-
Notifications
You must be signed in to change notification settings - Fork 0
/
mtl_train_eval.py
586 lines (507 loc) · 26.2 KB
/
mtl_train_eval.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
# coding=utf-8
'''
Created: 2021/12/3
@author: [email protected]
'''
# pytorch
import torch
import torch.optim as optim
# determined
from determined.tensorboard.metric_writers.pytorch import TorchWriter
from determined.pytorch import (
PyTorchTrial,
PyTorchTrialContext,
LRScheduler,
DataLoader
)
import numpy as np
# utils
from typing import (
Any,
Dict,
Union,
List,
Sequence
)
import os
# costume
from data.nyud import NYUD
from modules.VGG16 import DeepLabLargeFOVBN
from modules.ResNet import resnet50, resnet101
from modules.DeepLabV3Plus import DeepLabV3, DeepLabV3Plus
from modules.UNITENet import UNITENet
from utils.get_scheduler import get_lr_scheduler
from utils.MTLLosses import MTLLosses
from utils.MTLPredictionsReducer import MTLPredictionsReducer
from experiments.mtl import get_cfg_defaults
from optimizer.adabound import AdaBound, AdaBoundW
from utils.init_seed import init_seeds
from utils.metrics import (
get_mIoU,
get_conf_mat,
get_normal_cosine,
get_normal_metrics
)
from modules.test.NDDRNetAttention import NDDRNetAttention
# from utils.gpu_mem_track import MemTracker
# from pytorch_memlab import MemReporter
TorchData = Union[Dict[str, torch.Tensor], Sequence[torch.Tensor], torch.Tensor]
class UNITETrainEvalTrail(PyTorchTrial):
def __init__(self, context: PyTorchTrialContext) -> None:
self.context = context
# self.mem_tracker =MemTracker()
# cannot be used determined --test
self.det_exp_cfg = self.context.get_experiment_config()
print(self.det_exp_cfg)
cfg = get_cfg_defaults()
cfg.merge_from_file(self.context.get_hparam('config'))
cfg.freeze()
self.cfg = cfg
# print cfg
print(cfg)
self.last_epoch_idx = -1
# tensorboard logging
self.logger = TorchWriter()
self.dataset_name = cfg.DATASET_NAME
self.dataset_dir = cfg.DATASET_DIR
# mlt configs
self.tasks = cfg.MTL.TASKS
self.n_tasks = len(self.tasks)
self.n_output_per_task = cfg.MTL.N_OUTPUT_PER_TASK
self.backbone_type = cfg.MTL.BACKBONE_TYPE
self.weight_dir = cfg.MTL.WEIGHT_DIR
self.weight_per_task = cfg.MTL.WEIGHT_PER_TASK
self.loss_factors = cfg.MTL.LOSS_FACTORS
# unite cfg
self.n_patches = cfg.MODEL.N_PATCHES
self.similarity = cfg.MODEL.SIMILARITY
self.reconstruct_src = cfg.MODEL.RECONSTRUCT_SRC
self.init_method = cfg.MODEL.INIT_METHOD
self.init_weights = cfg.MODEL.INIT_WEIGHTS
self.shortcut = cfg.MODEL.SHORTCUT
self.bn_before_activation = cfg.MODEL.BN_BEFORE_ACTIVATION
# optimization cfg
self.base_lr = cfg.BASE_LR
self.base_factor = cfg.BASE_FACTOR
self.unites_factor = cfg.UNITES_FACTOR
self.fc8_w_factor = cfg.FC8_W_FACTOR # used for vgg
self.fc8_b_factor = cfg.FC8_B_FACTOR # used for vgg
self.head_factor = cfg.HEAD_FACTOR # used for r50v1/r101v1
self.classifier_factor = cfg.CLASSIFIER_FACTOR # used for resnet
self.weight_decay = cfg.WEIGHT_DECAY
self.momentum = cfg.MOMENTUM
self.n_train_units = self.det_exp_cfg['searcher']['max_length']['epochs'] * \
np.ceil(self.det_exp_cfg['records_per_epoch'] /
self.context.get_global_batch_size()) # not available for non-local test
self.n_warmup_units = cfg.N_WARMUP_UNITS
assert self.n_warmup_units < self.n_train_units
self.scheduler = cfg.SCHEDULER
self.poly_power = cfg.POLY_POWER
self.n_workers = self.context.get_hparam('n_workers')
assert isinstance(self.n_output_per_task, list) and len(self.n_output_per_task) == self.n_tasks
