-
Notifications
You must be signed in to change notification settings - Fork 3
/
torch_trainer.py
716 lines (533 loc) · 21.2 KB
/
torch_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
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
# __author__ = 'Vasudev Gupta'
# __author_email__ = '7vasudevgupta@gmail.com'
import numpy as np
import torch
from torch.utils.tensorboard import SummaryWriter
import os
import wandb
from tqdm import tqdm
from abc import ABC, abstractmethod
from dataclasses import dataclass
if torch.cuda.is_available():
m1 = 'available'
else:
m1 = "not available"
try:
import torch_xla
import torch_xla.core.xla_model as xm
import torch_xla.distributed.xla_multiprocessing as xmp
m2 = "available"
except:
m2 = 'not available'
"""
USAGE:
from torch_utils import TorchTrainer, DefaultArgs
class Trainer(TorchTrainer):
def __init__(self, model, args):
self.model = model
# call this at end only
super().__init__(args)
def configure_optimizers(self):
'''
....
'''
def training_step(self, batch, batch_idx):
'''
....
'''
def validation_step(self, batch):
'''
....
'''
# define model architecture
model = .....
# define dataset
tr_dataset = .....
val_dataset = .....
@dataclass
class Config(DefaultArgs):
# pass your args
lr: float = 2e-5
......
# If want to update defaut_args; just pass it here only
save_dir: str = 'weights'
.......
trainer = Trainer(model, args)
trainer.fit(tr_dataset, val_dataset)
Using TPU is pretty simple, run following command:
!pip install cloud-tpu-client==0.10 https://storage.googleapis.com/tpu-pytorch/wheels/torch_xla-1.6-cp36-cp36m-linux_x86_64.whl
& pass tpus=1 in DefaultArgs
"""
@dataclass
class DefaultArgs:
# root dir for any kind of saving
base_dir: str = "."
# args used in TorchTrainer
map_location: torch.device = torch.device("cuda:0")
# model weights will be in `.pt` file
# while other training stuff will be in `.tar`
save_dir: str = "resuming"
load_dir: str = None
save_epoch_dir: str = None
early_stop_n: int = None
epoch_saving_n: int = 3
fast_dev_run: bool = False
project_name: str = None
wandb_run_name: str = None
wandb_off: bool = False
# will be helpful in resuming
wandb_resume: bool = False
wandb_run_id: str = None
max_epochs: int = 5
accumulation_steps: int = 1
tpus: int = 0
precision: str = 'float32'
class TrainerSetup(object):
print(
"""
|-----------------------------------|
| Device | Status |
|-----------------------------------|
GPU | {}
|-----------------------------------|
TPU | {}
|-----------------------------------|
""".format(m1, m2)
)
def __init__(self):
"""
This class is mainly having setup methods for enable training
"""
@staticmethod
def assert_epoch_saving(val_metric: list, n: int = 3, mode: str = "min"):
"""
Allows saving if loss decreases / accuracy increases
n = 'min' corresponds to val_metric being loss-metric
n = 'max' corresponds to val_metric being accuracy-metric
Note:
val_metric should be having current value of loss/accuracy
"""
status = False
if len(val_metric) < n+1:
return True
current_val = val_metric[-1]
compr = val_metric[-n-2:-2]
if mode == "min":
compr = np.min(compr)
if current_val < compr:
status = True
elif mode == "max":
compr = np.max(compr)
if current_val > compr:
status = True
else:
raise ValueError("mode can be only either max or min")
return status
@staticmethod
def assert_early_stop(val_metric: list, n: int = None, mode="min"):
"""
If n is specified, then only early stopping will be enabled
n = 'min' corresponds to val_metric being loss-metric
n = 'max' corresponds to val_metric being accuracy-metric
"""
assert(mode in ["max", "min"])
stop_status = False
if n is None:
return stop_status
compr = np.max(val_metric[-n:]) if mode == "min" else np.min(val_metric[-n:])
if compr == val_metric[-1]:
stop_status = True
return stop_status
def stop_early(self, val_metric: list, n: int = None, mode="min"):
status = self.assert_early_stop(val_metric, n, mode)
if status:
raise KeyboardInterrupt("Model training stopped due to early-stopping")
@staticmethod
def _sanity_check(args: DefaultArgs):
if not hasattr(args, "__dict__"):
raise ValueError("Your argument class must have `dataclass` decorator")
for arg in DefaultArgs().__dict__:
if not hasattr(args, arg):
raise ValueError(f"Your config must have `{arg}`")
@staticmethod
def _setup_savedir(save_dir: str, base_dir: str):
if save_dir:
if save_dir not in os.listdir(base_dir):
os.mkdir(f"{base_dir}/{save_dir}")
return save_dir
return save_dir
@staticmethod
def _setup_basedir(base_dir: str):
if base_dir is None:
return "."
