-
Notifications
You must be signed in to change notification settings - Fork 0
/
xla_train.py
634 lines (550 loc) · 28.8 KB
/
xla_train.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
from pathlib import Path
from future import removesuffix
import time
import math
import pandas as pd
import numpy as np
import torchmetrics as tm
print("[ √ ] torchmetrics:", tm.__version__)
from metrics import get_tm_dist_args, is_listmetric, NegativeRate, AverageMeter, EmbeddingAveragePrecision
from torch_data import get_dataloaders, get_fakedata_loaders
import torch
from torch import nn
import torchvision.transforms.functional as TF
import torch.optim as optim
from torch.optim import lr_scheduler
from utils.schedulers import get_one_cycle_scheduler, maybe_step
from utils.general import listify
from models import is_bn
from torch import FloatTensor, LongTensor
# old_metrics
#from sklearn.metrics import f1_score, accuracy_score, average_precision_score
#try:
# from sklearn.metrics import top_k_accuracy_score # requires sklearn 0.24.2 or higher (kaggle: 0.23.2)
#except ImportError:
# pass
#from sklearn.metrics import label_ranking_average_precision_score
torch.set_default_tensor_type('torch.FloatTensor')
def train_fn(model, cfg, xm, epoch, dataloader, criterion, seg_crit, optimizer, scheduler, device):
# initialize
batch_start = time.perf_counter()
model.train()
optimizer.zero_grad()
loss_meter = AverageMeter(xm)
if cfg.use_aux_loss: seg_loss_meter = AverageMeter(xm)
# prepare batch_tfms (on device)
if cfg.use_batch_tfms:
try:
from torchvision.transforms.functional import InterpolationMode
except ImportError:
from configs.torchvision import InterpolationMode
#pre_size = TT.Resize(cfg.size)
#batch_tfms = torch.nn.Sequential(
#TT.RandomRotation(degrees=5),
#TT.GaussianBlur(kernel_size=5),
#TT.RandomHorizontalFlip(),
#TT.RandomGrayscale(0.2),
# TT.RandomResizedCrop(cfg.size, scale=(0.4, 1.0)),
# )
#batch_tfms.to(device)
#if cfg.use_batch_tfms:
# raise NotImplementedError('get aug_flags first!')
# xm.master_print("aug_flags:")
# for k, v in aug_flags.items():
# xm.master_print(f'{k:<30} {v}')
# horizontal_flip = aug_flags['horizontal_flip']
# vertical_flip = aug_flags['vertical_flip']
# p_grayscale = aug_flags['p_grayscale']
# jitter_brightness = aug_flags['jitter_brightness']
# jitter_saturation = aug_flags['jitter_saturation']
# max_rotate = aug_flags['max_rotate']
# p_rotation = 0.8 if max_rotate > 0 else 0
# diag = math.sqrt(cfg.size[0] ** 2 + cfg.size[1] ** 2)
# padding = [1 + int((diag - s) / 2) for s in cfg.size[::-1]] # [x,y]
# left, top = padding
# height, width = cfg.size
# training loop
n_iter = len(dataloader)
iterable = range(n_iter) if cfg.use_batch_tfms or (cfg.fake_data == 'on_device') else dataloader
sample_iterator = iter(dataloader) if cfg.use_batch_tfms else None
for batch_idx, batch in enumerate(iterable, start=1):
# extract inputs and labels
if cfg.fake_data == 'on_device':
inputs, labels = (
torch.zeros(cfg.bs, 3, *cfg.size, device=device),
torch.zeros(cfg.bs, dtype=torch.int64, device=device))
elif cfg.use_batch_tfms:
# resize and collate images, labels
#samples = [next(sample_iterator) for _ in range(cfg.bs)]
#for s in samples:
# assert s[0].device == device
# assert s[1].device == device
#inputs = torch.stack([pre_size(s[0]) for s in samples])
#labels = torch.stack([s[1] for s in samples])
#del samples
inputs, labels = [], []
for _ in range(cfg.bs):
s = next(sample_iterator)
inputs.append(TF.resize(s[0], cfg.size, InterpolationMode('nearest')))
labels.append(s[1])
inputs = torch.stack(inputs)
labels = torch.stack(labels)
#assert inputs.shape == (cfg.bs, 3, *cfg.size)
#assert labels.shape == (cfg.bs,)
#xm.master_print(inputs.device, labels.device)
elif cfg.filetype == 'tfds':
inputs, labels = FloatTensor(batch[0]['inp1']), LongTensor(batch[0]['inp2'])
inputs = inputs.permute((0, 3, 1, 2)) #.contiguous() # mem? speed?
