forked from DeepPSP/torch_ecg
-
Notifications
You must be signed in to change notification settings - Fork 0
/
trainer.py
482 lines (436 loc) · 16 KB
/
trainer.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
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
"""
"""
import argparse
import os
import sys
from copy import deepcopy
from pathlib import Path
from typing import Any, Dict, List, NoReturn, Optional, Tuple
import numpy as np
import torch
from torch import nn
from torch.nn.parallel import DataParallel as DP
from torch.nn.parallel import DistributedDataParallel as DDP # noqa: F401
from torch.utils.data import DataLoader, Dataset
try:
import torch_ecg # noqa: F401
except ModuleNotFoundError:
sys.path.insert(0, str(Path(__file__).absolute().parents[2]))
from aux_metrics import (
compute_main_task_metric,
compute_rpeak_metric,
compute_rr_metric,
)
from cfg import BaseCfg, ModelCfg, TrainCfg
from dataset import CPSC2021
from model import ( # noqa: F401
ECG_SEQ_LAB_NET_CPSC2021,
ECG_SUBTRACT_UNET_CPSC2021,
ECG_UNET_CPSC2021,
RR_LSTM_CPSC2021,
_qrs_detection_post_process,
)
from torch_ecg.cfg import CFG
from torch_ecg.components.trainer import BaseTrainer
from torch_ecg.utils.misc import str2bool
from torch_ecg.utils.utils_data import mask_to_intervals
from torch_ecg.utils.utils_nn import default_collate_fn as collate_fn
if BaseCfg.torch_dtype == torch.float64:
torch.set_default_tensor_type(torch.DoubleTensor)
__all__ = [
"CPSC2021Trainer",
]
class CPSC2021Trainer(BaseTrainer):
""" """
__DEBUG__ = True
__name__ = "CPSC2021Trainer"
def __init__(
self,
model: nn.Module,
model_config: dict,
train_config: dict,
device: Optional[torch.device] = None,
lazy: bool = True,
**kwargs: Any,
) -> NoReturn:
"""
Parameters
----------
model: Module,
the model to be trained
model_config: dict,
the configuration of the model,
used to keep a record in the checkpoints
train_config: dict,
the configuration of the training,
including configurations for the data loader, for the optimization, etc.
will also be recorded in the checkpoints.
`train_config` should at least contain the following keys:
"monitor": str,
"loss": str,
"n_epochs": int,
"batch_size": int,
"learning_rate": float,
"lr_scheduler": str,
"lr_step_size": int, optional, depending on the scheduler
"lr_gamma": float, optional, depending on the scheduler
"max_lr": float, optional, depending on the scheduler
"optimizer": str,
"decay": float, optional, depending on the optimizer
"momentum": float, optional, depending on the optimizer
device: torch.device, optional,
the device to be used for training
lazy: bool, default True,
whether to initialize the data loader lazily
"""
super().__init__(model, CPSC2021, model_config, train_config, device, lazy)
def _setup_dataloaders(
self,
train_dataset: Optional[Dataset] = None,
val_dataset: Optional[Dataset] = None,
) -> NoReturn:
"""
setup the dataloaders for training and validation
Parameters
----------
train_dataset: Dataset, optional,
the training dataset
val_dataset: Dataset, optional,
the validation dataset
"""
if train_dataset is None:
train_dataset = self.dataset_cls(
config=self.train_config,
task=self.train_config.task,
training=True,
lazy=False,
)
if self.train_config.debug:
val_train_dataset = train_dataset
else:
val_train_dataset = None
if val_dataset is None:
val_dataset = self.dataset_cls(
config=self.train_config,
task=self.train_config.task,
training=False,
lazy=False,
)
# https://discuss.pytorch.org/t/guidelines-for-assigning-num-workers-to-dataloader/813/4
num_workers = 4
self.train_loader = DataLoader(
dataset=train_dataset,
batch_size=self.batch_size,
shuffle=True,
num_workers=num_workers,
pin_memory=True,
drop_last=False,
collate_fn=collate_fn,
)
if self.train_config.debug:
self.val_train_loader = DataLoader(
dataset=val_train_dataset,
batch_size=self.batch_size,
shuffle=True,
num_workers=num_workers,
pin_memory=True,
drop_last=False,
collate_fn=collate_fn,
)
else:
self.val_train_loader = None
self.val_loader = DataLoader(
dataset=val_dataset,
batch_size=self.batch_size,
shuffle=True,
num_workers=num_workers,
pin_memory=True,
drop_last=False,
collate_fn=collate_fn,
)
def run_one_step(
self, *data: Tuple[torch.Tensor, torch.Tensor]
) -> Tuple[torch.Tensor, torch.Tensor]:
"""
Parameters
----------
data: tuple of Tensors,