elif base_dir == ".":
return base_dir
elif base_dir not in os.listdir():
os.mkdir(base_dir)
print(f"training stuff will be saved in {base_dir}")
return base_dir
@staticmethod
def setup_wandb(args: dict):
# useful for testing
if args["wandb_off"]:
try:
os.system('wandb offline')
except:
raise ValueError("wandb not available")
if args["wandb_resume"]:
if args["wandb_run_id"] is None:
raise ValueError('wandb-run-id must be mentioned for resuming training')
wandb.init(resume=args["wandb_run_id"])
else:
wandb.init(project=args["project_name"], name=args["wandb_run_name"], id=args["wandb_run_id"], config=args["wandb_config"], dir=args["wandb_dir"])
@staticmethod
def display_metrics(epochs: list, tr_metrics: list, val_metrics: list):
# round factors should be 3 for proper layout
results = """
|--------------------------------------------|
epoch | tr_metric | val_metric
|--------------------------------------------|"""
for e, tr, val in zip(range(epochs), tr_metrics, val_metrics):
res = """
{} | {} | {}
|--------------------------------------------|""".format(
np.round(e, 3), np.round(tr, 3), np.round(val, 3)
)
results += res
print(results)
@staticmethod
def model_summary(self, model: torch.nn.Module):
num = np.sum([p.nelement() for p in model.parameters()])
s = {"Net": num}
for n, layer in model.named_children():
num = np.sum([p.nelement() for p in layer.parameters()])
s.update({n: num})
print("Layers | Parameters")
for l, p in s.items():
print("{} | {}".format(l, p))
class TrainingLoop(ABC, TrainerSetup):
@abstractmethod
def configure_optimizers(self, **kwargs):
"""This method must be implemented in the class inherited from this class"""
@abstractmethod
def training_step(self, **kwargs):
"""This method must be implemented in the class inherited from this class"""
@abstractmethod
def validation_step(self, **kwargs):
"""This method must be implemented in the class inherited from this class"""
def training_batch_end(self, batch_idx):
"""This method is called at the end of batch-{batch_idx}"""
def training_epoch_end(self, epoch, losses):
"""This method is called at the end of epoch"""
def training_end(self):
"""This method is called at the end of complete training"""
def after_backward(self, batch_idx):
"""This method is called just after `loss.backward()`"""
def __init__(self, args: DefaultArgs):
super().__init__()
self._sanity_check(args)
# self.model = ?
self.base_dir = self._setup_basedir(args.base_dir)
self.precision = args.precision
self.load_dir = args.load_dir
self.save_dir = self._setup_savedir(args.save_dir, self.base_dir)
self.save_epoch_dir = self._setup_savedir(args.save_epoch_dir, self.base_dir)
self.early_stop_n = args.early_stop_n
self.epoch_saving_n = args.epoch_saving_n
self.device = self.configure_devices(args)
if self.gpus > 1:
self.model = torch.nn.DataParallel(self.model)
if self.load_dir:
self.map_location = args.map_location
self.load_model_state_dict(f"{self.base_dir}/{self.load_dir}")
self.model.to(self.device)
self.optimizer = self.configure_optimizers()
self.scaler = self._configure_scaler()
self.start_epoch = 0
self.start_batch_idx = 0
if self.load_dir:
self.load_training_state_dict(self.load_dir)
self.max_epochs = args.max_epochs
self.accumulation_steps = args.accumulation_steps
self.fast_dev_run = args.fast_dev_run
self.wandb_args = {
"wandb_config": args.__dict__,
"wandb_resume": args.wandb_resume,
"project_name": args.project_name,
"wandb_run_name": args.wandb_run_name,
"wandb_run_id": args.wandb_run_id,
"wandb_off": args.wandb_off,
"wandb_dir": self.base_dir
}
if self.precision == 'float16':
self._setup_half()
def train_step(self, batch, batch_idx):
if self.precision == 'mixed16':
return torch.cuda.amp.autocast(self.training_step)(batch, batch_idx)
return self.training_step(batch, batch_idx)
def val_step(self, batch):
if self.precision == 'mixed16':
return torch.cuda.amp.autocast(self.validation_step)(batch, batch_idx)
return self.validation_step(batch)
def _setup_half(self):
raise ValueError('Need to implement float16 with apex')
def configure_devices(self, args):
device = torch.device('cpu')
# If gpu is available, its automatically getting used
self.gpus = torch.cuda.device_count()
if self.gpus > 0:
device = torch.device('cuda')