elif cfg.use_aux_loss:
inputs, masks, labels = batch
masks = masks.to(device)
else:
inputs, labels = batch
del batch
#xm.master_print("inputs:", type(inputs), inputs.shape, inputs.dtype, inputs.device)
#assert inputs.shape == (cfg.bs, 3, *cfg.size), f'wrong inputs shape: {inputs.shape}'
#assert labels.shape == (cfg.bs,), f'wrong labels shape: {labels.shape}'
#print(f"rank {xm.get_ordinal()} labels: {labels}")
# send to device(s) if still on CPU (device_loaders do this automatically)
if True:
inputs = inputs.to(device)
labels = labels.to(device)
else:
inputs, labels = (
torch.zeros(cfg.bs, 3, *cfg.size, device=device),
torch.zeros(cfg.bs, dtype=torch.int64, device=device))
# image batch_tfms
#if cfg.use_batch_tfms:
# inputs = batch_tfms(inputs)
#if cfg.size[1] > cfg.size[0]:
# # train on 90-deg rotated nybg2021 images
# inputs = inputs.transpose(-2,-1)
#if horizontal_flip and random.random() > 0.5:
# inputs = TF.hflip(inputs)
#if vertical_flip and random.random() > 0.5:
# inputs = TF.vflip(inputs)
#if p_grayscale and random.random() > p_grayscale:
# inputs = TF.rgb_to_grayscale(inputs, num_output_channels=3)
# # try w/o num_output_channels
#if jitter_brightness:
# if isinstance(jitter_brightness, int):
# mu, sigma = 1.0, jitter_brightness
# else:
# mu = np.mean(jitter_brightness)
# sigma = (max(jitter_brightness) - min(jitter_brightness)) * 0.28868
# brightness_factor = random.normalvariate(mu, sigma)
# inputs = TF.adjust_brightness(inputs, brightness_factor)
#if jitter_saturation:
# if isinstance(jitter_saturation, int):
# mu, sigma = 1.0, jitter_saturation
# saturation_factor = random.normalvariate(mu, sigma)
# else:
# saturation_factor = random.uniform(*jitter_saturation)
# inputs = TF.adjust_saturation(inputs, saturation_factor)
#if p_rotation and random.random() > p_rotation:
# angle = random.randint(-max_rotate, max_rotate)
# inputs = TF.pad(inputs, padding, padding_mode='reflect')
# inputs = TF.rotate(inputs, angle, resample=0)
# inputs = TF.crop(inputs, top, left, height, width)
# forward and backward pass
preds = model(inputs, labels) if model.requires_labels else model(inputs)
if cfg.use_aux_loss:
seg_logits, preds = preds
seg_loss = seg_crit(seg_logits, masks)
cls_loss = criterion(preds, labels)
loss = cfg.seg_weight * seg_loss + (1 - cfg.seg_weight) * cls_loss
else:
loss = criterion(preds, labels)
loss = loss / cfg.n_acc # grads accumulate as 'sum' but loss reduction is 'mean'
loss.backward()
if batch_idx % cfg.n_acc == 0:
xm.optimizer_step(optimizer, barrier=True) # rendevouz, required for proper xmp shutdown
optimizer.zero_grad()
if hasattr(scheduler, 'step') and hasattr(scheduler, 'batchwise'):
maybe_step(scheduler, xm)
# aggregate loss locally
if cfg.use_aux_loss:
# loss components cls_loss, seg_loss were not divided by n_acc.
loss_meter.update(cls_loss.item(), inputs.size(0))
seg_loss_meter.update(seg_loss.item(), inputs.size(0))
else:
# undo "loss /= n_acc" because loss_meter reduction is 'mean'
loss_meter.update(loss.item() * cfg.n_acc, inputs.size(0)) # 1 aten/iter, but no performance drop
#loss_meter.update(loss.detach() * cfg.n_acc, inputs.size(0)) # recursion!
