-
Notifications
You must be signed in to change notification settings - Fork 9
/
train1.py
1130 lines (991 loc) · 53.9 KB
/
train1.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
import argparse
import torch.distributed as dist
import torch.optim as optim
import torch.optim.lr_scheduler as lr_scheduler
from torch.utils.tensorboard import SummaryWriter
import test # import test.py to get mAP after each epoch
from models import *
from utils.datasets import *
from utils.utils import *
from utils.prune_utils import *
import math
mixed_precision = True
try: # Mixed precision training https://github.com/NVIDIA/apex
from apex import amp
except:
print('Apex recommended for faster mixed precision training: https://github.com/NVIDIA/apex')
mixed_precision = False # not installed
wdir = 'weights' + os.sep # weights dir
last = wdir + 'v4_tiny_416_map61.pt'
best = wdir + 'best.pt'
results_file = 'results.txt'
# Hyperparameters (j-series, 50.5 mAP yolov3-320) evolved by @ktian08 https://github.com/ultralytics/yolov3/issues/310
hyp = {'giou': 3.582, # giou loss gain
'cls': 37.76, # cls loss gain (CE=~1.0, uCE=~20)
'cls_pw': 1.146, # cls BCELoss positive_weight
'obj': 64.35, # obj loss gain (*=80 for uBCE with 80 classes)
'obj_pw': 1.11, # obj BCELoss positive_weight
'iou_t': 0.20, # iou training threshold
'lr0': 0.0003, # initial learning rate (SGD=1E-3, Adam=9E-5)
'lrf': 0.0005, # final LambdaLR learning rate = lr0 * (10 ** lrf)
'momentum': 0.97, # SGD momentum
'weight_decay': 0.0005, # optimizer weight decay
'fl_gamma': 0.0, # focal loss gamma
'hsv_h': 0.0138, # image HSV-Hue augmentation (fraction)
'hsv_s': 0.678, # image HSV-Saturation augmentation (fraction)
'hsv_v': 0.36, # image HSV-Value augmentation (fraction)
'degrees': 1.113*0, # image rotation (+/- deg)
'translate': 0.06797*0, # image translation (+/- fraction)
'scale': 0.1059*0, # image scale (+/- gain)
'shear': 0.5768*0} # image shear (+/- deg)
'''
# Hyperparameters
hyp = {'giou': 3.54, # giou loss gain
'cls': 37.4, # cls loss gain
'cls_pw': 1.0, # cls BCELoss positive_weight
'obj': 64.3, # obj loss gain (*=img_size/320 if img_size != 320)
'obj_pw': 1.0, # obj BCELoss positive_weight
'iou_t': 0.20, # iou training threshold
'lr0': 0.01, # initial learning rate (SGD=5E-3, Adam=5E-4)
'lrf': 0.0005, # final learning rate (with cos scheduler)
'momentum': 0.98, # SGD momentum
'weight_decay': 0.0005, # optimizer weight decay
'fl_gamma': 0.0, # focal loss gamma (efficientDet default is gamma=1.5)
'hsv_h': 0.0138, # image HSV-Hue augmentation (fraction)
'hsv_s': 0.678, # image HSV-Saturation augmentation (fraction)
'hsv_v': 0.36, # image HSV-Value augmentation (fraction)
'degrees': 1.98 * 0, # image rotation (+/- deg)
'translate': 0.05 * 0, # image translation (+/- fraction)
'scale': 0.05 * 0, # image scale (+/- gain)
'shear': 0.641 * 0} # image shear (+/- deg)
'''
# Overwrite hyp with hyp*.txt (optional)
f = glob.glob('hyp*.txt')
if f:
print('Using %s' % f[0])
for k, v in zip(hyp.keys(), np.loadtxt(f[0])):
hyp[k] = v
# Print focal loss if gamma > 0
if hyp['fl_gamma']:
print('Using FocalLoss(gamma=%g)' % hyp['fl_gamma'])
def train(hyp):
cfg = opt.cfg
t_cfg = opt.t_cfg # teacher model cfg for knowledge distillation
data = opt.data
epochs = opt.epochs # 500200 batches at bs 64, 117263 images = 273 epochs
batch_size = opt.batch_size
accumulate = max(round(64 / batch_size), 1) # accumulate n times before optimizer update (bs 64)
