-
Notifications
You must be signed in to change notification settings - Fork 0
/
main.py
153 lines (131 loc) · 4.94 KB
/
main.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
import os
import sklearn
import torch
from torch import nn
import pandas
import pickle
import preprocess
import word2vec
import rnn
import cnn
from trainer import Trainer
def get_embedding_model(
embedding_dim: int, vocab_size: int, device: str
) -> (nn.Module, bool):
embedding_model_name = os.getenv("embedding_model", None)
if embedding_model_name is None:
return nn.Embedding(vocab_size, embedding_dim), True
elif embedding_model_name == "CBOW":
CONTEXT_SIZE = 4
train_word2vec_data_loader = word2vec.build_cbow_dataset(
train_dataset["tokens"].values.tolist(),
CONTEXT_SIZE,
BATCH_SIZE,
)
print("Train CBOW dataset is ready")
# val_word2vec_data_loader = word2vec.build_cbow_dataset(
# val_dataset["tokens"].values.tolist(), CONTEXT_SIZE, BATCH_SIZE
# )
print("Validate CBOW dataset is ready")
elif embedding_model_name == "skip-gram":
CONTEXT_SIZE = 4
train_word2vec_data_loader = word2vec.build_skip_gram_dataset(
train_dataset["tokens"].values.tolist(),
CONTEXT_SIZE,
BATCH_SIZE,
)
print("Train skip-gram dataset is ready")
# val_word2vec_data_loader = word2vec.build_skip_gram_dataset(
# val_dataset["tokens"].values.tolist(), CONTEXT_SIZE, BATCH_SIZE
# )
print("Validate skip-gram dataset is ready")
else:
print(f"There is no dataset for {embedding_model_name} word2vec model")
val_word2vec_data_loader = None # not used so far
# TODO: refactor it
word2vec_trainer = word2vec.Word2VecTrainer(
train_word2vec_data_loader,
val_word2vec_data_loader,
model_name=embedding_model_name,
vocab_size=vocab_size,
embedding_dim=embedding_dim,
context_size=CONTEXT_SIZE,
device=device,
)
word2vec_trainer.train(10)
return word2vec_trainer.model.emb, False
def get_model():
model_name = os.getenv("model_name", "rnn")
if model_name == "rnn":
return rnn.RNNModel(
embedding_model=embedding_model,
embedding_dim=EMBEDDING_DIM,
rnn_hidden_size=64,
rnn_num_layers=2,
rnn_dropout=0.8,
linear_sizes=[],
attention_heads=None,
)
elif model_name == "cnn":
return cnn.CNNModel(
embedding_model=embedding_model,
embedding_dim=EMBEDDING_DIM,
region_sizes=[2, 3, 4, 5],
feature_maps=100,
dropout=0.5,
)
else:
print(f"There is no a model named {model_name}")
exit()
def get_datasets():
cached_dataset = os.getenv("cached_dataset", True)
if cached_dataset:
try:
train_dataset = pandas.read_pickle("train_dataset.pkl")
val_dataset = pandas.read_pickle("val_dataset.pkl")
tokenizer = pickle.load(open("tokenizer.pkl", "rb"))
except FileNotFoundError as e:
print(
f'Caught exception while reading cached datasets "{e}".\nFallback to creating datasets'
)
clean_dataset = preprocess.get_prepared_dataset()
train_dataset, val_dataset = sklearn.model_selection.train_test_split(
clean_dataset, test_size=0.2
)
tokenizer = preprocess.FreqTokenizer(
train_dataset["preprocessed_text"].values.tolist()
)
preprocess.add_tokens(train_dataset, tokenizer)
preprocess.add_tokens(val_dataset, tokenizer)
train_dataset.to_pickle("train_dataset.pkl")
val_dataset.to_pickle("val_dataset.pkl")
pickle.dump(tokenizer, open("tokenizer.pkl", "wb"))
else:
print("Successfully loaded cached datasets")
return train_dataset, val_dataset, tokenizer.vocab_size
if __name__ == "__main__":
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
BATCH_SIZE = 512
print(f"Device = {DEVICE}")
train_dataset, val_dataset, vocab_size = get_datasets()
train_data_loader = preprocess.create_data_loader(train_dataset, BATCH_SIZE, DEVICE)
val_data_loader = preprocess.create_data_loader(val_dataset, BATCH_SIZE, DEVICE)
print("Datasets are ready")
EMBEDDING_DIM = 16
embedding_model, embedding_train = get_embedding_model(
EMBEDDING_DIM, vocab_size, DEVICE
)
model = get_model()
rnn_trainer = Trainer(
train_data_loader=train_data_loader,
val_data_loader=val_data_loader,
model=model,
embedding_train=embedding_train,
device=DEVICE,
)
print(
f"Model has {sum(p.numel() for p in rnn_trainer.model.parameters() if p.requires_grad)} parameters"
)
print(f"Initial validation_score = {rnn_trainer.validate(use_tqdm=False)}")
rnn_trainer.train(max_epochs=64, target_validation_score=90)
print(f"Final validation_score = {rnn_trainer.validate(use_tqdm=False)}")