-
Notifications
You must be signed in to change notification settings - Fork 2.9k
/
trainer.py
1510 lines (1338 loc) · 62.9 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
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
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import os
import sys
import copy
import time
import yaml
from tqdm import tqdm
import numpy as np
import typing
from PIL import Image, ImageOps, ImageFile
ImageFile.LOAD_TRUNCATED_IMAGES = True
import paddle
import paddle.nn as nn
import paddle.distributed as dist
from paddle.distributed import fleet
from paddle.static import InputSpec
from ppdet.optimizer import ModelEMA
from ppdet.core.workspace import create
from ppdet.utils.checkpoint import load_weight, load_pretrain_weight, convert_to_dict
from ppdet.utils.visualizer import visualize_results, save_result
from ppdet.metrics import get_infer_results, KeyPointTopDownCOCOEval, KeyPointTopDownCOCOWholeBadyHandEval, KeyPointTopDownMPIIEval, Pose3DEval
from ppdet.metrics import Metric, COCOMetric, VOCMetric, WiderFaceMetric, RBoxMetric, JDEDetMetric, SNIPERCOCOMetric, CULaneMetric
from ppdet.data.source.sniper_coco import SniperCOCODataSet
from ppdet.data.source.category import get_categories
import ppdet.utils.stats as stats
from ppdet.utils.fuse_utils import fuse_conv_bn
from ppdet.utils import profiler
from ppdet.modeling.post_process import multiclass_nms
from ppdet.modeling.lane_utils import imshow_lanes
from .callbacks import Callback, ComposeCallback, LogPrinter, Checkpointer, WiferFaceEval, VisualDLWriter, SniperProposalsGenerator, WandbCallback, SemiCheckpointer, SemiLogPrinter
from .export_utils import _dump_infer_config, _prune_input_spec, apply_to_static
from .naive_sync_bn import convert_syncbn
from paddle.distributed.fleet.utils.hybrid_parallel_util import fused_allreduce_gradients
from ppdet.utils.logger import setup_logger
logger = setup_logger('ppdet.engine')
__all__ = ['Trainer']
MOT_ARCH = ['JDE', 'FairMOT', 'DeepSORT', 'ByteTrack', 'CenterTrack']
class Trainer(object):
def __init__(self, cfg, mode='train'):
self.cfg = cfg.copy()
assert mode.lower() in ['train', 'eval', 'test'], \
"mode should be 'train', 'eval' or 'test'"
self.mode = mode.lower()
self.optimizer = None
self.is_loaded_weights = False
self.use_amp = self.cfg.get('amp', False)
self.amp_level = self.cfg.get('amp_level', 'O1')
self.custom_white_list = self.cfg.get('custom_white_list', None)
self.custom_black_list = self.cfg.get('custom_black_list', None)
self.use_master_grad = self.cfg.get('master_grad', False)
self.uniform_output_enabled = self.cfg.get('uniform_output_enabled', False)
if ('slim' in cfg and cfg['slim_type'] == 'PTQ') or self.uniform_output_enabled:
self.cfg['TestDataset'] = create('TestDataset')()
log_ranks = cfg.get('log_ranks', '0')
if isinstance(log_ranks, str):
self.log_ranks = [int(i) for i in log_ranks.split(',')]
elif isinstance(log_ranks, int):
self.log_ranks = [log_ranks]
train_results_path = os.path.abspath(os.path.join(self.cfg.save_dir, "train_results.json"))
if self.uniform_output_enabled:
if os.path.exists(train_results_path) and self.mode == 'train':
try:
os.remove(train_results_path)
except:
pass
if not os.path.exists(self.cfg.save_dir):
os.mkdir(self.cfg.save_dir)
with open(os.path.join(self.cfg.save_dir, "config.yaml"), "w") as f:
config_dict = convert_to_dict(self.cfg)
config_dict = {k: v for k, v in config_dict.items() if v != {}}
yaml.dump(config_dict, f)
# build data loader
capital_mode = self.mode.capitalize()
if cfg.architecture in MOT_ARCH and self.mode in [
'eval', 'test'
] and cfg.metric not in ['COCO', 'VOC']:
self.dataset = self.cfg['{}MOTDataset'.format(
capital_mode)] = create('{}MOTDataset'.format(capital_mode))()
else:
self.dataset = self.cfg['{}Dataset'.format(capital_mode)] = create(
'{}Dataset'.format(capital_mode))()
if cfg.architecture == 'DeepSORT' and self.mode == 'train':
logger.error('DeepSORT has no need of training on mot dataset.')
