dicts.S2I), embedding_len)
if choice == 'd':
self.concat_linear_layer = nn.Linear(embedding_len * 2, embedding_len)
# Rest of the model:
self.bi_lstm = nn.LSTM(input_size=embedding_len, hidden_size=lstm_out_dim, bidirectional=True, num_layers=2,
batch_first=True)
self.out = nn.Linear(2 * lstm_out_dim, output_dim)
self.softmax = nn.LogSoftmax(dim=0)
def embed_lstm_a(self, sentence):
return self.word_embed(sentence)
def embed_lstm_b(self, sentence, total_seq_length, batch_size):
char_input = create_char_input(sentence, self.dicts)
words_len = torch.tensor([get_size_without_pad(self.dicts.C2I[PAD], word) for word in char_input])
embed_chars = self.char_embed(char_input)
# It's chars packing time:
packed_chars_input = pack_padded_sequence(embed_chars, words_len, batch_first=True,
enforce_sorted=False)
_, (lstm_last_h_output, _) = self.chars_lstm(packed_chars_input)
return lstm_last_h_output.view(batch_size, total_seq_length, self.word_embed_dim)
def embed_lstm_c(self, sentence):
prefix_input, suffix_input = make_prefix_suffix_input(sentence, self.dicts)
embed_word_input = self.word_embed(sentence)
embed_prefix_input = self.prefix_embed(prefix_input)
embed_suffix_input = self.suffix_embed(suffix_input)
return embed_word_input + embed_prefix_input + embed_suffix_input
def embed_lstm_d(self, sentence, total_seq_length, batch_size):
lstm_a_output = self.embed_lstm_a(sentence)
lstm_b_output = self.embed_lstm_b(sentence, total_seq_length, batch_size)
concat_output = torch.cat((lstm_a_output, lstm_b_output), 2)
return self.concat_linear_layer(concat_output)
def forward(self, sentence):
# batch size for resize the shape at the end.
batch_size = len(sentence)
# get the len of each vector without padding. if no padding, return len of vector.
seq_lens_no_pad = torch.tensor([get_size_without_pad(self.dicts.F2I[PAD], element) for element in sentence])
total_seq_length = sentence.shape[1]
if self.choice == 'a':
embed_input = self.embed_lstm_a(sentence)
elif self.choice == 'b':
embed_input = self.embed_lstm_b(sentence, total_seq_length, batch_size)
elif self.choice == 'c':
embed_input = self.embed_lstm_c(sentence)
elif self.choice == 'd':
embed_input = self.embed_lstm_d(sentence, total_seq_length, batch_size)
# packing embed_input before model layers.
packed_x = pack_padded_sequence(embed_input, seq_lens_no_pad, batch_first=True, enforce_sorted=False)
packed_lstm_output, _ = self.bi_lstm(packed_x)
lstm_output, _ = pad_packed_sequence(packed_lstm_output, batch_first=True, padding_value=0,
total_length=total_seq_length)
# Rest of the model calculation
output = self.out(lstm_output)
output = self.softmax(output)
output = output.permute(0, 2, 1)
return output
def save(self, path):
torch.save(self.state_dict(), path)
def load(self, path):
state_dict = torch.load(path)
self.load_state_dict(state_dict)
class Dictionaries:
def __init__(self, data_set):
# Word <-> Index:
extend_vocab = [PAD, UNIQUE_WORD] + list(data_set.vocab)
extend_tags = [PAD, UNIQUE_WORD] + list(data_set.tags)
extend_chars = [PAD, UNIQUE_WORD] + list(data_set.chars)
extend_prefix = [PAD[:3], UNIQUE_WORD[:3]] + list(data_set.pref)
extend_suffix = [PAD[-3:], UNIQUE_WORD[-3:]] + list(data_set.suff)
self.F2I = {word: i for i, word in enumerate(extend_vocab)}
self.I2F = {i: word for i, word in enumerate(extend_vocab)}
# Label <-> Index:
self.L2I = {tag: i for i, tag in enumerate(extend_tags)}
self.I2L = {i: tag for i, tag in enumerate(extend_tags)}
# char <-> index:
self.C2I = {char: i for i, char in enumerate(extend_chars)}
self.I2C = {i: char for i, char in enumerate(extend_chars)}
# pref/suff <-> index:
self.P2I = {pref: i for i, pref in enumerate(extend_prefix)}
self.I2P = {i: pref for i, pref in enumerate(extend_prefix)}
self.S2I = {suff: i for i, suff in enumerate(extend_suffix)}
self.I2S = {i: suff for i, suff in enumerate(extend_suffix)}
def tensorize_sequence(sequence, F2I):
"""`
Convert sentence to indexed tensor representation matrix.
:param sequence: a sequence.
:param F2I: feature to index dictionary
:return: indexed tensor representation matrix
"""
tensor_vector = torch.zeros(len(sequence), len(F2I))
for i, feature in enumerate(sequence):
one_hot_vec = torch.zeros(len(F2I))
one_hot_vec[F2I[feature]] = 1
tensor_vector[i] = one_hot_vec
return tensor_vector
def get_part1_file_directory(data_kind):
"""
return the directory of the file from data directory.
:param data_kind: name of the data file
:return: local directory.
"""
return "./data/{0}".format(data_kind)
def get_part3_file_directory(data_name="pos", data_kind="train"):
"""
return the directory of the file from data directory.
:param data_kind: name of the data file
:param data_name weather is pos or ner.
:return: local directory.
"""
return "./data/{0}/{1}".format(data_name, data_kind)
def part1_parser(data_dir, is_test=False):
"""
Get's data directory, parse it's content with '\t' to sequence and tag.
if is_test is true, skip on creating tags.