torch.backends.cudnn.benchmark = True
if self.precision == 'mixed16':
if args.tpus > 0:
raise ValueError('Currently mixed_16 is not supported with TPUs')
elif self.gpus == 0:
raise ValueError("mixed_16 should not be used with cpu")
self.tpus = args.tpus
if args.tpus == 1:
try:
device = xm.xla_device()
self.gpus = 0
except:
raise ValueError("Can't set device to TPUs")
print(f"Using {device}")
args.device = device
return device
def fit(self,
tr_dataset: torch.utils.data.DataLoader,
val_dataset: torch.utils.data.DataLoader):
self.setup_wandb(self.wandb_args)
if self.fast_dev_run:
print("fast_dev_run is set to True")
tr_dataset = next(iter(tr_dataset))
val_dataset = next(iter(val_dataset))
try:
tr_metric, val_metric = self.train(tr_dataset, val_dataset)
self.display_metrics(self.max_epochs, tr_metric, val_metric)
except KeyboardInterrupt:
print('Interrupting through keyboard ======= Saving model weights')
torch.save(self.model.state_dict(), 'keyboard-interrupted_wts')
def _configure_scaler(self):
if self.precision == 'mixed16':
if not torch.cuda.is_available():
raise ValueError('CUDA is not available')
print('Training with mixed16')
return torch.cuda.amp.GradScaler()
@torch.no_grad()
def empty_grad(self):
for param in self.model.parameters():
param.grad = None
def train(self, tr_dataset, val_dataset):
tr_metric = []
val_metric = []
steps = 0 # updating under accumulation condition
# setting up epochs (handling resuming)
epochs = range(self.start_epoch, self.max_epochs)
for epoch in epochs:
# setting up tr_loss for accumulation
tr_loss = 0
losses = []
# helping in resuming
self.start_epoch = epoch
# setting up progress bar to display
desc = f"running epoch-{epoch}"
pbar = tqdm(enumerate(tr_dataset), total=len(tr_dataset), desc=desc, initial=0, leave=False)
for batch_idx, batch in pbar:
# will help in resuming training from last-saved batch_idx
if batch_idx != self.start_batch_idx:
steps += 1
pbar.write(f'training will start from batch_idx-{self.start_batch_idx}')
continue
self.start_batch_idx += 1
self.model.train(True)
# simply doing forward-propogation
loss = self.train_step(batch, batch_idx)
loss /= self.accumulation_steps
# accumulating tr_loss for logging (helpful when accumulation-steps > 1)
tr_loss += loss.item()
# configuring for mixed-precision
if self.precision == 'mixed16':
loss = self.scaler.scale(loss)
loss.backward()
self.after_backward(batch_idx)
# gradient accumulation handler
if (batch_idx+1)%self.accumulation_steps == 0:
# configuring for mixed-precision
if self.precision == 'mixed16':
self.mixed_optimizer_step(self, self.optimizer)
else:
if self.tpus == 1:
xm.optimizer_step(self.optimizer, barrier=True)
else:
self.optimizer.step()
wandb.log({
'global_steps': steps,
'step_tr_loss': tr_loss
}, commit=True)
steps += 1
pbar.set_postfix(tr_loss=tr_loss)
# emptying gradients in very efficient way
self.empty_grad()
# accumulating losses for training-loss at epoch end
losses.append(tr_loss)
# emptying tr_loss
tr_loss = 0
self.training_batch_end(batch_idx)
# clearing batch_idx for next epoch
self.start_batch_idx = 0
# val_loss at training epoch end for logging
val_loss = self.evaluate(val_dataset)
wandb.log({
'epoch': epoch,
'tr_loss': np.mean(losses),
'val_loss': val_loss.item()
}, commit=False)
tr_metric.append(np.mean(losses))
val_metric.append(val_loss.item())
if self.save_epoch_dir:
save_status = self.assert_epoch_saving(val_metric, n=self.epoch_saving_n, mode="min")
if save_status:
self.save_model_state_dict(f"{self.base_dir}/{self.save_epoch_dir}")
self.save_training_state_dict(f"{self.base_dir}/{self.save_epoch_dir}")
self.training_epoch_end(epoch, losses)
if self.early_stop_n:
self.stop_early(val_metric, self.early_stop_n, model="min")
self.start_epoch += 1
self.training_end()
if self.save_dir:
print("Saving model and training related stuff")
self.save_model_state_dict(f"{self.base_dir}/{self.save_dir}")
self.save_training_state_dict(f"{self.base_dir}/{self.save_dir}")
return tr_metric, val_metric
def mixed_optimizer_step(self):
self.scaler.step(self.optimizer)
self.scaler.update()
def evaluate(self, val_dataset):
# disabling layers like dropout, batch-normalization
self.model.train(False)
desc = 'Validating ....'