sys.exit(1)
if cfg.architecture == 'FairMOT' and self.mode == 'eval':
images = self.parse_mot_images(cfg)
self.dataset.set_images(images)
if self.mode == 'train':
self.loader = create('{}Reader'.format(capital_mode))(
self.dataset, cfg.worker_num)
if cfg.architecture == 'JDE' and self.mode == 'train':
self.cfg['JDEEmbeddingHead'][
'num_identities'] = self.dataset.num_identities_dict[0]
# JDE only support single class MOT now.
if cfg.architecture == 'FairMOT' and self.mode == 'train':
self.cfg['FairMOTEmbeddingHead'][
'num_identities_dict'] = self.dataset.num_identities_dict
# FairMOT support single class and multi-class MOT now.
# build model
if 'model' not in self.cfg:
self.model = create(cfg.architecture)
else:
self.model = self.cfg.model
self.is_loaded_weights = True
if cfg.architecture == 'YOLOX':
for k, m in self.model.named_sublayers():
if isinstance(m, nn.BatchNorm2D):
m._epsilon = 1e-3 # for amp(fp16)
m._momentum = 0.97 # 0.03 in pytorch
# reset norm param attr for setting them in optimizer
if 'reset_norm_param_attr' in cfg and cfg['reset_norm_param_attr']:
self.model = self.reset_norm_param_attr(
self.model, weight_attr=None, bias_attr=None)
# normalize params for deploy
if 'slim' in cfg and cfg['slim_type'] == 'OFA':
self.model.model.load_meanstd(cfg['TestReader'][
'sample_transforms'])
elif 'slim' in cfg and cfg['slim_type'] == 'Distill':
self.model.student_model.load_meanstd(cfg['TestReader'][
'sample_transforms'])
elif 'slim' in cfg and cfg[
'slim_type'] == 'DistillPrune' and self.mode == 'train':
self.model.student_model.load_meanstd(cfg['TestReader'][
'sample_transforms'])
else:
self.model.load_meanstd(cfg['TestReader']['sample_transforms'])
# EvalDataset build with BatchSampler to evaluate in single device
# TODO: multi-device evaluate
if self.mode == 'eval':
if cfg.architecture == 'FairMOT':
self.loader = create('EvalMOTReader')(self.dataset, 0)
elif cfg.architecture == "METRO_Body":
reader_name = '{}Reader'.format(self.mode.capitalize())
self.loader = create(reader_name)(self.dataset, cfg.worker_num)
else:
self._eval_batch_sampler = paddle.io.BatchSampler(
self.dataset, batch_size=self.cfg.EvalReader['batch_size'])
reader_name = '{}Reader'.format(self.mode.capitalize())
# If metric is VOC, need to be set collate_batch=False.
if cfg.metric == 'VOC':
self.cfg[reader_name]['collate_batch'] = False
self.loader = create(reader_name)(self.dataset, cfg.worker_num,
self._eval_batch_sampler)
# TestDataset build after user set images, skip loader creation here
# get Params
print_params = self.cfg.get('print_params', False)
if print_params:
params = sum([
p.numel() for n, p in self.model.named_parameters()
if all([x not in n for x in ['_mean', '_variance', 'aux_']])
]) # exclude BatchNorm running status
logger.info('Model Params : {} M.'.format((params / 1e6).numpy()[
0]))
# build optimizer in train mode
if self.mode == 'train':
steps_per_epoch = len(self.loader)
if steps_per_epoch < 1:
logger.warning(
"Samples in dataset are less than batch_size, please set smaller batch_size in TrainReader."
)
self.lr = create('LearningRate')(steps_per_epoch)
self.optimizer = create('OptimizerBuilder')(self.lr, self.model)