:param data_dir: data location.
:param is_test: imply whether to tag or not
:return: dataset (sequence, [int tag])
"""
data_set = []
lines_list = open(data_dir).read().splitlines()
for line in lines_list:
split_line = line.split('\t')
sentence = split_line[0]
if not is_test:
current_tag = split_line[1]
data_set.append((sentence, [int(current_tag)]))
else:
data_set.append(sentence)
return data_set
def pos_ner_parser(data_dir, data_name="pos", data_kind="train", to_lower=False, convert_digits=False):
"""
Get's data directory, parse it's content with '\t' to sequence and tag.
if is_test is true, skip on creating tags.
:param data_dir: data location.
:param data_kind: imply whether to tag or not.
:param data_name imply whether is pos or ner.
:param to_lower if lower case the letters
:param convert_digits if convert digits to '*DG*'
:return: dataset (sequence, [int tag])
"""
sentences_data_set = []
current_sentence_list = []
# parse by spaces if post, if ner parse by tab.
delimiter = ' ' if data_name == "pos" else '\t'
data_set = []
lines_list = open(data_dir).read().splitlines()
for line in lines_list:
raw_splitted = line.split(delimiter)
word = raw_splitted[0]
if word != '':
# convert all chars to lower case.
if to_lower:
word = word.lower()
# if we want to convert each digit to be DG for similarity, '300' = '400'.
if convert_digits:
word = re.sub('[0-9]', '*DG*', word)
if data_kind != "test":
tag = raw_splitted[1]
current_sentence_list.append((word, tag))
else:
current_sentence_list.append(word)
else:
# finished iterate over one single sentence:
sentences_data_set.append(current_sentence_list.copy())
# reset list for next sentence.
current_sentence_list.clear()
return sentences_data_set
def make_loader(data, F2I, L2I, batch_size):
# split the tupled given data to x and y.
max_sequence_len = max(len(tup[0]) for tup in data)
x = torch.LongTensor([convert_to_padded_indexes(sentence, F2I, max_sequence_len) for sentence, _ in data])
y = torch.LongTensor([convert_to_padded_indexes(tags, L2I, max_sequence_len) for _, tags in data])
dataset = TensorDataset(x, y)
return DataLoader(dataset, batch_size=batch_size, shuffle=True)
def make_test_loader(data, F2I, batch_size):
# split the tupled given data to x and y.
max_sequence_len = max(len(word) for word in data)
x = torch.LongTensor([convert_to_padded_indexes(sentence, F2I, max_sequence_len) for sentence in data])
dataset = TensorDataset(x)
return DataLoader(dataset, batch_size=batch_size, shuffle=False)
def get_size_without_pad(value, tensor_arr):
# get the size witout pad. if it zero convert it to 1 ( The packing doe's not execpt zero as word size)
return (tensor_arr.tolist().index(value) or 1) if value in tensor_arr else len(tensor_arr)
def convert_to_padded_indexes(sequence, index_dict, max_len):
indexed_list = list()
for sub_sequence in sequence:
indexed_list.append(index_dict[sub_sequence] if sub_sequence in index_dict else index_dict[UNIQUE_WORD])
pad_index = index_dict[PAD]
indexed_list.extend([pad_index] * (max_len - len(indexed_list)))
return indexed_list
def get_max_word_size(keys_list):
return len(max(keys_list, key=len))
def create_char_input(input, dicts):
F2I, I2F, C2I = dicts.F2I, dicts.I2F, dicts.C2I
max_word_len = get_max_word_size(list(F2I.keys()))
# input shape is (batch_size, num_sequences)
word_input = input.view(-1)
# input shape is (batch_size * num_sequences)
char_input = torch.zeros(len(word_input), max_word_len, dtype=torch.long)
# words_length = []
for i, idx in enumerate(word_input):
word = I2F[int(idx)]
if word != PAD:
# words_length.append(len(word))
char_input[i] = torch.LongTensor(convert_to_padded_indexes(word, C2I, max_word_len))
# else:
# # doesn't matter because in the word embedding rapper it will skip them.
# words_length.append(1)
return char_input
def make_prefix(index, dicts):
I2F, P2I = dicts.I2F, dicts.P2I
word = I2F[int(index)]
prefix = word[:3]
return P2I[prefix] if prefix in P2I else P2I[UNIQUE_WORD[:3]]
def make_suffix(index, dicts):
I2F, S2I = dicts.I2F, dicts.S2I
word = I2F[int(index)]
prefix = word[-3:]
return S2I[prefix] if prefix in S2I else S2I[UNIQUE_WORD[-3:]]
def make_prefix_suffix_input(input, dicts):
# [[item + 1 for item in list] for list in list_of_lists]
# input shape is (batch_size, num_sequences)
# prefix_input = torch.LongTensor(input.shape)
# suffix_input = torch.LongTensor(len(input), len(input[0]))
prefixes = [[make_prefix(word, dicts) for word in sentence] for sentence in input]
suffixes = [[make_suffix(word, dicts) for word in sentence] for sentence in input]
prefix_input = torch.LongTensor(prefixes)
suffix_input = torch.LongTensor(suffixes)
return prefix_input, suffix_input
def save_model_and_data_sets(model, train_dataset, model_file_path, train_dataset_save_dir):
model.save(model_file_path)
with open(train_dataset_save_dir, 'wb') as file:
pickle.dump(train_dataset, file, pickle.HIGHEST_PROTOCOL)
print('\nData sets and model were saved successfully!')
def load_dataset(data_set_dir):
with open(data_set_dir, 'rb') as file:
data_set = pickle.load(file)
print('\nData set was loaded successfully!')
return data_set