# assume that the order of # of outputs is along with taht of weight_per_task
assert isinstance(self.weight_per_task, list) and len(self.weight_per_task) == self.n_tasks
assert self.dataset_name in ['nyud', 'cityscapes', 'pascal_context', 'taskonomy']
assert self.backbone_type in ['VGG16V1', 'R50V1', 'R50V3', 'R50V3P', 'R101V1', 'R101V3', 'R101V3P']
assert isinstance(self.n_patches, list)
assert isinstance(self.init_weights, list)
# reproducibility
# determined <= 0.7.5 is not able to set all used random seed
# see determined\pytorch\_pytorch_trial.py#Line78 for details
init_seeds(self.det_exp_cfg['reproducibility']['experiment_seed'])
if self.dataset_name == 'nyud':
self.train_datasets = NYUD(data_dir=self.dataset_dir,
data_list_1=os.path.join(self.dataset_dir, 'list/training_seg.txt'),
data_list_2=os.path.join(self.dataset_dir, 'list/training_normal_mask.txt'),
output_size=self.cfg.TRAIN.OUTPUT_SIZE,
random_scale=self.cfg.TRAIN.RANDOM_SCALE,
random_mirror=self.cfg.TRAIN.RANDOM_MIRROR,
random_crop=self.cfg.TRAIN.RANDOM_CROP,
color_jitter=self.cfg.TRAIN.RANDOM_JITTER,
ignore_label=self.cfg.IGNORE_LABEL)
self.test_datasets = NYUD(data_dir=self.dataset_dir,
data_list_1=os.path.join(self.dataset_dir, 'list/testing_seg.txt'),
data_list_2=os.path.join(self.dataset_dir, 'list/testing_normal_mask.txt'),
output_size=self.cfg.TEST.OUTPUT_SIZE,
random_scale=self.cfg.TEST.RANDOM_SCALE,
random_mirror=self.cfg.TEST.RANDOM_MIRROR,
random_crop=self.cfg.TEST.RANDOM_CROP,
color_jitter=self.cfg.TEST.RANDOM_JITTER,
ignore_label=self.cfg.IGNORE_LABEL)
else:
raise NotImplementedError('[EE] Notimplemented datasets: {}'.format(self.dataset_name))
# multi-task
if self.n_tasks > 1:
backbones = {}
is_deeplabv3 = False
for i in range(self.n_tasks):
if self.backbone_type == 'VGG16V1':
_backbone = DeepLabLargeFOVBN(in_dim=3,
out_dim=self.n_output_per_task[i],
weights_dir=self.weight_dir,
weights=self.weight_per_task[i],
dataset=self.dataset_name)
backbones.update({'task_{}'.format(i): _backbone})
elif self.backbone_type == 'R50V1':
_backbone = resnet50(pretrained=False,
weights_dir=self.weight_dir,
weights=self.weight_per_task[i],
num_classes=self.n_output_per_task[i],
dataset=self.dataset_name,
# os=16
# replace_stride_with_dilation=[False, False, True],
# os=8
replace_stride_with_dilation=[False, True, True],
)
backbones.update({'task_{}'.format(i): _backbone})
elif self.backbone_type == 'R101V1':
_backbone = resnet101(pretrained=False,
weights_dir=self.weight_dir,
weights=self.weight_per_task[i],
num_classes=self.n_output_per_task[i],
dataset=self.dataset_name,
# os=16
# replace_stride_with_dilation=[False, False, True],
# os=8
replace_stride_with_dilation=[False, True, True],
)
backbones.update({'task_{}'.format(i): _backbone})
elif self.backbone_type == 'R50V3':
_backbone = DeepLabV3(backbone_name='resnet50',
pretrained=False,
weights_dir=self.weight_dir,
weight=self.weight_per_task[i],
num_classes=self.n_output_per_task[i],
# os=16
# replace_stride_with_dilation=[False, False, True],
# aspp_dilated=[6, 12, 18],
# os=8
replace_stride_with_dilation=[False, True, True],
aspp_dilated=[12, 24, 36]
)
backbones.update({'task_{}'.format(i): _backbone})
is_deeplabv3=True
elif self.backbone_type == 'R101V3':
_backbone = DeepLabV3(backbone_name='resnet101',
pretrained=False,
weights_dir=self.weight_dir,
weight=self.weight_per_task[i],
num_classes=self.n_output_per_task[i],
# os=16
# replace_stride_with_dilation=[False, False, True],
# aspp_dilated=[6, 12, 18],
# os=8
replace_stride_with_dilation=[False, True, True],
aspp_dilated=[12, 24, 36]
)