# Unstructured pruner is only enabled in the train mode.
if self.cfg.get('unstructured_prune'):
self.pruner = create('UnstructuredPruner')(self.model,
steps_per_epoch)
if self.use_amp and self.amp_level == 'O2':
paddle_version = paddle.__version__[:3]
# paddle version >= 2.5.0 or develop
if paddle_version in ["2.5", "0.0"]:
self.model, self.optimizer = paddle.amp.decorate(
models=self.model,
optimizers=self.optimizer,
level=self.amp_level,
master_grad=self.use_master_grad)
else:
self.model, self.optimizer = paddle.amp.decorate(
models=self.model,
optimizers=self.optimizer,
level=self.amp_level)
# support sync_bn for npu/xpu
if (paddle.get_device()[:3]=='npu' or paddle.get_device()[:3]=='xpu'):
use_npu = ('use_npu' in cfg and cfg['use_npu'])
use_xpu = ('use_xpu' in cfg and cfg['use_xpu'])
norm_type = ('norm_type' in cfg and cfg['norm_type'])
if norm_type == 'sync_bn' and (use_npu or use_xpu) and dist.get_world_size() > 1:
convert_syncbn(self.model)
self.use_ema = ('use_ema' in cfg and cfg['use_ema'])
if self.use_ema:
ema_decay = self.cfg.get('ema_decay', 0.9998)
ema_decay_type = self.cfg.get('ema_decay_type', 'threshold')
cycle_epoch = self.cfg.get('cycle_epoch', -1)
ema_black_list = self.cfg.get('ema_black_list', None)
ema_filter_no_grad = self.cfg.get('ema_filter_no_grad', False)
self.ema = ModelEMA(
self.model,
decay=ema_decay,
ema_decay_type=ema_decay_type,
cycle_epoch=cycle_epoch,
ema_black_list=ema_black_list,
ema_filter_no_grad=ema_filter_no_grad)
self._nranks = dist.get_world_size()
self._local_rank = dist.get_rank()
self.status = {}
self.start_epoch = 0
self.end_epoch = 0 if 'epoch' not in cfg else cfg.epoch
# initial default callbacks
self._init_callbacks()
# initial default metrics
self._init_metrics()
self._reset_metrics()
def _init_callbacks(self):
if self.mode == 'train':
if self.cfg.get('ssod_method',
False) and self.cfg['ssod_method'] == 'Semi_RTDETR':
self._callbacks = [SemiLogPrinter(self), SemiCheckpointer(self)]
else:
self._callbacks = [LogPrinter(self), Checkpointer(self)]
if self.cfg.get('use_vdl', False):
self._callbacks.append(VisualDLWriter(self))
if self.cfg.get('save_proposals', False):
self._callbacks.append(SniperProposalsGenerator(self))
if self.cfg.get('use_wandb', False) or 'wandb' in self.cfg:
self._callbacks.append(WandbCallback(self))
self._compose_callback = ComposeCallback(self._callbacks)
elif self.mode == 'eval':
self._callbacks = [LogPrinter(self)]
if self.cfg.metric == 'WiderFace':
self._callbacks.append(WiferFaceEval(self))
self._compose_callback = ComposeCallback(self._callbacks)
elif self.mode == 'test' and self.cfg.get('use_vdl', False):
self._callbacks = [VisualDLWriter(self)]
self._compose_callback = ComposeCallback(self._callbacks)
else:
self._callbacks = []
self._compose_callback = None
def _init_metrics(self, validate=False):
if self.mode == 'test' or (self.mode == 'train' and not validate):
self._metrics = []
return
classwise = self.cfg['classwise'] if 'classwise' in self.cfg else False
if self.cfg.metric == 'COCO' or self.cfg.metric == "SNIPERCOCO":
# TODO: bias should be unified
bias = 1 if self.cfg.get('bias', False) else 0
output_eval = self.cfg['output_eval'] \
if 'output_eval' in self.cfg else None
save_prediction_only = self.cfg.get('save_prediction_only', False)
# pass clsid2catid info to metric instance to avoid multiple loading
# annotation file
clsid2catid = {v: k for k, v in self.dataset.catid2clsid.items()} \
if self.mode == 'eval' else None
save_threshold = self.cfg.get('save_threshold', 0)
# when do validation in train, annotation file should be get from
# EvalReader instead of self.dataset(which is TrainReader)
if self.mode == 'train' and validate:
eval_dataset = self.cfg['EvalDataset']
eval_dataset.check_or_download_dataset()
anno_file = eval_dataset.get_anno()
dataset = eval_dataset
else:
dataset = self.dataset
anno_file = dataset.get_anno()
IouType = self.cfg['IouType'] if 'IouType' in self.cfg else 'bbox'
if self.cfg.metric == "COCO":
self._metrics = [
COCOMetric(