#xm.add_step_closure(loss_meter.update, args=(loss.item(), cfg.n_acc * inputs.size(0))) # recursion!
# print batch_verbose information
if cfg.batch_verbose and (batch_idx % cfg.batch_verbose == 0):
info_strings = [
f' batch {batch_idx} / {n_iter}',
f'current_loss {loss_meter.current:.5f}']
if cfg.use_aux_loss:
info_strings.append(f'seg_loss {seg_loss_meter.current:.5f}')
info_strings.append(f'avg_loss {loss_meter.average:.5f}')
info_strings.append(f'lr {optimizer.param_groups[-1]["lr"] / cfg.n_replicas:7.1e}')
info_strings.append(f'mom {optimizer.param_groups[-1]["betas"][0]:.3f}')
info_strings.append(f'time {(time.perf_counter() - batch_start) / 60:.2f} min')
xm.master_print(', '.join(info_strings))
if hasattr(scheduler, 'get_last_lr'):
current_lr = optimizer.param_groups[-1]['lr']
assert scheduler.get_last_lr()[-1] == current_lr, f'scheduler: {scheduler.get_last_lr()[-1]}, opt: {current_lr}'
batch_start = time.perf_counter()
if cfg.DEBUG and batch_idx == 1:
xm.master_print(f"train inputs: {inputs.shape}, value range: {inputs.min():.2f} ... {inputs.max():.2f}")
# scheduler step after epoch
if hasattr(scheduler, 'step') and not hasattr(scheduler, 'batchwise'):
maybe_step(scheduler, xm)
return loss_meter.average
def valid_fn(model, cfg, xm, epoch, dataloader, criterion, device, metrics=None):
# initialize
model.eval()
if not cfg.pudae_valid:
loss_meter = AverageMeter(xm)
if metrics:
metrics.to(device)
# validation loop
n_iter = len(dataloader)
iterable = range(n_iter) if cfg.use_batch_tfms or (cfg.fake_data == 'on_device') else dataloader
for batch_idx, batch in enumerate(iterable, start=1):
# extract inputs and labels
if cfg.fake_data == 'on_device':
inputs, labels = (
torch.zeros(cfg.bs, 3, *cfg.size, device=device),
torch.zeros(cfg.bs, dtype=torch.int64, device=device))
elif cfg.filetype == 'tfds':
inputs, labels = FloatTensor(batch[0]['inp1']), LongTensor(batch[0]['inp2'])
inputs = inputs.permute((0, 3, 1, 2)) #.contiguous() # mem? speed?
else:
inputs, labels = batch
# send to device(s) if still on CPU (device_loaders do this automatically)
if True:
inputs = inputs.to(device)
labels = labels.to(device)
else:
inputs, labels = (
torch.zeros(cfg.bs, 3, *cfg.size, device=device),
torch.zeros(cfg.bs, dtype=torch.int64, device=device))
# forward
with torch.no_grad():
if cfg.use_aux_loss:
seg_preds, preds = model(inputs)
else:
preds = model(inputs, labels) if model.requires_labels else model(inputs)
if cfg.DEBUG and batch_idx == 1:
xm.master_print(f"valid inputs: {inputs.shape}, value range {inputs.min():.2f} ... {inputs.max():.2f}")
# pudae's ArcFace validation
if cfg.pudae_valid:
assert preds.size()[1] == 512, f'preds have wrong shape {preds.detach().size()}'
all_scores.append(preds.detach().to(torch.float16)) # default: float32
all_preds.append(torch.zeros_like(labels, dtype=torch.int8))
all_labels.append(labels.to(torch.int16)) # default: int64
continue # skip loss
# compute local loss
assert preds.detach().dim() == 2, f'preds have wrong dim {preds.detach().dim()}'
assert preds.detach().size()[1] == cfg.n_classes, f'preds have wrong shape {preds.detach().size()}'
assert labels.max() < cfg.n_classes, f'largest label out of bound: {labels.max()}'
loss = criterion(preds, labels)
loss_meter.update(loss.item(), inputs.size(0)) # 1 aten/iter but no performance drop
#loss_meter.update(loss.detach(), inputs.size(0)) # recursion!