the data to be processed for training one step (batch),
should be of the following order:
signals, labels, *extra_tensors
Returns
-------
preds: Tensor,
the predictions of the model for the given data
labels: Tensor,
the labels of the given data
"""
if self.train_config.task == "rr_lstm":
signals, labels, weight_masks = data
# (batch_size, seq_len, n_channel) -> (seq_len, batch_size, n_channel)
signals = signals.permute(1, 0, 2)
weight_masks = weight_masks.to(device=self.device, dtype=self.dtype)
elif self.train_config.task == "qrs_detection":
signals, labels = data
else: # main task
signals, labels, weight_masks = data
weight_masks = weight_masks.to(device=self.device, dtype=self.dtype)
signals = signals.to(device=self.device, dtype=self.dtype)
labels = labels.to(device=self.device, dtype=self.dtype)
# print(f"signals: {signals.shape}")
# print(f"labels: {labels.shape}")
preds = self.model(signals)
if self.train_config.task == "qrs_detection":
return preds, labels
else:
return preds, labels, weight_masks
@torch.no_grad()
def evaluate(self, data_loader: DataLoader) -> Dict[str, float]:
""" """
self.model.eval()
if self.train_config.task == "qrs_detection":
all_rpeak_preds = []
all_rpeak_labels = []
for signals, labels in data_loader:
signals = signals.to(device=self.device, dtype=self.dtype)
labels = labels.numpy()
labels = [
mask_to_intervals(item, 1) for item in labels
] # intervals of qrs complexes
labels = [ # to indices of rpeaks in the original signal sequence
(
self.train_config.qrs_detection.reduction
* np.array([itv[0] + itv[1] for itv in item])
/ 2
).astype(int)
for item in labels
]
labels = [
item[
np.where(
(item >= self.train_config.rpeaks_dist2border)
& (
item
< self.train_config.qrs_detection.input_len
- self.train_config.rpeaks_dist2border
)
)[0]
]
for item in labels
]
all_rpeak_labels += labels
if torch.cuda.is_available():
torch.cuda.synchronize()
model_output = self._model.inference(signals)
all_rpeak_preds += model_output.rpeak_indices
eval_res = compute_rpeak_metric(
rpeaks_truths=all_rpeak_labels,
rpeaks_preds=all_rpeak_preds,
fs=self.train_config.fs,
thr=self.train_config.qrs_mask_bias / self.train_config.fs,
)
# in case possible memeory leakage?
del all_rpeak_preds, all_rpeak_labels
elif self.train_config.task == "rr_lstm":
all_preds = np.array([]).reshape(
(0, self.train_config[self.train_config.task].input_len)
)
all_labels = np.array([]).reshape(
(0, self.train_config[self.train_config.task].input_len)
)
all_weight_masks = np.array([]).reshape(
(0, self.train_config[self.train_config.task].input_len)
)
for signals, labels, weight_masks in data_loader:
signals = signals.to(device=self.device, dtype=self.dtype)
labels = labels.numpy().squeeze(
-1
) # (batch_size, seq_len, 1) -> (batch_size, seq_len)
weight_masks = weight_masks.numpy().squeeze(
-1
) # (batch_size, seq_len, 1) -> (batch_size, seq_len)
all_labels = np.concatenate((all_labels, labels))
all_weight_masks = np.concatenate((all_weight_masks, weight_masks))
if torch.cuda.is_available():
torch.cuda.synchronize()
model_output = self._model.inference(signals)
all_preds = np.concatenate(
(all_preds, model_output.af_mask)
) # or model_output.prob ?
eval_res = compute_rr_metric(all_labels, all_preds, all_weight_masks)
# in case possible memeory leakage?
del all_preds, all_labels, all_weight_masks
elif self.train_config.task == "main":
all_preds = np.array([]).reshape(
(
0,
self.train_config.main.input_len
// self.train_config.main.reduction,
)
)
all_labels = np.array([]).reshape(
(
0,
self.train_config.main.input_len
// self.train_config.main.reduction,
)
)
all_weight_masks = np.array([]).reshape(
(
0,
self.train_config.main.input_len
// self.train_config.main.reduction,
)
)
for signals, labels, weight_masks in data_loader:
signals = signals.to(device=self.device, dtype=self.dtype)
labels = labels.numpy().squeeze(
-1
) # (batch_size, seq_len, 1) -> (batch_size, seq_len)
weight_masks = weight_masks.numpy().squeeze(
-1
) # (batch_size, seq_len, 1) -> (batch_size, seq_len)
all_labels = np.concatenate((all_labels, labels))
all_weight_masks = np.concatenate((all_weight_masks, weight_masks))
if torch.cuda.is_available():
torch.cuda.synchronize()
model_output = self._model.inference(signals)
all_preds = np.concatenate(
(all_preds, model_output.af_mask)
) # or model_output.prob ?
