-
Notifications
You must be signed in to change notification settings - Fork 0
/
word2vec.py
436 lines (350 loc) · 14.2 KB
/
word2vec.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
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
"""
Word2Vec from Scratch
All in one file
"""
from dataclasses import dataclass
from collections import Counter
import pickle
import os
import re
import time
import json
from datasets import load_dataset
import numpy as np
import torch
from torch import nn
from torch.utils.data import DataLoader
from torch import optim
from torch.optim.lr_scheduler import LambdaLR
##########################################################################
# A couple classes for typing hygiene
##########################################################################
@dataclass
class ModelConfig:
"""class to store Model configuration"""
name: str
model: nn.Module
n_words: int
@dataclass
class Vocab:
"""class to store Vocabulary"""
word_to_idx: dict
idx_to_word: list
##########################################################################
# Some configurable hyper-parameters
##########################################################################
# constants to pre-process data and build vocab
MIN_WORD_FREQUENCY = 50
MAX_SEQUENCE_LENGTH = 256
# embedding constants
EMBED_DIMENSION = 300
EMBED_MAX_NORM = 1 # keep embedding values low
##########################################################################
# The actual models, which are very simple
##########################################################################
class SkipGramModel(nn.Module):
"""
Implementation of SkipGram model described in paper:
https://arxiv.org/abs/1301.3781
Takes a single word
Returns logprobs for all possible neighbors
inputs_ is a one_hot encoding of a word in vocabulary
inputs_ is 1 by n; embeddings is n by m; linear is m by n
input_ dot embeddings is 1 by m => x
and x dot linear is 1 by n
"""
def __init__(self, vocab_size: int):
super().__init__()
# n by m Matrix
self.embeddings = nn.Embedding(
num_embeddings=vocab_size, # n
embedding_dim=EMBED_DIMENSION, # m
max_norm=EMBED_MAX_NORM,
)
# m by n Matrix
self.linear = nn.Linear(
in_features=EMBED_DIMENSION, # m
out_features=vocab_size, # n
)
def forward(self, inputs_):
"""forward pass"""
x = self.embeddings(inputs_)
x = self.linear(x)
return x
class CBOWModel(SkipGramModel):
"""
Implementation of CBOW model described in paper:
https://arxiv.org/abs/1301.3781
Note that this class extends SkipGram because the
network is identical with the exception that the input
in this case is a group of words instead of one
which is handled differently in the forward pass
by averaging the embeddings for each input word
Takes N*2 neighbors from each side
Returns logprobs for entire vocabulary
inputs_ is a one_hot encoding of BOWs (e.g. N_WORDS * 2) in vocabulary
inputs_ is (N_WORDS * 2) by n; embeddings is n by m; linear is m by n
input_ dot embeddings is (N_WORDS * 2) by m => x
we then take the mean of x over the 1st axis (e.g the m'th axis or columns)
then x becomes 1 by m
and x dot linear is 1 by n
"""
def forward(self, inputs_):
"""forward pass"""
x = self.embeddings(inputs_)
x = x.mean(axis=1)
x = self.linear(x)
return x
##########################################################################
# The data loader fetches dataset and returns an iterable dataset
# It also builds and returns the vocabulary
##########################################################################