#xm.add_step_closure(loss_meter.update, args=(loss.item(), inputs.size(0))) # recursion!
# torchmetrics
if metrics:
metrics.update(preds.detach(), labels)
# mesh_reduce loss
if cfg.pudae_valid:
avg_loss = 0
else:
avg_loss = loss_meter.average
# mesh_reduce metrics
if metrics:
avg_metrics = metrics.compute()
avg_metrics = {k: v.item() if v.ndim == 0 else v.tolist() for k, v in avg_metrics.items()}
if cfg.DEBUG and 'acc' in metrics:
counters = 'tp fp tn fn'.split()
vals = [getattr(metrics['acc'], a).item() for a in counters]
xm.master_print(f'metrics.acc {counters}: {vals}, sum: {sum(vals)}')
metrics.reset()
else:
avg_metrics = {}
return avg_loss, avg_metrics
def get_valid_labels(cfg, metadata):
class_column = metadata.columns[1] # convention, defined in metadata.get_metadata
is_valid = metadata.is_valid
is_shared = (metadata.fold == cfg.shared_fold) if cfg.shared_fold is not None else False
return metadata.loc[is_valid | is_shared, class_column].values
def _mp_fn(rank, cfg, metadata, wrapped_model, serial_executor, xm, use_fold):
"Distributed training loop master function"
# XLA device setup
device = xm.xla_device()
if cfg.xla:
import torch_xla.distributed.parallel_loader as pl
loader_prefetch_size = 1
device_prefetch_size = 1
cfg.deviceloader = cfg.deviceloader or 'mp' # 'mp' performs better than 'pl' on kaggle
# Dataloaders
if cfg.fake_data == 'on_device':
train_loader, valid_loader = None, None
elif cfg.fake_data:
train_loader, valid_loader = get_fakedata_loaders(cfg, device)
else:
train_loader, valid_loader = get_dataloaders(cfg, use_fold, metadata, xm)
#batch = next(iter(valid_loader)) # OK
#xm.master_print("test batch:", len(batch), batch[0].shape, batch[1].shape)
# Send model to device
model = wrapped_model.to(device)
# Criterion (default reduction: 'mean'), Metrics
criterion = nn.BCEWithLogitsLoss() if cfg.multilabel else nn.CrossEntropyLoss()
if cfg.use_aux_loss:
from segmentation_models_pytorch.losses.dice import DiceLoss
seg_crit = DiceLoss('binary') if cfg.use_aux_loss else None
#old_metrics = [] # OK: [pct_N, eap5, acc, macro_acc, top3, f1, macro_f1]
# cfg.metrics and metrics need to have identical keys: replace aliases in cfg.metrics
aliases = {
'micro_acc': 'acc',
'f1': 'F1',
'macro_f1': 'macro_F1',
'class_f1': 'class_F1',
'f2': 'F2',
'neg_rate': 'pct_N',
'map': 'mAP',
'eap5': 'eAP5',
}
cfg.metrics = cfg.metrics or []
cfg.metrics = [k.replace(k, aliases[k]) if k in aliases else k for k in cfg.metrics]
if cfg.save_best and cfg.save_best in aliases:
cfg.save_best = aliases[cfg.save_best]
# torchmetrics
dist_args = get_tm_dist_args(xm)
metrics = {}
if 'acc' in cfg.metrics:
metrics['acc'] = tm.Accuracy(**dist_args)
if 'macro_acc' in cfg.metrics:
metrics['macro_acc'] = tm.Accuracy(average='macro', num_classes=cfg.n_classes, **dist_args)
if 'top5' in cfg.metrics:
metrics['top5'] = tm.Accuracy(top_k=5, **dist_args)
if 'top3' in cfg.metrics:
metrics['top3'] = tm.Accuracy(top_k=3, **dist_args)
if 'F1' in cfg.metrics:
metrics['F1'] = tm.F1Score(num_classes=cfg.n_classes, average='micro', **dist_args)
if 'macro_F1' in cfg.metrics:
metrics['macro_F1'] = tm.F1Score(num_classes=cfg.n_classes, average='macro', **dist_args)
if 'class_F1' in cfg.metrics:
metrics['class_F1'] = tm.F1Score(num_classes=cfg.n_classes, average=None, **dist_args)
if 'F2' in cfg.metrics:
metrics['F2'] = tm.FBetaScore(num_classes=cfg.n_classes, average='micro', beta=2.0, **dist_args)
if 'pct_N' in cfg.metrics:
negative_class = (
0 if cfg.negative_class is None else
cfg.negative_class if isinstance(cfg.negative_class, int) else
cfg.vocab.transform([cfg.negative_class])[0])
metrics['pct_N'] = NegativeRate(num_classes=cfg.n_classes, negative_class=negative_class,
threshold=cfg.negative_thres or 0.5)
if 'mAP' in cfg.metrics:
metrics['mAP'] = tm.AveragePrecision(average='macro', num_classes=cfg.n_classes, **dist_args)
if 'eAP5' in cfg.metrics:
metrics['eAP5'] = EmbeddingAveragePrecision(xm, k=5) # happywhale
metrics = tm.MetricCollection(metrics, compute_groups=cfg.metric_compute_groups)