backbones.update({'task_{}'.format(i): _backbone})
is_deeplabv3=True
elif self.backbone_type == 'R50V3P':
_backbone = DeepLabV3Plus(backbone_name='resnet50',
pretrained=False,
weights_dir=self.weight_dir,
weight=self.weight_per_task[i],
num_classes=self.n_output_per_task[i],
# os=16
# replace_stride_with_dilation=[False, False, True],
# aspp_dilated=[6, 12, 18],
# os=8
replace_stride_with_dilation=[False, True, True],
aspp_dilated=[12, 24, 36]
)
backbones.update({'task_{}'.format(i): _backbone})
is_deeplabv3=True
elif self.backbone_type == 'R101V3P':
_backbone = DeepLabV3Plus(backbone_name='resnet101',
pretrained=False,
weights_dir=self.weight_dir,
weight=self.weight_per_task[i],
num_classes=self.n_output_per_task[i],
# os=16
# replace_stride_with_dilation=[False, False, True],
# aspp_dilated=[6, 12, 18],
# os=8
replace_stride_with_dilation=[False, True, True],
aspp_dilated=[12, 24, 36]
)
backbones.update({'task_{}'.format(i): _backbone})
is_deeplabv3=True
else:
raise NotImplementedError('[EE] Unimplemented backbone: {}'.format(self.backbone_type))
_mtl_net = UNITENet(backbones=backbones,
n_tasks=self.n_tasks,
n_patch=self.n_patches,
reconstruct_src=self.reconstruct_src,
init_weights=self.init_weights,
init_method=self.init_method,
shortcut=self.shortcut,
bn_before_activation=self.bn_before_activation,
similarity=self.similarity,
is_deeplabv3=is_deeplabv3)
# self.mem_tracker.track()
self.model = self.context.wrap_model(_mtl_net)
# self.mem_tracker.track()
# self.reporter = MemReporter(self.model)
# print('========= after to =========')
# self.reporter.report(verbose=True)
no_decay = []
base_param = []
base_nd = []
fc8_weights = []
fc8_bias = []
head_param=[]
head_nd=[]
unites_param = []
unites_nd = []
classifier_params = []
classifier_nd = []
for k, v in self.model.named_parameters():
if 'unites' in k:
if any(nd in k for nd in no_decay):
unites_nd.append(v)
else:
unites_param.append(v)
elif 'head' in k:
# not head
if 'head.8' in k:
if 'weight' in k:
fc8_weights.append(v)
else:
assert 'bias' in k
fc8_bias.append(v)
else:
if any(nd in k for nd in no_decay):
head_nd.append(v)
else:
head_param.append(v)
# only v3 and v3p use this
elif 'classifier' in k:
if any(nd in k for nd in no_decay):
classifier_nd.append(v)
else:
classifier_params.append(v)
else:
# filter out fc
if 'fc' in k:
print('skip fc param with name {}'.format(k))
continue
if any(nd in k for nd in no_decay):
base_nd.append(v)
else:
base_param.append(v)
optimizer_grouped_parameters = [
{'params': base_param,
'lr': self.base_lr * self.base_factor},
{'params': unites_param,
'lr': self.base_lr * self.unites_factor},
]
# for onecos
max_lrs=[
self.base_lr * self.base_factor,
self.base_lr * self.unites_factor,
]
if len(no_decay)>0:
optimizer_grouped_parameters+=[
{'params': base_nd,
'lr': self.base_lr * self.base_factor,
'weight_decay': 0.0},
{'params': unites_nd,
'lr': self.base_lr * self.unites_factor,
'weight_decay': 0.0},
]
max_lrs+=[
self.base_lr * self.base_factor,
self.base_lr * self.unites_factor,
]
if 'V1' in self.backbone_type:
optimizer_grouped_parameters += [
{'params': fc8_weights,
'lr': self.base_lr * self.fc8_w_factor},
{'params': fc8_bias,
'lr': self.base_lr * self.fc8_b_factor,
'weight_decay': 0.0 if len(no_decay)>0 else self.weight_decay},
]
max_lrs += [
self.base_lr * self.fc8_w_factor,
self.base_lr * self.fc8_b_factor,
]
optimizer_grouped_parameters += [
{'params': head_param,
'lr': self.base_lr * self.head_factor},
]
max_lrs += [
self.base_lr * self.head_factor,
]
if len(no_decay) > 0:
optimizer_grouped_parameters += [
{'params': head_nd,
'lr': self.base_lr * self.head_factor,
'weight_decay': 0.0},
]
max_lrs += [
self.base_lr * self.head_factor,
]
elif 'V3' in self.backbone_type:
optimizer_grouped_parameters += [
{'params': classifier_params,
'lr': self.base_lr * self.classifier_factor},
]
max_lrs += [
self.base_lr * self.classifier_factor,
]
if len(no_decay) > 0:
optimizer_grouped_parameters += [
{'params': classifier_nd,
'lr': self.base_lr * self.classifier_factor,
'weight_decay': 0.0},
]
max_lrs += [
self.base_lr * self.classifier_factor,
]
self.optimizer = self.context.wrap_optimizer(optim.SGD(optimizer_grouped_parameters,
lr=self.base_lr,
weight_decay=self.weight_decay,
momentum=self.momentum))
self.lr_scheduler = self.context.wrap_lr_scheduler(
optim.lr_scheduler.OneCycleLR(self.optimizer,
# need to be same with the num of parameter groups
max_lr=max_lrs,
total_steps=self.det_exp_cfg['searcher']['max_length']['epochs']
),
LRScheduler.StepMode.STEP_EVERY_EPOCH,
frequency=1)
self.loss_func = MTLLosses(tasks=self.tasks, weight=self.loss_factors)
self.training_reducer = self.context.wrap_reducer(MTLPredictionsReducer(self.tasks, is_training=True),
for_training=True, for_validation=False)
else:
pass
def build_training_data_loader(self) -> DataLoader:
return DataLoader(self.train_datasets, batch_size=self.context.get_per_slot_batch_size(), shuffle=True,
num_workers=self.n_workers, pin_memory=True)
def build_validation_data_loader(self) -> DataLoader:
# not compatible for eval_full_datasets
# self.eval_reducer = self.context.wrap_reducer(MTLPredictionsReducer(self.tasks, is_training=False),
# for_training=False, for_validation=True)
return DataLoader(self.test_datasets, batch_size=self.context.get_per_slot_batch_size(), shuffle=False,
num_workers=self.n_workers, pin_memory=True)
def train_batch(self, batch: TorchData, epoch_idx: int, batch_idx: int) -> Dict[str, Any]:
# self.mem_tracker.track()
# print('========= after data to =========')
# self.reporter.report(verbose=True)
# if batch_idx == 0 or self.last_epoch_idx < epoch_idx:
# current_lr = self.lr_scheduler.get_last_lr()[0]
# print("Epoch: {} lr {}".format(epoch_idx, current_lr))
# self.last_epcoch_idx = epoch_idx
img = batch[0]
seg_idx = self.tasks.index('seg')
labels = {'task_{}'.format(t): batch[t + 1] if t != seg_idx else batch[t + 1].squeeze(1) for t in
range(self.n_tasks)}
features = {'task_{}'.format(i): img for i in range(self.n_tasks)}
out = self.model(features)
loss = self.loss_func(out, labels)
# print('========= before backward =========')
# self.reporter.report(verbose=True)
self.context.backward(loss['weighted_loss'])
# print('========= after backward =========')
# self.reporter.report(verbose=True)
self.context.step_optimizer(self.optimizer)
# print('========= after train =========')
# self.reporter.report(verbose=True)
# the lr scheduler step has already adopted by pytorch trial
# see determined\pytorch\_pytorch_trial.py#Line404
# self.lr_scheduler.step()
self.training_reducer.update_train({k: loss['task_{}'.format(i)] for i, k in enumerate(self.tasks)})
# print(torch.cuda.memory_summary())
# print(torch.cuda.memory_stats_as_nested_dict())
# the loss will be mean-reduced across the epoch
return {k: v if k != 'weighted_loss' else v.item() for k, v in loss.items()}
