-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtrain_base.py
232 lines (214 loc) · 7.81 KB
/
train_base.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
import argparse
from pathlib import Path
import flair
import torch
from flair.data import Corpus
from flair.models import SequenceTagger
from flair.embeddings import (
FastTextEmbeddings,
StackedEmbeddings,
FlairEmbeddings,
TokenEmbeddings,
CharacterEmbeddings,
)
from flair.trainers import ModelTrainer
from train_grid import UD_UKRAINIAN
from torch.optim.adam import Adam
def choochoo(
hidden_size: int,
rnn_layers: int,
embeddings: TokenEmbeddings,
config_name: str,
optimize_lr: bool = False,
learning_rate: float = 0.1,
mini_batch_size: int = 32,
dropout: float = 0.0,
) -> None:
corpus: Corpus = UD_UKRAINIAN()
# 2. what tag do we want to predict?
tag_type = "upos"
# 3. make the tag dictionary from the corpus
tag_dictionary = corpus.make_label_dictionary(tag_type)
tagger: SequenceTagger = SequenceTagger(
hidden_size=hidden_size,
rnn_layers=rnn_layers,
embeddings=embeddings(),
tag_dictionary=tag_dictionary,
tag_type=tag_type,
use_crf=True,
dropout=dropout,
)
# 6. initialize trainer
trainer = ModelTrainer(tagger, corpus)
results_path = Path(f"./pos-tests/{config_name}/")
checkpoint_path = results_path / "checkpoint.pt"
tensorboard_path = results_path / "tensorboard"
tensorboard_path.mkdir(parents=True, exist_ok=True)
if checkpoint_path.exists():
trained_model = SequenceTagger.load(checkpoint_path)
trainer.resume(
trained_model,
base_path=results_path,
learning_rate=learning_rate,
mini_batch_size=mini_batch_size,
checkpoint=True,
train_with_dev=True,
monitor_test=True,
max_epochs=150,
embeddings_storage_mode="cpu",
use_tensorboard=True,
tensorboard_log_dir=tensorboard_path,
tensorboard_comment=f"Flair UK: {config_name}",
)
else:
# 7. start training
if optimize_lr:
trainer.find_learning_rate(results_path, Adam)
else:
trainer.train(
results_path,
learning_rate=learning_rate,
mini_batch_size=mini_batch_size,
checkpoint=True,
train_with_dev=True,
monitor_test=True,
max_epochs=150,
embeddings_storage_mode="cpu",
use_tensorboard=True,
tensorboard_log_dir=tensorboard_path,
tensorboard_comment=f"Flair UK: {config_name}",
)
if __name__ == "__main__":
flair.device = torch.device("cuda:0")
parser = argparse.ArgumentParser(
description="""That is the simple trainer that can accept a base dir
with embeddings and the name of the config to train the model"""
)
parser.add_argument("--embeddings-dir", type=Path, help="Path base dir with embeddings", default=Path("/data/"))
parser.add_argument("config")
args = parser.parse_args()
config = {
"fb.fasttext": {
"embeddings": lambda: FastTextEmbeddings(args.embeddings_dir / "fasttext/uk/cc.uk.300.bin"),
"hidden_size": 256,
"rnn_layers": 1,
},
"uk.flairembeddings": {
"embeddings": lambda: StackedEmbeddings(
[
FlairEmbeddings(args.embeddings_dir / "flair/uk/backward/best-lm.pt"),
FlairEmbeddings(args.embeddings_dir / "flair/uk/forward/best-lm.pt"),
]
),
"hidden_size": 256,
"rnn_layers": 1,