# metric_compute_groups: don't use if metrics use modifying kwargs like 'num_classes', 'threshold', 'top_k'
# Issue: If different metrics yield same values, set compute_groups=False!
# MetricCollection.__setitem__ is broken (only first update works?)
#xm.master_print('Metrics:', *[m.__name__ for m in old_metrics])
xm.master_print('Metrics:', metrics)
# Don't Scale LRs (optimal lrs don't scale linearly with step size)
lr_head, lr_bn, lr_body = cfg.lr_head, cfg.lr_bn, cfg.lr_body
# Parameter Groups
use_parameter_groups = False if lr_head == lr_bn == lr_body else True
if use_parameter_groups:
xm.master_print(f"Using parameter groups. lr_head={lr_head}, lr_body={lr_body}, lr_bn={lr_bn}")
parameter_groups = {
'body': (p for name, p in model.body.named_parameters() if not is_bn(name)),
'head': model.head.parameters(),
'bn': (p for name, p in model.body.named_parameters() if is_bn(name)),
}
max_lrs = {'body': lr_body, 'head': lr_head, 'bn': lr_bn}
params = [{'params': parameter_groups[g], 'lr': max_lrs[g]}
for g in parameter_groups.keys()]
max_lrs = list(max_lrs.values())
else:
max_lrs = lr_head
params = model.parameters()
# Optimizer
optimizer = (
optim.AdamW(params, lr=lr_head, betas=cfg.betas, weight_decay=cfg.wd) if cfg.optimizer == 'AdamW' else
optim.Adam(params, lr=lr_head, betas=cfg.betas) if cfg.optimizer == 'Adam' else
optim.SGD(params, lr=lr_head, momentum=cfg.betas[0], dampening=1 - cfg.betas[1]))
rst_epoch = 0
if cfg.rst_name:
fn = Path(cfg.rst_path) / f'{removesuffix(cfg.rst_name, ".pth")}.opt'
if fn.exists() and not cfg.reset_opt:
checkpoint = torch.load(fn, map_location='cpu')
optimizer.load_state_dict(checkpoint['optimizer_state_dict'])
rst_epoch = checkpoint['epoch'] + 1
xm.master_print("Restarting from previous opt state")
# Scheduler
if cfg.one_cycle:
scheduler = get_one_cycle_scheduler(optimizer, max_lrs, cfg,
xm, rst_epoch, train_loader)
elif cfg.reduce_on_plateau:
# ReduceLROnPlateau must be called after validation
scheduler = lr_scheduler.ReduceLROnPlateau(optimizer, 'min', factor=0.5,
patience=5, verbose=True, eps=1e-6)
elif isinstance(cfg.step_lr_after, int) and cfg.step_lr_factor > 0:
scheduler = lr_scheduler.StepLR(optimizer, step_size=cfg.step_lr_after,
gamma=cfg.step_lr_factor)
elif hasattr(cfg.step_lr_after, '__iter__') and cfg.step_lr_factor > 0:
scheduler = lr_scheduler.MultiStepLR(optimizer, milestones=cfg.step_lr_after,
gamma=cfg.step_lr_factor)
else:
scheduler = None
if scheduler:
xm.master_print(f"Scheduler: {scheduler.__class__.__name__}")
#_lrs = [p["lr"] / cfg.n_replicas for p in optimizer.param_groups]
_lrs = [p["lr"] for p in optimizer.param_groups]
xm.master_print(f"""Initial lrs: {', '.join(f'{lr:7.2e}' for lr in _lrs)}""")
#_lrs = [lr / cfg.n_replicas for lr in listify(max_lrs)]
_lrs = [lr for lr in listify(max_lrs)]
xm.master_print(f"Max lrs: {', '.join(f'{lr:7.2e}' for lr in _lrs)}")
# Maybe freeze body
if hasattr(model, 'body') and lr_body == 0:
xm.master_print("Freezing body of pretrained model")
for n, p in model.body.named_parameters():
if not is_bn(n):
p.requires_grad = False
model_name = f'{cfg.name}_fold{use_fold}'