# fixme: may be there are some undefined behaviors for segmentation validation
# due to the distributed determined-ai reducers
# def evaluate_batch(self, batch: TorchData, batch_idx: int) -> Dict[str, Any]:
# img = batch[0]
# seg_idx = self.tasks.index('seg')
# labels = {'task_{}'.format(t): batch[t + 1] if t != seg_idx else batch[t + 1].squeeze(1) for t in
# range(self.n_tasks)}
#
# seg_pred, normal_pred = self.model(img, img)
# out = {'task_0': seg_pred, 'task_1': normal_pred}
# loss = self.loss_func(out, labels)
#
# self.eval_reducer.update_eval(out, labels, {k: loss['task_{}'.format(i)] for i, k in enumerate(self.tasks)})
#
# metrics = {k: v for k, v in loss.items() if k != 'weighted_loss'}
# metrics.update({'validation_loss': loss['weighted_loss'].item()})
#
# return metrics
def evaluate_full_dataset(self, data_loader: DataLoader) -> Dict[str, Any]:
# only evaluate model with single slot
_cnt = 0
# fixme: currently only implemented for normal, segmentation
if 'normal' in self.tasks:
cosine = [] # type: List[np.array]
cosine_mask = [] # type: List[np.array]
_cnt += 1
if 'seg' in self.tasks:
conf_mat_sum = 0
_cnt += 1
assert self.n_tasks == _cnt
n_batch = len(data_loader)
batch_metrics = [] # type: List[Dict]
for _, batch in enumerate(data_loader):
# nddr_attention
# img = self.context.to_device(batch[0])
# seg_idx = self.tasks.index('seg')
# labels = {
# 'task_{}'.format(t): self.context.to_device(batch[t + 1]) if t != seg_idx
# else self.context.to_device(batch[t + 1]).squeeze(1) for t in range(self.n_tasks)}
#
# seg_pred, normal_pred = self.model(img, img)
# out = {'task_0': seg_pred, 'task_1': normal_pred}
# unites
img = self.context.to_device(batch[0])
seg_idx = self.tasks.index('seg')
labels = {
'task_{}'.format(t): self.context.to_device(batch[t + 1]) if t != seg_idx
else self.context.to_device(batch[t + 1]).squeeze(1) for t in range(self.n_tasks)}
features = {'task_{}'.format(i): img for i in range(self.n_tasks)}
out = self.model(features)
loss = self.loss_func(out, labels)
# loss for each task has been applied with .detach().item()
_metrics = {k: v for k, v in loss.items() if k != 'weighted_loss'}
_metrics.update({'validation_loss': loss['weighted_loss'].item()})
batch_metrics.append(_metrics)
for i, t in enumerate(self.tasks):
if t == 'seg':
# fixme: make the num of classes be configurable
_conf_mat = get_conf_mat(out['task_{}'.format(i)], labels['task_{}'.format(i)], 40)
conf_mat_sum += _conf_mat.cpu().numpy()
elif t == 'normal':
_cosine, _mask = get_normal_cosine(prediction=out['task_{}'.format(i)],
gt=labels['task_{}'.format(i)], normalization=False,
ignore_label=255)
cosine.append(_cosine.cpu().numpy())
cosine_mask.append(_mask.cpu().numpy())
else:
raise NotImplementedError('[EE] Unimplemented tasks: {}'.format(i))
# statistics
_reduced_metrics = {'task_losses': {}}
_reduced_metrics.update({'validation_loss': sum([p['validation_loss'] for p in batch_metrics]) / n_batch})
_cnt = 0
if 'normal' in self.tasks:
normal_idx = self.tasks.index('normal')
_normal_loss = sum([p['task_{}'.format(normal_idx)] for p in batch_metrics]) / n_batch
_reduced_metrics['task_losses'].update({'normal': _normal_loss})
_cos_angels = [c[m] for c, m in zip(cosine, cosine_mask)]
normal_metrics = get_normal_metrics(_cos_angels)
_reduced_metrics.update({**normal_metrics})
_cnt += 1
if 'seg' in self.tasks:
seg_idx = self.tasks.index('seg')
_seg_loss = sum([p['task_{}'.format(seg_idx)] for p in batch_metrics]) / n_batch
_reduced_metrics['task_losses'].update({'seg': _seg_loss})
_mIoU = get_mIoU(conf_mat_sum) * 100.
_pixel_acc = np.diag(conf_mat_sum).sum() / np.sum(conf_mat_sum) * 100.
_reduced_metrics.update({'mIoU': _mIoU, 'pixel_acc': _pixel_acc})
_cnt += 1
assert self.n_tasks == _cnt
return _reduced_metrics