pbar = tqdm(val_dataset, total=len(val_dataset), desc=desc, initial=0, leave=False)
for batch in pbar:
val_loss = self.val_step(batch)
pbar.set_postfix(val_loss=val_loss.item())
return val_loss
def save_training_state_dict(self, save_dir: str):
path = f"{save_dir}/training.tar"
# defining what all to save
state_dict = {
'optimizer': self.optimizer.state_dict(),
'start_epoch': self.start_epoch,
'start_batch_idx': self.start_batch_idx
}
# mixed-precision states saving, if saving enabled
if self.precision == 'mixed16':
state_dict.update({
'scaler': self.scaler.state_dict()
})
torch.save(state_dict, path)
def save_model_state_dict(self, save_dir: str):
path = f"{save_dir}/model.pt"
module = self.model.module if hasattr(self.model, "module") else self.model
state_dict = module.state_dict()
if self.tpus > 0:
xm.save(state_dict, path)
else:
torch.save(state_dict, path)
def load_model_state_dict(self, load_dir: str):
path = f"{load_dir}/model.pt"
"""
Note:
`map_function` is very memory expensive if you are changing the device
"""
print(
"""loading:
1) model state_dict
"""
)
model = torch.load(path, map_location=self.map_location)
if hasattr(self.model, "module"):
self.model.module.load_state_dict(model)
else:
self.model.load_state_dict(model)
def load_training_state_dict(self, load_dir: str):
path = f"{load_dir}/training.tar"
print(
"""loading:
1) optimizer-state-dict
2) scaler-state-dict (if mixed-precision)
3) start_epoch
4) start_batch_idx
"""
)
checkpoint = torch.load(path)
self.optimizer.load_state_dict(checkpoint.pop('optimizer'))
if self.precision == 'mixed16':
self.scaler.load_state_dict(checkpoint.pop('scaler'))
# helpful in resuming training from particular step
self.start_epoch = checkpoint.pop('start_epoch')
self.start_batch_idx = checkpoint.pop('start_batch_idx')
print(f'loading successful (start-epoch-{self.start_epoch}, start_batch_idx-{self.start_batch_idx})')
class TorchTrainer(TrainingLoop):
def __init__(self, args):
TrainingLoop.__init__(self, args)
@abstractmethod
def configure_optimizers(self):
"""
Return:
`torch.optim` object
"""
@abstractmethod
def training_step(self, batch, batch_idx):
"""
This method should look something like this
batch = batch.to(self.device)
out = self(batch)
loss = out.mean()
return loss
"""
@abstractmethod
def validation_step(self, batch):
"""
This method should look something like this
batch = batch.to(self.device)
with torch.no_grad():
out = self(batch)
loss = out.mean()
return loss
"""
def training_batch_end(self, batch_idx):
"""This method is called at the end of batch-{batch_idx}"""
def training_epoch_end(self, epoch, losses):
"""This method is called at the end of epoch"""
def training_end(self):
"""This method is called at the end of complete training"""
def after_backward(self, batch_idx):
"""This method is called just after `loss.backward()`"""
def histogram_params(self, logdir="tb_params"):
"""
You need to call this method yourself
"""
writer = SummaryWriter(log_dir=f"{self.base_dir}/{logdir}")
params = self.model.named_parameters()
for n, param in params:
writer.add_histogram(n, param)
writer.close()
# tensorboard --logdir "{tb_params}"
def histogram_grads(self, logdir="tb_grads"):
"""
You need to call this method yourself
Remember to call this only after `backward`
"""
writer = SummaryWriter(log_dir=f"{self.base_dir}/{logdir}")
params = self.model.named_parameters()
for n, param in params:
if param.grad is not None:
writer.add_histogram(n, param.grad)
else:
writer.add_scalar(n, 0.0)
writer.close()
# tensorboard --logdir "{tb_grads}"
if __name__ == '__main__':
"""
Peace max .....
"""