if opt.quantized != 0:
weights = "weights/v4_tiny_416_map61.pt"
else:
weights = opt.weights # initial training weights
t_weights = opt.t_weights # teacher model weights
imgsz_min, imgsz_max, imgsz_test = opt.img_size # img sizes (min, max, test)
# Image Sizes
gs = 32 # (pixels) grid size
assert math.fmod(imgsz_min, gs) == 0, '--img-size %g must be a %g-multiple' % (imgsz_min, gs)
opt.multi_scale |= imgsz_min != imgsz_max # multi if different (min, max)
if opt.multi_scale:
if imgsz_min == imgsz_max:
imgsz_min //= 1.5
imgsz_max //= 0.667
grid_min, grid_max = imgsz_min // gs, imgsz_max // gs
imgsz_min, imgsz_max = int(grid_min * gs), int(grid_max * gs)
img_size = imgsz_max # initialize with max size
# Configure run
init_seeds()
# seed_torch() # modifiey on 2020 dec-31;
data_dict = parse_data_cfg(data)
train_path = data_dict['train']
test_path = data_dict['valid']
nc = 1 if opt.single_cls else int(data_dict['classes']) # number of classes
hyp['cls'] *= nc / 80 # update coco-tuned hyp['cls'] to current dataset
# Remove previous results
for f in glob.glob('*_batch*.jpg') + glob.glob(results_file):
os.remove(f)
# Initialize model
model = Darknet(cfg, quantized=opt.quantized, a_bit=opt.a_bit, w_bit=opt.w_bit, BN_Fold=opt.BN_Fold,
FPGA=opt.FPGA).to(device)
if t_cfg:
t_model = Darknet(t_cfg).to(device)
# Optimizer
pg0, pg1, pg2 = [], [], [] # optimizer parameter groups
for k, v in dict(model.named_parameters()).items():
if '.bias' in k:
pg2 += [v] # biases
elif 'Conv2d.weight' in k:
pg1 += [v] # apply weight_decay
else:
pg0 += [v] # all else
if opt.adam:
# hyp['lr0'] *= 0.1 # reduce lr (i.e. SGD=5E-3, Adam=5E-4)
optimizer = optim.Adam(pg0, lr=hyp['lr0'])
# optimizer = AdaBound(pg0, lr=hyp['lr0'], final_lr=0.1)
else:
optimizer = optim.SGD(pg0, lr=hyp['lr0'], momentum=hyp['momentum'], nesterov=True)
optimizer.add_param_group({'params': pg1, 'weight_decay': hyp['weight_decay']}) # add pg1 with weight_decay
optimizer.add_param_group({'params': pg2}) # add pg2 (biases)
print('Optimizer groups: %g .bias, %g Conv2d.weight, %g other' % (len(pg2), len(pg1), len(pg0)))
del pg0, pg1, pg2
# print('<.....................using gridmask.......................>')
# seed = int(img_size / 32)
# gridmask = GridMask(d1=96, d2=224, rotate=360, ratio=0.6, mode=1, prob=0.8)
print('<.....................using fencemask.......................>')
seed = int(img_size / 32)
fencemask = FenceMask(seed, seed * 3, seed * 4, seed * 8, [0, 0, 0], 0.8)
max_epoch = int(epochs * 0.8)
start_epoch = 0
best_fitness = 0.0
if weights != 'None':
attempt_download(weights)
if weights.endswith('.pt'): # pytorch format
# possible weights are '*.pt', 'yolov3-spp.pt', 'yolov3-tiny.pt' etc.
chkpt = torch.load(weights, map_location=device)
# load model
try:
chkpt['model'] = {k: v for k, v in chkpt['model'].items() if model.state_dict()[k].numel() == v.numel()}
model.load_state_dict(chkpt['model'], strict=False)
except KeyError as e:
s = "%s is not compatible with %s. Specify --weights '' or specify a --cfg compatible with %s. " \
"See https://github.com/ultralytics/yolov3/issues/657" % (opt.weights, opt.cfg, opt.weights)
raise KeyError(s) from e
# load optimizer
if chkpt['optimizer'] is not None:
optimizer.load_state_dict(chkpt['optimizer'])
best_fitness = chkpt['best_fitness']
# load results
if chkpt.get('training_results') is not None:
with open(results_file, 'w') as file:
file.write(chkpt['training_results']) # write results.txt
start_epoch = chkpt['epoch'] + 1
del chkpt
elif len(weights) > 0: # darknet format
# possible weights are '*.weights', 'yolov3-tiny.conv.15', 'darknet53.conv.74' etc.
load_darknet_weights(model, weights, pt=opt.pt, BN_Fold=opt.BN_Fold)
if t_cfg:
if t_weights.endswith('.pt'):
t_model.load_state_dict(torch.load(t_weights, map_location=device)['model'])
elif t_weights.endswith('.weights'):
load_darknet_weights(t_model, t_weights)
else:
raise Exception('pls provide proper teacher weights for knowledge distillation')
if not mixed_precision:
t_model.eval()
print('<.....................using knowledge distillation.......................>')
print('teacher model:', t_weights, '\n')
# Mixed precision training https://github.com/NVIDIA/apex
if mixed_precision:
model, optimizer = amp.initialize(model, optimizer, opt_level='O1', verbosity=0)
# Scheduler https://arxiv.org/pdf/1812.01187.pdf
lf = lambda x: (((1 + math.cos(x * math.pi / epochs)) / 2) ** 1.0) * 0.95 + 0.05 # cosine
scheduler = lr_scheduler.LambdaLR(optimizer, lr_lambda=lf)
scheduler.last_epoch = start_epoch - 1 # see link below
# https://discuss.pytorch.org/t/a-problem-occured-when-resuming-an-optimizer/28822
# Plot lr schedule
# y = []
# for _ in range(epochs):
# scheduler.step()
# y.append(optimizer.param_groups[0]['lr'])
# plt.plot(y, '.-', label='LambdaLR')
# plt.xlabel('epoch')
# plt.ylabel('LR')
# plt.tight_layout()
# plt.savefig('LR.png', dpi=300)
# Initialize distributed training
if device.type != 'cpu' and torch.cuda.device_count() > 1 and torch.distributed.is_available():
dist.init_process_group(backend='nccl', # 'distributed backend'
init_method='tcp://127.0.0.1:9999', # distributed training init method
world_size=1, # number of nodes for distributed training
rank=0) # distributed training node rank
model = torch.nn.parallel.DistributedDataParallel(model, find_unused_parameters=True)
model.yolo_layers = model.module.yolo_layers # move yolo layer indices to top level
# Dataset
dataset = LoadImagesAndLabels(train_path, img_size, batch_size,
augment=True,
hyp=hyp, # augmentation hyperparameters
rect=opt.rect, # rectangular training
cache_images=opt.cache_images,
single_cls=opt.single_cls)
testset = LoadImagesAndLabels(test_path, imgsz_test, batch_size // 4,
hyp=hyp,
rect=True,
cache_images=opt.cache_images,
single_cls=opt.single_cls,
#rank=opt.local_rank,
)
# 获得要剪枝的层
if hasattr(model, 'module'):
print('muti-gpus sparse')
if opt.prune == 0:
print('normal sparse training ')
_, _, prune_idx = parse_module_defs(model.module.module_defs)
elif opt.prune == 1:
print('shortcut sparse training')
_, _, prune_idx, _, _ = parse_module_defs2(model.module.module_defs)
elif opt.prune == 2:
print('layer sparse training')
_, _, prune_idx = parse_module_defs4(model.module.module_defs)
else:
print('single-gpu sparse')
if opt.prune == 0:
print('normal sparse training')
_, _, prune_idx = parse_module_defs(model.module_defs)
elif opt.prune == 1:
print('shortcut sparse training')
_, _, prune_idx, _, _ = parse_module_defs2(model.module_defs)
elif opt.prune == 2:
print('layer sparse training')
_, _, prune_idx = parse_module_defs4(model.module_defs)
# Dataloader
batch_size = min(batch_size, len(dataset))
nw = min([os.cpu_count(), batch_size if batch_size > 1 else 0, 8]) # number of workers
dataloader = torch.utils.data.DataLoader(dataset,
batch_size=batch_size,
num_workers=nw,
shuffle=not opt.rect, # Shuffle=True unless rectangular training is used
