-
Notifications
You must be signed in to change notification settings - Fork 0
/
cbow.py
228 lines (186 loc) · 8.04 KB
/
cbow.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
import csv
import os
from datetime import datetime
from typing import Tuple, Generator, List, Callable, Dict
import torch
from torch import Tensor, nn
from torchtext import data
from torch.utils.tensorboard import SummaryWriter
import torch.optim as optim
import pandas as pd
from torchtext.data import TabularDataset, BucketIterator, get_tokenizer
from torchtext.vocab import Vocab
from tqdm import tqdm
from src.config import CBOWConfig, Config
from src.early_stopping import EarlyStopping
from src.models.cbow import CBOWModel
from src.train_model import TrainingLoop
from src.utils import write_losses, cosine_similarity, train_val_test_split, tokenize_english_text
def load_dataset(output_dir: str) -> Tuple[Tuple[TabularDataset, TabularDataset, TabularDataset], Vocab]:
"""Load train, val, and test datasets."""
target_text = data.LabelField(lower=True)
context_text = data.Field(lower=True)
fields = [
('target', target_text),
('context', context_text),
]
train, val, test = data.TabularDataset.splits(
path=output_dir,
train='train.tsv',
validation='val.tsv',
test='test.tsv',
format='tsv',
fields=fields,
skip_header=True
)
target_text.build_vocab(train.target, train.context, val.target, val.context,
specials=['<pad>', '<unk>'])
context_text.vocab = target_text.vocab
return (train, val, test), target_text.vocab
def get_device() -> torch.device:
"""Get the device depending on the system's compute hardware."""
return torch.device('cuda' if torch.cuda.is_available() else 'cpu')
def extract_eng_from_parallel_text_file() -> None:
"""Extract just the english text from a TSV file containing eng-fra parallel corpus text."""
parallel_text_path = 'data/eng-fra.txt'
df = pd.read_csv(parallel_text_path, sep='\t')
df.columns = ['eng', 'fra']
df[['eng']].to_csv('data/eng.txt', index=False)
class DataManager:
def __init__(self, batch_size: int):
self.batch_size = batch_size
self.train_size = None
self.val_size = None
self.test_size = None
self.train_ds, self.val_ds, self.test_ds = None, None, None
self.train_iterator, self.val_iterator, self.test_iterator = None, None, None
self.source_vocab = None
def load_data(self, input_dir: str):
datasets, self.source_vocab = load_dataset(input_dir)
(self.train_ds, self.val_ds, self.test_ds) = datasets
self.train_iterator, self.val_iterator, self.test_iterator = self.get_iterators(
(self.train_ds, self.val_ds, self.test_ds)
)
self.train_size = len(self.train_ds)
self.val_size = len(self.val_ds)
self.test_size = len(self.test_ds)
def get_iterators(self, datasets: Tuple[TabularDataset, TabularDataset, TabularDataset]
) -> Tuple[BucketIterator, BucketIterator, BucketIterator]:
return data.BucketIterator.splits(
datasets,
# shuffle samples between epochs
shuffle=True,
device=get_device(),
batch_sizes=(self.batch_size, self.batch_size, self.batch_size),
sort_key=lambda x: len(x.context)
)
def save_vocab(self, output_path: str) -> None:
"""Save vocabulary generated during model training."""
with open(output_path, 'w+') as f:
writer = csv.writer(f, delimiter='\t')
writer.writerow(['index', 'token'])
for token, index in self.source_vocab.stoi.items():
writer.writerow([index, token])
@staticmethod
def read_vocab(path: str):
vocab = dict()
with open(path, 'r') as f:
reader = csv.reader(f, delimiter='\t')
next(reader) # skip header
for line in f:
index, token = line[0], line[1]
vocab[token.strip()] = int(index)
return vocab
def create_cbow_features(input_dir: str, output_dir: str, context_size: int) -> None:
"""Creates CBOW features by identifying surrounding context words."""
assert context_size % 2 == 0, "Context size should be even so we pick the same number of words" \
" to the left and right of the target word."