eval_res = compute_main_task_metric(
mask_truths=all_labels,
mask_preds=all_preds,
fs=self.train_config.fs,
reduction=self.train_config.main.reduction,
weight_masks=all_weight_masks,
)
# in case possible memeory leakage?
del all_preds, all_labels, all_weight_masks
self.model.train()
return eval_res
@property
def batch_dim(self) -> int:
"""
batch dimension, usually 0,
but can be 1 for some models, e.g. RR_LSTM
"""
# return 1 if self.train_config.task in ["rr_lstm"] else 0
return 0
@property
def extra_required_train_config_fields(self) -> List[str]:
""" """
return [
"task",
]
@property
def save_prefix(self) -> str:
if self.train_config.task in ["rr_lstm"]:
return f"task-{self.train_config.task}_{self._model.__name__}_epoch"
else:
return f"task-{self.train_config.task}_{self._model.__name__}_{self.model_config.cnn_name}_epoch"
def extra_log_suffix(self) -> str:
if self.train_config.task in ["rr_lstm"]:
return f"task-{self.train_config.task}_{super().extra_log_suffix()}"
else:
return f"task-{self.train_config.task}_{super().extra_log_suffix()}_{self.model_config.cnn_name}"
def get_args(**kwargs: Any):
"""NOT checked,"""
cfg = deepcopy(kwargs)
parser = argparse.ArgumentParser(
description="Train the Model on CPSC2021",
formatter_class=argparse.ArgumentDefaultsHelpFormatter,
)
parser.add_argument(
"-b",
"--batch-size",
type=int,
default=128,
help="the batch size for training",
dest="batch_size",
)
# parser.add_argument(
# "-c", "--cnn-name",
# type=str, default="multi_scopic_leadwise",
# help="choice of cnn feature extractor",
# dest="cnn_name")
# parser.add_argument(
# "-r", "--rnn-name",
# type=str, default="none",
# help="choice of rnn structures",
# dest="rnn_name")
# parser.add_argument(
# "-a", "--attn-name",
# type=str, default="se",
# help="choice of attention structures",
# dest="attn_name")
parser.add_argument(
"--keep-checkpoint-max",
type=int,
default=20,
help="maximum number of checkpoints to keep. If set 0, all checkpoints will be kept",
dest="keep_checkpoint_max",
)
# parser.add_argument(
# "--optimizer", type=str, default="adam",
# help="training optimizer",
# dest="train_optimizer")
parser.add_argument(
"--debug",
type=str2bool,
default=False,
help="train with more debugging information",
dest="debug",
)
args = vars(parser.parse_args())
cfg.update(args)
return CFG(cfg)
_MODEL_MAP = {
"seq_lab": ECG_SEQ_LAB_NET_CPSC2021,
"unet": ECG_UNET_CPSC2021,
"lstm_crf": RR_LSTM_CPSC2021,
"lstm": RR_LSTM_CPSC2021,
}
def _set_task(task: str, config: CFG) -> NoReturn:
""" """
assert task in config.tasks
config.task = task
for item in [
"classes",
"monitor",
"final_model_name",
"loss",
]:
config[item] = config[task][item]
if __name__ == "__main__":
# WARNING: most training were done in notebook,
# NOT in cli
train_config = get_args(**TrainCfg)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
# TODO: adjust for CPSC2021
for task in train_config.tasks:
model_cls = _MODEL_MAP[train_config[task].model_name]
model_cls.__DEBUG__ = False
_set_task(task, train_config)
model_config = deepcopy(ModelCfg[task])
model = model_cls(config=model_config)
if torch.cuda.device_count() > 1 and task not in [
"rr_lstm",
]:
model = DP(model)
# model = DDP(model)
model.to(device=device)
trainer = CPSC2021Trainer(
model=model,
model_config=model_config,
train_config=train_config,
device=device,
lazy=False,
)
try:
best_model_state_dict = trainer.train()
except KeyboardInterrupt:
try:
sys.exit(0)
except SystemExit:
os._exit(0)