anno_file=anno_file,
clsid2catid=clsid2catid,
classwise=classwise,
output_eval=output_eval,
bias=bias,
IouType=IouType,
save_prediction_only=save_prediction_only,
save_threshold=save_threshold)
]
elif self.cfg.metric == "SNIPERCOCO": # sniper
self._metrics = [
SNIPERCOCOMetric(
anno_file=anno_file,
dataset=dataset,
clsid2catid=clsid2catid,
classwise=classwise,
output_eval=output_eval,
bias=bias,
IouType=IouType,
save_prediction_only=save_prediction_only)
]
elif self.cfg.metric == 'RBOX':
# TODO: bias should be unified
bias = self.cfg['bias'] if 'bias' in self.cfg else 0
output_eval = self.cfg['output_eval'] \
if 'output_eval' in self.cfg else None
save_prediction_only = self.cfg.get('save_prediction_only', False)
imid2path = self.cfg.get('imid2path', None)
# when do validation in train, annotation file should be get from
# EvalReader instead of self.dataset(which is TrainReader)
anno_file = self.dataset.get_anno()
if self.mode == 'train' and validate:
eval_dataset = self.cfg['EvalDataset']
eval_dataset.check_or_download_dataset()
anno_file = eval_dataset.get_anno()
self._metrics = [
RBoxMetric(
anno_file=anno_file,
classwise=classwise,
output_eval=output_eval,
bias=bias,
save_prediction_only=save_prediction_only,
imid2path=imid2path)
]
elif self.cfg.metric == 'VOC':
output_eval = self.cfg['output_eval'] \
if 'output_eval' in self.cfg else None
save_prediction_only = self.cfg.get('save_prediction_only', False)
self._metrics = [
VOCMetric(
label_list=self.dataset.get_label_list(),
class_num=self.cfg.num_classes,
map_type=self.cfg.map_type,
classwise=classwise,
output_eval=output_eval,
save_prediction_only=save_prediction_only)
]
elif self.cfg.metric == 'WiderFace':
multi_scale = self.cfg.multi_scale_eval if 'multi_scale_eval' in self.cfg else True
self._metrics = [
WiderFaceMetric(
image_dir=os.path.join(self.dataset.dataset_dir,
self.dataset.image_dir),
anno_file=self.dataset.get_anno(),
multi_scale=multi_scale)
]
elif self.cfg.metric == 'KeyPointTopDownCOCOEval':
eval_dataset = self.cfg['EvalDataset']
eval_dataset.check_or_download_dataset()
anno_file = eval_dataset.get_anno()
save_prediction_only = self.cfg.get('save_prediction_only', False)
self._metrics = [
KeyPointTopDownCOCOEval(
anno_file,
len(eval_dataset),
self.cfg.num_joints,
self.cfg.save_dir,
save_prediction_only=save_prediction_only)
]
elif self.cfg.metric == 'KeyPointTopDownCOCOWholeBadyHandEval':
eval_dataset = self.cfg['EvalDataset']
eval_dataset.check_or_download_dataset()
anno_file = eval_dataset.get_anno()
save_prediction_only = self.cfg.get('save_prediction_only', False)
self._metrics = [
KeyPointTopDownCOCOWholeBadyHandEval(
anno_file,
len(eval_dataset),
self.cfg.num_joints,
self.cfg.save_dir,
save_prediction_only=save_prediction_only)
]
elif self.cfg.metric == 'KeyPointTopDownMPIIEval':
eval_dataset = self.cfg['EvalDataset']
eval_dataset.check_or_download_dataset()
anno_file = eval_dataset.get_anno()
save_prediction_only = self.cfg.get('save_prediction_only', False)
self._metrics = [
KeyPointTopDownMPIIEval(
anno_file,
len(eval_dataset),
self.cfg.num_joints,
self.cfg.save_dir,
save_prediction_only=save_prediction_only)
]
elif self.cfg.metric == 'Pose3DEval':
save_prediction_only = self.cfg.get('save_prediction_only', False)
self._metrics = [
Pose3DEval(
self.cfg.save_dir,
save_prediction_only=save_prediction_only)
]
elif self.cfg.metric == 'MOTDet':
self._metrics = [JDEDetMetric(), ]
elif self.cfg.metric == 'CULaneMetric':
output_eval = self.cfg.get('output_eval', None)
self._metrics = [
CULaneMetric(
cfg=self.cfg,
output_eval=output_eval,
split=self.dataset.split,
dataset_dir=self.cfg.dataset_dir)
]
else:
logger.warning("Metric not support for metric type {}".format(
self.cfg.metric))
self._metrics = []
def _reset_metrics(self):
for metric in self._metrics:
metric.reset()
def register_callbacks(self, callbacks):
callbacks = [c for c in list(callbacks) if c is not None]
for c in callbacks:
assert isinstance(c, Callback), \
"metrics shoule be instances of subclass of Metric"
