-
Notifications
You must be signed in to change notification settings - Fork 3
/
pretrain.py
295 lines (265 loc) · 7.87 KB
/
pretrain.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
"""Train the model."""
import argparse
import os
import sys
from typing import Any, Dict
import pytorch_lightning as pl
import torch
from lightning.pytorch.loggers import WandbLogger
from pytorch_lightning.callbacks import LearningRateMonitor, ModelCheckpoint
from pytorch_lightning.strategies.ddp import DDPStrategy
from sklearn.model_selection import train_test_split
from torch.utils.data import DataLoader
from odyssey.data.dataset import PretrainDataset, PretrainDatasetDecoder
from odyssey.data.tokenizer import ConceptTokenizer
from odyssey.models.cehr_bert.model import BertPretrain
from odyssey.models.cehr_big_bird.model import BigBirdPretrain
from odyssey.models.ehr_mamba.model import MambaPretrain
from odyssey.models.model_utils import (
get_run_id,
load_config,
load_pretrain_data,
)
from odyssey.utils.utils import seed_everything
def main(args: argparse.Namespace, model_config: Dict[str, Any]) -> None:
"""Train the model."""
seed_everything(args.seed)
os.environ["CUDA_LAUNCH_BLOCKING"] = "1"
torch.cuda.empty_cache()
torch.set_float32_matmul_precision("medium")
pre_data = load_pretrain_data(
args.data_dir,
args.sequence_file,
args.id_file,
)
# Split data
pre_train, pre_val = train_test_split(
pre_data,
test_size=args.val_size,
random_state=args.seed,
)
# Train Tokenizer
tokenizer = ConceptTokenizer(data_dir=args.vocab_dir)
tokenizer.fit_on_vocab()
# Load datasets
if args.is_decoder: # e.g. Mamba
train_dataset = PretrainDatasetDecoder(
data=pre_train,
tokenizer=tokenizer,
max_len=args.max_len,
)
val_dataset = PretrainDatasetDecoder(
data=pre_val,
tokenizer=tokenizer,
max_len=args.max_len,
)
else:
train_dataset = PretrainDataset(
data=pre_train,
tokenizer=tokenizer,
max_len=args.max_len,
mask_prob=args.mask_prob,
)
val_dataset = PretrainDataset(
data=pre_val,
tokenizer=tokenizer,
max_len=args.max_len,
mask_prob=args.mask_prob,
)
train_loader = DataLoader(
train_dataset,
batch_size=args.batch_size,
num_workers=args.num_workers,
persistent_workers=args.persistent_workers,
shuffle=True,
pin_memory=args.pin_memory,
)
val_loader = DataLoader(
val_dataset,
batch_size=args.batch_size,
num_workers=args.num_workers,
persistent_workers=args.persistent_workers,
pin_memory=args.pin_memory,
)
callbacks = [
ModelCheckpoint(
monitor="val_loss",
mode="min",
filename="best",
save_top_k=1,
save_last=True,
verbose=True,
dirpath=args.checkpoint_dir,
),
LearningRateMonitor(logging_interval="step"),
]
# Create model
if args.model_type == "cehr_bert":
model = BertPretrain(
args=args,
vocab_size=tokenizer.get_vocab_size(),
padding_idx=tokenizer.get_pad_token_id(),
**model_config,
)
elif args.model_type == "cehr_bigbird":
model = BigBirdPretrain(
vocab_size=tokenizer.get_vocab_size(),
padding_idx=tokenizer.get_pad_token_id(),
**model_config,
)
elif args.model_type == "ehr_mamba":
model = MambaPretrain(
vocab_size=tokenizer.get_vocab_size(),
padding_idx=tokenizer.get_pad_token_id(),
cls_idx=tokenizer.get_class_token_id(),
**model_config,
)
run_id = get_run_id(args.checkpoint_dir)
wandb_logger = WandbLogger(
project=args.exp_name,
save_dir=args.log_dir,
entity=args.workspace_name,
id=run_id,
resume="allow",
)
# Setup PyTorchLightning trainer
trainer = pl.Trainer(
accelerator="gpu",
num_nodes=args.nodes,
devices=args.gpus,
strategy=DDPStrategy(find_unused_parameters=True)
if args.gpus > 1
else "auto", # DeepSpeedStrategy(stage=2, offload_optimizer=False)
precision="16-mixed",
check_val_every_n_epoch=1,
max_epochs=args.max_epochs,
callbacks=callbacks,
deterministic=False,
enable_checkpointing=True,
enable_progress_bar=True,
enable_model_summary=True,
logger=wandb_logger,
log_every_n_steps=args.log_every_n_steps,
accumulate_grad_batches=args.acc,
gradient_clip_val=1.0,
)
# Train the model
trainer.fit(
model=model,
train_dataloaders=train_loader,
val_dataloaders=val_loader,
ckpt_path=args.resume_checkpoint,
)
if __name__ == "__main__":
parser = argparse.ArgumentParser()
# project configuration
parser.add_argument(
"--model-type",
type=str,
required=True,
help="Model type: 'cehr_bert' or 'cehr_bigbird' or 'ehr_mamba'",
)
parser.add_argument(
"--exp-name",
type=str,
required=True,
help="Path to model config file",
)
parser.add_argument(
"--workspace-name",
type=str,
default=None,
help="Name of the Wandb workspace",
)
parser.add_argument(
"--config-dir",
type=str,
required=True,
help="Path to model config file",
)
parser.add_argument(
"--is-decoder",
type=bool,
default=False,
help="Is the model a decoder (e.g. Mamba) or not",
)
# data-related arguments
parser.add_argument(
"--data-dir",
type=str,
required=True,
help="Path to the data directory",
)
parser.add_argument(
"--sequence-file",
type=str,
required=True,
help="Path to the patient sequence file",
)
parser.add_argument(
"--id-file",
type=str,
required=True,
help="Path to the patient id file",
)
parser.add_argument(
"--vocab-dir",
type=str,
required=True,
help="Path to the vocabulary directory of json files",
)
parser.add_argument(
"--val-size",
type=float,
default=0.1,
help="Validation set size for splitting the data",
)
# checkpointing and loggig arguments
parser.add_argument(
"--checkpoint-dir",
type=str,
required=True,
help="Path to the checkpoint directory",
)
parser.add_argument(
"--log_dir",
type=str,
default="logs",
help="Path to the log directory",
)
parser.add_argument(
"--resume_checkpoint",
type=str,
default=None,
help="Checkpoint to resume pretraining from",
)
parser.add_argument(
"--log_every_n_steps",
type=int,
default=10,
help="Number of steps to log the training",
)
# Other arguments
parser.add_argument(
"--seed",
type=int,
default=42,
help="Random seed for reproducibility",
)
args = parser.parse_args()
if args.model_type not in ["cehr_bert", "cehr_bigbird", "ehr_mamba"]:
print(
"Invalid model type. Choose 'cehr_bert' or 'cehr_bigbird' or 'ehr_mamba'."
)
sys.exit(1)
args.checkpoint_dir = os.path.join(args.checkpoint_dir, args.exp_name)
os.makedirs(args.checkpoint_dir, exist_ok=True)
os.makedirs(args.log_dir, exist_ok=True)
config = load_config(args.config_dir, args.model_type)
train_config = config["train"]
for key, value in train_config.items():
if not hasattr(args, key) or getattr(args, key) is None:
setattr(args, key, value)
model_config = config["model"]
args.max_len = model_config["max_seq_length"]
main(args, model_config)