@dataclass
class W2VDataLoader:
"""builds dataloader and vocabulary"""
model_config: ModelConfig
batch_size: int
ds_name: str
shuffle: bool = True
vocab: Vocab = None
model_dir: str = None
def __post_init__(self):
self.tokenizer = re.compile(r"\w+")
self.dataset = load_dataset("wikitext", self.ds_name)
if self.vocab is None:
self.build_vocab()
self.save_vocab()
self.train = DataLoader(
self.dataset["train"],
batch_size=self.batch_size,
shuffle=self.shuffle,
collate_fn=self.collate,
)
self.val = DataLoader(
self.dataset["validation"],
batch_size=self.batch_size,
shuffle=self.shuffle,
collate_fn=self.collate,
)
def tokenize(self, text):
"""basic tokenizer"""
# the pervious solution was word_tokenizer from nltk
return self.tokenizer.findall(text.lower())
def build_vocab(self):
"""Builds vocabulary from dataset"""
word_freq = Counter()
all_tokens = []
for text in self.dataset["train"]["text"]:
# could add tokenize(text) directly to
# word_freq.update() but it's a bit slower
all_tokens.extend(self.tokenize(text))
word_freq.update(all_tokens)
vocab = {
word: freq for word, freq in word_freq.items() if freq >= MIN_WORD_FREQUENCY
}
vocab["unk"] = 10**18 # very large number so it's idx=0
word_to_idx = {
word: idx
for idx, (word, freq) in enumerate(
# sort by the [1]st element in the tuple ('word', freq)
sorted(vocab.items(), key=lambda item: item[1], reverse=True),
)
}
idx_to_word = [word for word, idx in word_to_idx.items()]
self.vocab = Vocab(word_to_idx, idx_to_word)
def save_vocab(self):
"""save vocabulary to disk"""
if self.model_dir:
# save vocabulary to disk
vocab_path = os.path.join(self.model_dir, "vocab.pkl")
with open(vocab_path, "wb") as file:
pickle.dump(self.vocab, file)
def collate(self, batch):
"""
Prepares examples for trainer
`batch` is a list of text paragraphs
Long paragraphs will be truncated to MAX_SEQUENCE_LENGTH
Context for i-th word is model_config.n_words before and after
if model_config.name == "CBOWModel":
Each element in `batch_input` is model_config.n_words * 2 words
Each element in `batch_output` is the middle / i-th word
elif model_config.name == "SkipGramModel":
Each element in `batch_input` is the middle / i-th word
Each element in `batch_output` is one of model_config.n_words* 2 context words
"""
batch_input, batch_output = [], []
for b in batch:
text_tokens_ids = [
self.vocab.word_to_idx.get(word, 0) for word in self.tokenize(b["text"])
]
# if the sentence is too short to extract CBOW, skip
if len(text_tokens_ids) < self.model_config.n_words * 2 + 1:
continue
if MAX_SEQUENCE_LENGTH:
text_tokens_ids = text_tokens_ids[:MAX_SEQUENCE_LENGTH]
# iterate for as long as the sliding window allows
for idx in range(len(text_tokens_ids) - self.model_config.n_words * 2):
token_id_sequence = text_tokens_ids[
idx : (idx + self.model_config.n_words * 2 + 1)
]
if self.model_config.name == "CBOWModel":
output_ = token_id_sequence.pop(
self.model_config.n_words
) # get the middle word
inputs_ = token_id_sequence # the remaining words
batch_input.append(inputs_)
batch_output.append(output_)
elif self.model_config.name == "SkipGramModel":
input_ = token_id_sequence.pop(
self.model_config.n_words
) # get the middle word
outputs_ = token_id_sequence # the remaining words
# make one example per output
for output_ in outputs_:
batch_input.append(input_)
batch_output.append(output_)
batch_input = torch.tensor(batch_input, dtype=torch.long)
batch_output = torch.tensor(batch_output, dtype=torch.long)