self._callbacks.extend(callbacks)
self._compose_callback = ComposeCallback(self._callbacks)
def register_metrics(self, metrics):
metrics = [m for m in list(metrics) if m is not None]
for m in metrics:
assert isinstance(m, Metric), \
"metrics shoule be instances of subclass of Metric"
self._metrics.extend(metrics)
def load_weights(self, weights, ARSL_eval=False):
if self.is_loaded_weights:
return
self.start_epoch = 0
load_pretrain_weight(self.model, weights, ARSL_eval)
logger.debug("Load weights {} to start training".format(weights))
def load_weights_sde(self, det_weights, reid_weights):
if self.model.detector:
load_weight(self.model.detector, det_weights)
if self.model.reid:
load_weight(self.model.reid, reid_weights)
else:
load_weight(self.model.reid, reid_weights)
def resume_weights(self, weights):
# support Distill resume weights
if hasattr(self.model, 'student_model'):
self.start_epoch = load_weight(self.model.student_model, weights,
self.optimizer)
else:
self.start_epoch = load_weight(self.model, weights, self.optimizer,
self.ema if self.use_ema else None)
logger.debug("Resume weights of epoch {}".format(self.start_epoch))
def train(self, validate=False):
assert self.mode == 'train', "Model not in 'train' mode"
Init_mark = False
if validate:
self.cfg['EvalDataset'] = self.cfg.EvalDataset = create(
"EvalDataset")()
model = self.model
if self.cfg.get('to_static', False):
model = apply_to_static(self.cfg, model)
sync_bn = (getattr(self.cfg, 'norm_type', None) == 'sync_bn' and
(self.cfg.use_gpu or self.cfg.use_mlu) and self._nranks > 1)
if sync_bn:
model = paddle.nn.SyncBatchNorm.convert_sync_batchnorm(model)
# enabel auto mixed precision mode
if self.use_amp:
scaler = paddle.amp.GradScaler(
enable=self.cfg.use_gpu or self.cfg.use_npu or self.cfg.use_mlu,
init_loss_scaling=self.cfg.get('init_loss_scaling', 1024))
# get distributed model
if self.cfg.get('fleet', False):
model = fleet.distributed_model(model)
self.optimizer = fleet.distributed_optimizer(self.optimizer)
elif self._nranks > 1:
find_unused_parameters = self.cfg[
'find_unused_parameters'] if 'find_unused_parameters' in self.cfg else False
model = paddle.DataParallel(
model, find_unused_parameters=find_unused_parameters)
self.status.update({
'epoch_id': self.start_epoch,
'step_id': 0,
'steps_per_epoch': len(self.loader)
})
self.status['batch_time'] = stats.SmoothedValue(
self.cfg.log_iter, fmt='{avg:.4f}')
self.status['data_time'] = stats.SmoothedValue(
self.cfg.log_iter, fmt='{avg:.4f}')
self.status['training_staus'] = stats.TrainingStats(self.cfg.log_iter)
if self.cfg.get('print_flops', False):
flops_loader = create('{}Reader'.format(self.mode.capitalize()))(
self.dataset, self.cfg.worker_num)
self._flops(flops_loader)
profiler_options = self.cfg.get('profiler_options', None)
self._compose_callback.on_train_begin(self.status)
use_fused_allreduce_gradients = self.cfg[
'use_fused_allreduce_gradients'] if 'use_fused_allreduce_gradients' in self.cfg else False
for epoch_id in range(self.start_epoch, self.cfg.epoch):
self.status['mode'] = 'train'
self.status['epoch_id'] = epoch_id
self._compose_callback.on_epoch_begin(self.status)
self.loader.dataset.set_epoch(epoch_id)
model.train()
iter_tic = time.time()
for step_id, data in enumerate(self.loader):
def deep_pin(blob, blocking):
if isinstance(blob, paddle.Tensor):
return blob.cuda(blocking=blocking)
elif isinstance(blob, dict):
return {k: deep_pin(v, blocking) for k, v in blob.items()}
elif isinstance(blob, (list, tuple)):
return type(blob)([deep_pin(x, blocking) for x in blob])
else:
return blob
if paddle.base.core.is_compiled_with_cuda():
data = deep_pin(data, False)
self.status['data_time'].update(time.time() - iter_tic)
self.status['step_id'] = step_id
profiler.add_profiler_step(profiler_options)
self._compose_callback.on_step_begin(self.status)
data['epoch_id'] = epoch_id
if self.cfg.get('to_static',
False) and 'image_file' in data.keys():
data.pop('image_file')
if self.use_amp:
if isinstance(
model, paddle.