xm.master_print(f'Checkpoints will be saved as {cfg.out_dir}/{model_name}_ep*')
step_size = cfg.bs * cfg.n_replicas * cfg.n_acc
xm.master_print(f'Training {cfg.arch_name}, size={cfg.size}, replica_bs={cfg.bs}, '
f'step_size={step_size}, lr={cfg.lr_head} on fold {use_fold}')
#
#
### Training Loop ---------------------------------------------------------
lrs = []
metrics_dicts = []
minutes = []
best_model_score = -np.inf
epoch_summary_header = ''.join([
#'epoch ', ' train_loss ', ' valid_loss ', ' '.join([f'{m.__name__:^8}' for m in metrics]),
'epoch ', ' train_loss ', ' valid_loss ',
' '.join([f'{key:^8}' for key in cfg.metrics if not is_listmetric(metrics[key])]),
' lr ', 'min_train min_total'])
xm.master_print("\n", epoch_summary_header)
xm.master_print("=" * (len(epoch_summary_header) + 2))
for epoch in range(rst_epoch, rst_epoch + cfg.epochs):
# Data for verbose info
epoch_start = time.perf_counter()
current_lr = optimizer.param_groups[-1]["lr"]
if hasattr(scheduler, 'get_last_lr'):
assert scheduler.get_last_lr()[-1] == current_lr, f'scheduler: {scheduler.get_last_lr()[-1]}, opt: {current_lr}'
# Update train_loader shuffling
if hasattr(train_loader, 'sampler') and hasattr(train_loader.sampler, 'set_epoch'):
train_loader.sampler.set_epoch(epoch)
# Training
if cfg.xla and (cfg.deviceloader == 'pl') and (cfg.fake_data != 'on_device'):
# ParallelLoader requires instantiation per epoch
dataloader = pl.ParallelLoader(train_loader, [device],
loader_prefetch_size=loader_prefetch_size,
device_prefetch_size=device_prefetch_size
).per_device_loader(device)
else:
dataloader = train_loader
train_loss = train_fn(model, cfg, xm,
epoch = epoch + 1,
dataloader = dataloader,
criterion = criterion,
seg_crit = seg_crit,
optimizer = optimizer,
scheduler = scheduler,
device = device)
# Validation
valid_start = time.perf_counter()
if cfg.train_on_all:
valid_loss, valid_metrics = 0, []
else:
if cfg.xla and (cfg.deviceloader == 'pl') and (cfg.fake_data != 'on_device'):
# ParallelLoader requires instantiation per epoch
dataloader = pl.ParallelLoader(valid_loader, [device],
loader_prefetch_size=loader_prefetch_size,
device_prefetch_size=device_prefetch_size
).per_device_loader(device)
else:
dataloader = valid_loader
#valid_loss, old_valid_metrics, valid_metrics = valid_fn(model, cfg, xm,
valid_loss, valid_metrics = valid_fn(model, cfg, xm,
epoch = epoch + 1,
dataloader = dataloader,
criterion = criterion,
device = device,
#old_metrics = old_metrics,
metrics = metrics)
#old_metrics_dict = {'train_loss': train_loss, 'valid_loss': valid_loss}
metrics_dict = {'train_loss': train_loss, 'valid_loss': valid_loss}
#old_metrics_dict.update({m.__name__: val for m, val in zip(old_metrics, old_valid_metrics)})
metrics_dict.update(valid_metrics)
last_lr = optimizer.param_groups[-1]["lr"] if hasattr(scheduler, 'batchwise') else current_lr
avg_lr = 0.5 * (current_lr + last_lr)
# Old epoch summary
#epoch_summary_strings = [f'{epoch + 1:>2} / {rst_epoch + cfg.epochs:<2}'] # ep/epochs
#epoch_summary_strings.append(f'{train_loss:10.5f}') # train_loss