# window_size is the number of words to the left and right
for filename in ['train.tsv', 'val.tsv', 'test.tsv']:
window_size = context_size // 2
with open(os.path.join(input_dir, filename)) as f_in, \
open(os.path.join(output_dir, filename), 'w') as f_out:
reader = csv.reader(f_in, delimiter='\t')
writer = csv.writer(f_out, delimiter='\t')
next(reader) # skip the header
lines = [line[0].strip().split() for line in reader]
for words in lines:
for i, target in enumerate(words):
words_to_the_left = words[max(0, i - window_size):i]
words_to_the_right = words[i + 1:i + window_size + 1]
padded_left = ['<pad>'] * (window_size - len(words_to_the_left)) + words_to_the_left
padded_right = words_to_the_right + ['<pad>'] * (window_size - len(words_to_the_right))
context = padded_left + padded_right
assert len(context) == window_size * 2
writer.writerow([target, ' '.join(context)])
def cbow_preprocessing(config: CBOWConfig):
# Split into train/val/test
train_val_test_split(config.raw_data_path, config.data_dir)
# Tokenize files in data_dir
for filename in ['test.tsv', 'val.tsv', 'train.tsv']:
data_path = os.path.join(config.data_dir, filename)
tokenize_english_text(data_path, data_path)
# Create features
create_cbow_features(config.data_dir, config.features_dir, config.context_size)
class CBOWTrainingLoop(TrainingLoop):
def compute_loss_from_batch(self, batch) -> Tensor:
# Do a forward pass and compute the loss
return self.model.compute_loss(batch.context, batch.target)
def train():
conf = CBOWConfig()
cbow_preprocessing(conf)
# Load data and compute vocabulary
# TODO: remove sentences that exceed the maximum sequence length
data_manager = DataManager(conf.batch_size)
data_manager.load_data(conf.features_dir)
data_manager.save_vocab(conf.model_vocab_path)
# Initialize model
model = CBOWModel(
vocab_size=len(data_manager.source_vocab),
embedding_dim=conf.embedding_dim,
context_size=conf.context_size
)
# Initialize optimizer
# TODO: move LR to config
optimizer = optim.Adam(model.parameters(), lr=0.001, weight_decay=1e-4)
loop = CBOWTrainingLoop(conf, model, optimizer, data_manager.train_iterator,
data_manager.val_iterator)
loop.train()
write_losses(os.path.join(conf.model_checkpoint_dir, 'losses.csv'),
loop.train_losses, loop.val_losses)
class InferenceServer:
def __init__(self, config: Config, model: nn.Module, vocab: Dict[str, int]):
self.config = config
self.model = model
self.vocab = vocab
self.oov_index = self.vocab['<oov>']
def embed_tokens(self, tokens: List[str]) -> Tensor:
sentence_indices = [self.vocab.get(word, self.oov_index) for word in tokens]
sentence_tensor = torch.LongTensor(sentence_indices)
return self.model.embeddings(sentence_tensor)
def embed_word(self, word: str) -> Tensor:
return self.embed_tokens([word]).squeeze()
def calculate_similarity(self, word1: str, word2: str,
similarity_func: Callable = cosine_similarity) -> float:
e1 = self.embed_word(word1)
e2 = self.embed_word(word2)
return similarity_func(e1, e2).item()
def load_cbow_model(model_path: str, vocab: Dict[str, int], config: Config):
model = CBOWModel(
vocab_size=len(vocab),
embedding_dim=config.embedding_dim,
context_size=config.context_size
)
model_info = torch.load(model_path)
model.load_state_dict(model_info['model_state_dict'])
model.eval()
return model
def inference():
config = CBOWConfig(model_date=get_latest_model(CBOWConfig.model_root_dir))
data_manager = DataManager(config.batch_size)
vocab = data_manager.read_vocab(config.model_vocab_path)
model = load_cbow_model(config.model_best_checkpoint_path, vocab, config)
server = InferenceServer(config, model, vocab)
word1, word2 = 'hello', 'hi'
similarity = server.calculate_similarity(word1, word2, similarity_func=cosine_similarity)
print(f'Similarity between {word1} and {word2} is {similarity}')