DataParallel) and use_fused_allreduce_gradients:
with model.no_sync():
with paddle.amp.auto_cast(
enable=self.cfg.use_gpu or
self.cfg.use_npu or self.cfg.use_mlu,
custom_white_list=self.custom_white_list,
custom_black_list=self.custom_black_list,
level=self.amp_level):
# model forward
outputs = model(data)
loss = outputs['loss']
# model backward
scaled_loss = scaler.scale(loss)
scaled_loss.backward()
fused_allreduce_gradients(
list(model.parameters()), None)
else:
with paddle.amp.auto_cast(
enable=self.cfg.use_gpu or self.cfg.use_npu or
self.cfg.use_mlu,
custom_white_list=self.custom_white_list,
custom_black_list=self.custom_black_list,
level=self.amp_level):
# model forward
outputs = model(data)
loss = outputs['loss']
# model backward
scaled_loss = scaler.scale(loss)
scaled_loss.backward()
# in dygraph mode, optimizer.minimize is equal to optimizer.step
scaler.minimize(self.optimizer, scaled_loss)
else:
if isinstance(
model, paddle.
DataParallel) and use_fused_allreduce_gradients:
with model.no_sync():
# model forward
outputs = model(data)
loss = outputs['loss']
# model backward
loss.backward()
fused_allreduce_gradients(
list(model.parameters()), None)
else:
# model forward
outputs = model(data)
loss = outputs['loss']
# model backward
loss.backward()
self.optimizer.step()
curr_lr = self.optimizer.get_lr()
self.lr.step()
if self.cfg.get('unstructured_prune'):
self.pruner.step()
self.optimizer.clear_grad()
self.status['learning_rate'] = curr_lr
if self._nranks < 2 or self._local_rank in self.log_ranks:
self.status['training_staus'].update(outputs)
self.status['batch_time'].update(time.time() - iter_tic)
self._compose_callback.on_step_end(self.status)
if self.use_ema:
self.ema.update()
iter_tic = time.time()
if self.cfg.get('unstructured_prune'):
self.pruner.update_params()
is_snapshot = (self._nranks < 2 or (self._local_rank == 0 or self.cfg.metric == "Pose3DEval")) \
and ((epoch_id + 1) % self.cfg.snapshot_epoch == 0 or epoch_id == self.end_epoch - 1)
if is_snapshot and self.use_ema:
# apply ema weight on model
weight = copy.deepcopy(self.model.state_dict())
self.model.set_dict(self.ema.apply())
self.status['weight'] = weight
self._compose_callback.on_epoch_end(self.status)
if validate and is_snapshot:
if not hasattr(self, '_eval_loader'):
# build evaluation dataset and loader
self._eval_dataset = self.cfg.EvalDataset
self._eval_batch_sampler = \
paddle.io.BatchSampler(
self._eval_dataset,
batch_size=self.cfg.EvalReader['batch_size'])
# If metric is VOC, need to be set collate_batch=False.
if self.cfg.metric == 'VOC':
self.cfg['EvalReader']['collate_batch'] = False
if self.cfg.metric == "Pose3DEval":
self._eval_loader = create('EvalReader')(
self._eval_dataset, self.cfg.worker_num)
else:
self._eval_loader = create('EvalReader')(
self._eval_dataset,
self.cfg.worker_num,
batch_sampler=self._eval_batch_sampler)
# if validation in training is enabled, metrics should be re-init
# Init_mark makes sure this code will only execute once
if validate and Init_mark == False:
Init_mark = True
self._init_metrics(validate=validate)
self._reset_metrics()
with paddle.no_grad():
self.status['save_best_model'] = True
self._eval_with_loader(self._eval_loader)
if is_snapshot and self.use_ema:
# reset original weight
self.model.set_dict(weight)
self.status.pop('weight')
self._compose_callback.on_train_end(self.status)
def _eval_with_loader(self, loader):
sample_num = 0
tic = time.time()
self._compose_callback.on_epoch_begin(self.status)
self.status['mode'] = 'eval'
self.model.eval()
if self.cfg.get('print_flops', False):
flops_loader = create('{}Reader'.format(self.mode.capitalize()))(
self.dataset, self.cfg.worker_num, self._eval_batch_sampler)
self._flops(flops_loader)
for step_id, data in enumerate(loader):
self.status['step_id'] = step_id
self._compose_callback.on_step_begin(self.status)
# forward
if self.use_amp:
with paddle.amp.auto_cast(
enable=self.cfg.use_gpu or self.cfg.use_npu or
self.cfg.use_mlu,
custom_white_list=self.custom_white_list,