},
"uk.flairembeddings.lr0.5": {
"embeddings": lambda: StackedEmbeddings(
[
FlairEmbeddings(args.embeddings_dir / "flair/uk/backward/best-lm.pt"),
FlairEmbeddings(args.embeddings_dir / "flair/uk/forward/best-lm.pt"),
]
),
"hidden_size": 256,
"rnn_layers": 1,
"learning_rate": 0.5,
},
"uk.flairembeddings.find_lr": {
"embeddings": lambda: StackedEmbeddings(
[
FlairEmbeddings(args.embeddings_dir / "flair/uk/backward/best-lm.pt"),
FlairEmbeddings(args.embeddings_dir / "flair/uk/forward/best-lm.pt"),
]
),
"hidden_size": 256,
"rnn_layers": 1,
"optimize_lr": True,
},
"uk.flairembeddings.fasttext": {
"embeddings": lambda: StackedEmbeddings(
[
FlairEmbeddings(args.embeddings_dir / "flair/uk/backward/best-lm.pt"),
FlairEmbeddings(args.embeddings_dir / "flair/uk/forward/best-lm.pt"),
FastTextEmbeddings(args.embeddings_dir / "fasttext/uk/cc.uk.300.bin"),
]
),
"hidden_size": 256,
"rnn_layers": 1,
},
"uk.flairembeddings.x2": {
"embeddings": lambda: StackedEmbeddings(
[
FlairEmbeddings(args.embeddings_dir / "flair/uk/backward/best-lm.pt"),
FlairEmbeddings(args.embeddings_dir / "flair/uk/forward/best-lm.pt"),
]
),
"hidden_size": 256,
"rnn_layers": 2,
},
"uk.flairembeddings.charembeddings": {
"embeddings": lambda: StackedEmbeddings(
[
CharacterEmbeddings(),
FlairEmbeddings(args.embeddings_dir / "flair/uk/backward/best-lm.pt"),
FlairEmbeddings(args.embeddings_dir / "flair/uk/forward/best-lm.pt"),
]
),
"hidden_size": 256,
"rnn_layers": 1,
},
"uk.flairembeddings.large": {
"embeddings": lambda: StackedEmbeddings(
[
FlairEmbeddings(args.embeddings_dir / "flair/uk/backward/best-lm.pt"),
FlairEmbeddings(args.embeddings_dir / "flair/uk/forward/best-lm.pt"),
]
),
"hidden_size": 512,
"rnn_layers": 1,
},
"uk.flairembeddings.xlarge": {
"embeddings": lambda: StackedEmbeddings(
[
FlairEmbeddings(args.embeddings_dir / "flair/uk/backward/best-lm.pt"),
FlairEmbeddings(args.embeddings_dir / "flair/uk/forward/best-lm.pt"),
]
),
"hidden_size": 1024,
"rnn_layers": 1,
},
"uk.flairembeddings.champ": {
"embeddings": lambda: StackedEmbeddings(
[
FlairEmbeddings(args.embeddings_dir / "flair/uk/backward/best-lm.pt"),
FlairEmbeddings(args.embeddings_dir / "flair/uk/forward/best-lm.pt"),
]
),
"hidden_size": 256,
"rnn_layers": 1,
"learning_rate": 0.1,
"mini_batch_size": 16,
"dropout": 0.005334913013756493,
},
}
# evaluation run 10
# dropout: 0.09721289436761998
# embeddings: StackedEmbeddings [0-/data/flair/uk/backward/best-lm.pt,1-/data/flair/uk/forward/best-lm.pt]
# hidden_size: 128
# learning_rate: 0.25
# mini_batch_size: 16
# rnn_layers: 2
# loss: 0.07583423372772007
# variance: 5.048710420923001e-12
# test_score: 0.9765132040196307
# evaluation run 2
# dropout: 0.005334913013756493
# embeddings: StackedEmbeddings [0-/data/flair/uk/backward/best-lm.pt,1-/data/flair/uk/forward/best-lm.pt]
# hidden_size: 256
# learning_rate: 0.1
# mini_batch_size: 16
# rnn_layers: 1
# loss: 0.07414293454753029
# variance: 6.252252625092456e-11
# test_score: 0.9780322505258238
choochoo(config_name=args.config, **config[args.config])