#epoch_summary_strings.append(f'{valid_loss:10.5f}') # valid_loss
#for val in old_valid_metrics: # metrics
# if isinstance(val, list): xm.master_print(val)
# epoch_summary_strings.append(f'{val:7.5f}')
#xm.master_print(' '.join(epoch_summary_strings))
# Print epoch summary
epoch_summary_strings = [f'{epoch + 1:>2} / {rst_epoch + cfg.epochs:<2}'] # ep/epochs
epoch_summary_strings.append(f'{train_loss:10.5f}') # train_loss
epoch_summary_strings.append(f'{valid_loss:10.5f}') # valid_loss
for key in cfg.metrics: # metrics
# cannot use valid_metric.items() because MetricCollection re-orders keys alphabetically
val = valid_metrics[key]
if isinstance(val, list):
if getattr(metrics[key], 'average', '') is not None:
xm.master_print(f'Warning: metric {key} returned list but "average" attribute is not None')
xm.master_print(key + ':\t' + "\t".join(f'{v:.5f}' for v in val))
else:
epoch_summary_strings.append(f'{val:7.5f}')
epoch_summary_strings.append(f'{avg_lr:7.1e}') # lr
epoch_summary_strings.append(f'{(valid_start - epoch_start) / 60:7.2f}') # Wall train
epoch_summary_strings.append(f'{(time.perf_counter() - epoch_start) / 60:7.2f}') # Wall total
xm.master_print(' '.join(epoch_summary_strings))
# Save weights, optimizer state, scheduler state
# Note: xm.save must not be inside an if statement that may validate differently on
# different TPU cores. Reason: rendezvous inside them will hang if any core
# does not arrive at the rendezvous.
if cfg.save_best:
assert cfg.save_best in metrics_dict, f'{cfg.save_best} not in {list(metrics_dict)}'
model_score = metrics_dict[cfg.save_best] if cfg.save_best else -valid_loss
if cfg.save_best and 'loss' in cfg.save_best: model_score = -model_score
if model_score > best_model_score or not cfg.save_best:
if cfg.save_best:
best_model_score = model_score
#xm.master_print(f'{cfg.save_best or "valid_loss"} improved.')
fn = cfg.out_dir / f'{model_name}_best_{cfg.save_best}'
else:
fn = cfg.out_dir / f'{model_name}_ep{epoch+1}'
#xm.master_print(f'saving {model_name}_ep{epoch+1}.pth ...')
xm.save(model.state_dict(), f'{fn}.pth')
#xm.master_print(f'saving {model_name}_ep{epoch+1}.opt ...')
xm.save({'optimizer_state_dict': optimizer.state_dict(),
'epoch': epoch}, f'{fn}.opt')
if hasattr(scheduler, 'state_dict'):
#xm.master_print(f'saving {model_name}_ep{epoch+1}.sched ...')
xm.save({'scheduler_state_dict': {
k: v for k, v in scheduler.state_dict().items() if k != 'anneal_func'}},
f'{fn}.sched')
# Save losses, metrics
#train_losses.append(train_loss)
#valid_losses.append(valid_loss)
metrics_dicts.append(metrics_dict)
lrs.append(avg_lr)
minutes.append((time.perf_counter() - epoch_start) / 60)
if cfg.n_replicas > 1:
serial_executor.run(lambda: save_metrics(metrics_dicts, lrs, minutes, rst_epoch, use_fold, cfg.out_dir))
else:
save_metrics(metrics_dicts, lrs, minutes, rst_epoch, use_fold, cfg.out_dir)
def save_metrics(metrics_dicts, lrs, minutes, rst_epoch, fold, out_dir):
df = pd.DataFrame(metrics_dicts)
df['lr'] = lrs
df['Wall'] = minutes
df['epoch'] = df.index + rst_epoch + 1
df.set_index('epoch', inplace=True)
df.to_json(Path(out_dir) / f'metrics_fold{fold}.json')