custom_black_list=self.custom_black_list,
level=self.amp_level):
outs = self.model(data)
else:
outs = self.model(data)
# update metrics
for metric in self._metrics:
metric.update(data, outs)
# multi-scale inputs: all inputs have same im_id
if isinstance(data, typing.Sequence):
sample_num += data[0]['im_id'].numpy().shape[0]
else:
sample_num += data['im_id'].numpy().shape[0]
self._compose_callback.on_step_end(self.status)
self.status['sample_num'] = sample_num
self.status['cost_time'] = time.time() - tic
# accumulate metric to log out
for metric in self._metrics:
metric.accumulate()
metric.log()
self._compose_callback.on_epoch_end(self.status)
# reset metric states for metric may performed multiple times
self._reset_metrics()
def evaluate(self):
# get distributed model
if self.cfg.get('fleet', False):
self.model = fleet.distributed_model(self.model)
self.optimizer = fleet.distributed_optimizer(self.optimizer)
elif self._nranks > 1:
find_unused_parameters = self.cfg[
'find_unused_parameters'] if 'find_unused_parameters' in self.cfg else False
self.model = paddle.DataParallel(
self.model, find_unused_parameters=find_unused_parameters)
with paddle.no_grad():
self._eval_with_loader(self.loader)
def _eval_with_loader_slice(self,
loader,
slice_size=[640, 640],
overlap_ratio=[0.25, 0.25],
combine_method='nms',
match_threshold=0.6,
match_metric='iou'):
sample_num = 0
tic = time.time()
self._compose_callback.on_epoch_begin(self.status)
self.status['mode'] = 'eval'
self.model.eval()
if self.cfg.get('print_flops', False):
flops_loader = create('{}Reader'.format(self.mode.capitalize()))(
self.dataset, self.cfg.worker_num, self._eval_batch_sampler)
self._flops(flops_loader)
merged_bboxs = []
for step_id, data in enumerate(loader):
self.status['step_id'] = step_id
self._compose_callback.on_step_begin(self.status)
# forward
if self.use_amp:
with paddle.amp.auto_cast(
enable=self.cfg.use_gpu or self.cfg.use_npu or
self.cfg.use_mlu,
custom_white_list=self.custom_white_list,
custom_black_list=self.custom_black_list,
level=self.amp_level):
outs = self.model(data)
else:
outs = self.model(data)
shift_amount = data['st_pix']
outs['bbox'][:, 2:4] = outs['bbox'][:, 2:4] + shift_amount
outs['bbox'][:, 4:6] = outs['bbox'][:, 4:6] + shift_amount
merged_bboxs.append(outs['bbox'])
if data['is_last'] > 0:
# merge matching predictions
merged_results = {'bbox': []}
if combine_method == 'nms':
final_boxes = multiclass_nms(
np.concatenate(merged_bboxs), self.cfg.num_classes,
match_threshold, match_metric)
merged_results['bbox'] = np.concatenate(final_boxes)
elif combine_method == 'concat':
merged_results['bbox'] = np.concatenate(merged_bboxs)
else:
raise ValueError(
"Now only support 'nms' or 'concat' to fuse detection results."
)
merged_results['im_id'] = np.array([[0]])
merged_results['bbox_num'] = np.array(
[len(merged_results['bbox'])])
merged_bboxs = []
data['im_id'] = data['ori_im_id']
# update metrics
for metric in self._metrics:
metric.update(data, merged_results)
# multi-scale inputs: all inputs have same im_id
if isinstance(data, typing.Sequence):
sample_num += data[0]['im_id'].numpy().shape[0]
else:
sample_num += data['im_id'].numpy().shape[0]
self._compose_callback.on_step_end(self.status)
self.status['sample_num'] = sample_num
self.status['cost_time'] = time.time() - tic
# accumulate metric to log out
for metric in self._metrics:
metric.accumulate()
metric.log()
self._compose_callback.on_epoch_end(self.status)
# reset metric states for metric may performed multiple times
self._reset_metrics()
def evaluate_slice(self,
slice_size=[640, 640],
overlap_ratio=[0.25, 0.25],
combine_method='nms',
match_threshold=0.6,
match_metric='iou'):
with paddle.no_grad():
self._eval_with_loader_slice(self.loader, slice_size, overlap_ratio,
combine_method, match_threshold,
match_metric)
def slice_predict(self,
images,
slice_size=[640, 640],
overlap_ratio=[0.25, 0.25],
combine_method='nms',
match_threshold=0.6,
match_metric='iou',