return batch_input, batch_output
##########################################################################
# The Trainer class orchestrates the trainer.
# This is where GPUs do most of the work (e.g. gradient descent)
##########################################################################
@dataclass
# pylint: disable=too-many-instance-attributes
class Trainer:
"""Main class for model training"""
model: nn.Module
epochs: int
learning_rate: float
dataloader: W2VDataLoader
model_dir: str
train_steps: int = None # limit the number of steps
val_steps: int = None # limit the number of steps
checkpoint_frequency: int = None # number of steps between saves
def __post_init__(self):
self.loss = {"train": [], "val": []}
self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
# if torch.cuda.device_count() > 1:
# self.model = nn.DataParallel(self.model)
# print(f"Using {torch.cuda.device_count()} GPUs")
self.criterion = nn.CrossEntropyLoss() # loss function
# set up optimizer (e.g. gradient descent calculation) and learning rate scheduler
self.optimizer = optim.Adam(self.model.parameters(), lr=self.learning_rate)
self.lr_scheduler = LambdaLR(
self.optimizer,
# variable learning rate that ends at 0 in the last epoch
lr_lambda=lambda epoch: (self.epochs - epoch) / self.epochs,
verbose=True,
)
self.model.to(self.device)
def train(self):
"""Train model for `self.epochs` epochs"""
start_time = time.time()
print(f"running on {self.device}")
for epoch in range(self.epochs):
self._train_epoch()
self._validate_epoch()
current_time = time.time()
print(
f"""Epoch: {epoch+1}/{self.epochs}
Train Loss={self.loss['train'][-1]:.5f},
Val Loss={self.loss['val'][-1]:.5f}
Seconds Elapsed={current_time-start_time}"""
)
self.lr_scheduler.step() # adjust learning rate
if self.checkpoint_frequency:
self._save_checkpoint(epoch)
def _train_epoch(self):
self.model.train() # train mode for model
running_loss = []
batches = len(self.dataloader.train)
for i, batch_data in enumerate(self.dataloader.train, 1):
inputs = batch_data[0].to(self.device)
labels = batch_data[1].to(self.device)
self.optimizer.zero_grad() # clear gradient
outputs = self.model(inputs) # generate predictions from current model
loss = self.criterion(outputs, labels) # calculate loss
loss.backward() # calculate derivative
self.optimizer.step() # apply gradient descent to model
running_loss.append(loss.item())
if i % int(batches / 25) == 0:
print(f"Batch: {i} / {batches}")
if i == self.train_steps:
break
epoch_loss = np.mean(running_loss)
self.loss["train"].append(epoch_loss)
def _validate_epoch(self):
self.model.eval()
running_loss = []
with torch.no_grad():
for i, batch_data in enumerate(self.dataloader.val, 1):
inputs = batch_data[0].to(self.device)
labels = batch_data[1].to(self.device)
outputs = self.model(inputs)
loss = self.criterion(outputs, labels)
running_loss.append(loss.item())
if i == self.val_steps:
break
epoch_loss = np.mean(running_loss)
self.loss["val"].append(epoch_loss)
def _save_checkpoint(self, epoch):
"""Save model checkpoint to `self.model_dir` directory"""
epoch_num = epoch + 1
if epoch_num % self.checkpoint_frequency == 0:
model_path = f"checkpoint_{str(epoch_num).zfill(3)}.pt"
model_path = os.path.join(self.model_dir, model_path)
torch.save(self.model, model_path)
def save_model(self):
"""Save final model to `self.model_dir` directory"""
model_path = os.path.join(self.model_dir, "model.pt")
torch.save(self.model, model_path)
def save_loss(self):
"""Save train/val loss as json to `self.model_dir` directory"""
loss_path = os.path.join(self.model_dir, "loss.json")
with open(loss_path, "w", encoding="utf-8") as file:
json.dump(self.loss, file)
##########################################################################
# Short routine to initialize model and dataloader, and run the trainer
##########################################################################
MODEL_CONFIG = ModelConfig(name="CBOWModel", model=CBOWModel, n_words=4)
# MODEL_CONFIG = ModelConfig(name="SkipGramModel", model=SkipGramModel, n_words=4)
DS_NAME = "wikitext-2-v1" # "wikitext-103-v1" (large) , "wikitext-2-v1" (small)
def train():
"""main function to coordinate training"""
model_config = MODEL_CONFIG
ds_name = DS_NAME
model_dir = f"{model_config.name}_{ds_name}_data"
# these could be moved to top as global hyper-parameters
batch_size = 96
shuffle = True
learning_rate = 0.025
epochs = 5
checkpoint_frequency = 5_000
if not os.path.exists(model_dir):
os.makedirs(model_dir)
dataloader = W2VDataLoader(
batch_size=batch_size,
ds_name=ds_name,
shuffle=shuffle,
model_config=model_config,
model_dir=model_dir,
)
vocab_size = len(dataloader.vocab.word_to_idx)
print(f"Vocabulary size: {vocab_size}")
# initialize model instance
model = model_config.model(vocab_size=vocab_size)
trainer = Trainer(
model=model,
model_dir=model_dir,
epochs=epochs,
learning_rate=learning_rate,
dataloader=dataloader,
checkpoint_frequency=checkpoint_frequency,
)
trainer.train()
print("Training Finished.")
trainer.save_model()
trainer.save_loss()
if __name__ == "__main__":
train()