draw_threshold=0.5,
output_dir='output',
save_results=False,
visualize=True):
if not os.path.exists(output_dir):
os.makedirs(output_dir)
self.dataset.set_slice_images(images, slice_size, overlap_ratio)
loader = create('TestReader')(self.dataset, 0)
imid2path = self.dataset.get_imid2path()
def setup_metrics_for_loader():
# mem
metrics = copy.deepcopy(self._metrics)
mode = self.mode
save_prediction_only = self.cfg[
'save_prediction_only'] if 'save_prediction_only' in self.cfg else None
output_eval = self.cfg[
'output_eval'] if 'output_eval' in self.cfg else None
# modify
self.mode = '_test'
self.cfg['save_prediction_only'] = True
self.cfg['output_eval'] = output_dir
self.cfg['imid2path'] = imid2path
self._init_metrics()
# restore
self.mode = mode
self.cfg.pop('save_prediction_only')
if save_prediction_only is not None:
self.cfg['save_prediction_only'] = save_prediction_only
self.cfg.pop('output_eval')
if output_eval is not None:
self.cfg['output_eval'] = output_eval
self.cfg.pop('imid2path')
_metrics = copy.deepcopy(self._metrics)
self._metrics = metrics
return _metrics
if save_results:
metrics = setup_metrics_for_loader()
else:
metrics = []
anno_file = self.dataset.get_anno()
clsid2catid, catid2name = get_categories(
self.cfg.metric, anno_file=anno_file)
# Run Infer
self.status['mode'] = 'test'
self.model.eval()
if self.cfg.get('print_flops', False):
flops_loader = create('TestReader')(self.dataset, 0)
self._flops(flops_loader)
results = [] # all images
merged_bboxs = [] # single image
for step_id, data in enumerate(tqdm(loader)):
self.status['step_id'] = step_id
# forward
outs = self.model(data)
outs['bbox'] = outs['bbox'].numpy() # only in test mode
shift_amount = data['st_pix']
outs['bbox'][:, 2:4] = outs['bbox'][:, 2:4] + shift_amount.numpy()
outs['bbox'][:, 4:6] = outs['bbox'][:, 4:6] + shift_amount.numpy()
merged_bboxs.append(outs['bbox'])
if data['is_last'] > 0:
# merge matching predictions
merged_results = {'bbox': []}
if combine_method == 'nms':
final_boxes = multiclass_nms(
np.concatenate(merged_bboxs), self.cfg.num_classes,
match_threshold, match_metric)
merged_results['bbox'] = np.concatenate(final_boxes)
elif combine_method == 'concat':
merged_results['bbox'] = np.concatenate(merged_bboxs)
else:
raise ValueError(
"Now only support 'nms' or 'concat' to fuse detection results."
)
merged_results['im_id'] = np.array([[0]])
merged_results['bbox_num'] = np.array(
[len(merged_results['bbox'])])
merged_bboxs = []
data['im_id'] = data['ori_im_id']
for _m in metrics:
_m.update(data, merged_results)
for key in ['im_shape', 'scale_factor', 'im_id']:
if isinstance(data, typing.Sequence):
merged_results[key] = data[0][key]
else:
merged_results[key] = data[key]
for key, value in merged_results.items():
if hasattr(value, 'numpy'):
merged_results[key] = value.numpy()
results.append(merged_results)
for _m in metrics:
_m.accumulate()
_m.reset()
if visualize:
for outs in results:
batch_res = get_infer_results(outs, clsid2catid)
bbox_num = outs['bbox_num']
start = 0
for i, im_id in enumerate(outs['im_id']):
image_path = imid2path[int(im_id)]
image = Image.open(image_path).convert('RGB')
image = ImageOps.exif_transpose(image)
self.status['original_image'] = np.array(image.copy())
end = start + bbox_num[i]
bbox_res = batch_res['bbox'][start:end] \
if 'bbox' in batch_res else None
mask_res = batch_res['mask'][start:end] \
if 'mask' in batch_res else None
segm_res = batch_res['segm'][start:end] \
if 'segm' in batch_res else None
keypoint_res = batch_res['keypoint'][start:end] \
if 'keypoint' in batch_res else None
pose3d_res = batch_res['pose3d'][start:end] \
if 'pose3d' in batch_res else None
image = visualize_results(
image, bbox_res, mask_res, segm_res, keypoint_res,
pose3d_res, int(im_id), catid2name, draw_threshold)
self.status['result_image'] = np.array(image.copy())
if self._compose_callback:
self._compose_callback.on_step_end(self.status)
# save image with detection
save_name = self._get_save_image_name(output_dir,
image_path)