pin_memory=True,
collate_fn=dataset.collate_fn)
# Testloader
testloader = torch.utils.data.DataLoader(LoadImagesAndLabels(test_path, imgsz_test, batch_size,
hyp=hyp,
rect=True,
cache_images=opt.cache_images,
single_cls=opt.single_cls),
batch_size=batch_size,
num_workers=nw,
pin_memory=True,
collate_fn=dataset.collate_fn)
if opt.sr :
for idx in prune_idx:
if hasattr(model, 'module'):
bn_weights = gather_bn_weights(model.module.module_list, [idx])
else:
bn_weights = gather_bn_weights(model.module_list, [idx])
tb_writer.add_histogram('before_train_perlayer_bn_weights/hist', bn_weights.numpy(), idx, bins='doane')
# Model parameters
model.nc = nc # attach number of classes to model
model.hyp = hyp # attach hyperparameters to model
model.gr = 1.0 # giou loss ratio (obj_loss = 1.0 or giou)
model.class_weights = labels_to_class_weights(dataset.labels, nc).to(device) # attach class weights
# Model EMA
if opt.ema:
ema = torch_utils.ModelEMA(model)
# Start training
nb = len(dataloader) # number of batches
n_burn = max(3 * nb, 500) # burn-in iterations, max(3 epochs, 500 iterations)
maps = np.zeros(nc) # mAP per class
# torch.autograd.set_detect_anomaly(True)
results = (0, 0, 0, 0, 0, 0, 0) # 'P', 'R', 'mAP', 'F1', 'val GIoU', 'val Objectness', 'val Classification'
t0 = time.time()
print('Image sizes %g - %g train, %g test' % (imgsz_min, imgsz_max, imgsz_test))
print('Using %g dataloader workers' % nw)
print('Starting training for %g epochs...' % epochs)
for epoch in range(start_epoch, epochs): # epoch ------------------------------------------------------------------
fencemask.set_prob(epoch, max_epoch)
# gridmask.set_prob(epoch, max_epoch)
model.train()
print("learning rate lr: {:.6f}".format(optimizer.param_groups[0]['lr']))
# 稀疏化标志
sr_flag = get_sr_flag(epoch, opt.sr)
# Update image weights (optional)
if dataset.image_weights:
w = model.class_weights.cpu().numpy() * (1 - maps) ** 2 # class weights
image_weights = labels_to_image_weights(dataset.labels, nc=nc, class_weights=w)
dataset.indices = random.choices(range(dataset.n), weights=image_weights, k=dataset.n) # rand weighted idx
mloss = torch.zeros(4).to(device) # mean losses
print(('\n' + '%10s' * 8) % ('Epoch', 'gpu_mem', 'GIoU', 'obj', 'cls', 'total', 'targets', 'img_size'))
pbar = tqdm(enumerate(dataloader), total=nb) # progress bar
for i, (imgs, targets, paths, _) in pbar: # batch -------------------------------------------------------------
ni = i + nb * epoch # number integrated batches (since train start)
imgs = imgs.to(device).float() / 255.0 # uint8 to float32, 0 - 255 to 0.0 - 1.0
# Burn-in
if ni <= n_burn:
xi = [0, n_burn] # x interp
model.gr = np.interp(ni, xi, [0.0, 1.0]) # giou loss ratio (obj_loss = 1.0 or giou)
accumulate = max(1, np.interp(ni, xi, [1, 64 / batch_size]).round())
for j, x in enumerate(optimizer.param_groups):
# bias lr falls from 0.1 to lr0, all other lrs rise from 0.0 to lr0
x['lr'] = np.interp(ni, xi, [0.1 if j == 2 else 0.0, x['initial_lr'] * lf(epoch)])
x['weight_decay'] = np.interp(ni, xi, [0.0, hyp['weight_decay'] if j == 1 else 0.0])
if 'momentum' in x:
x['momentum'] = np.interp(ni, xi, [0.9, hyp['momentum']])
# Multi-Scale
if opt.multi_scale:
if ni / accumulate % 1 == 0: # adjust img_size (67% - 150%) every 1 batch
img_size = random.randrange(grid_min, grid_max + 1) * gs
sf = img_size / max(imgs.shape[2:]) # scale factor
if sf != 1:
ns = [math.ceil(x * sf / gs) * gs for x in imgs.shape[2:]] # new shape (stretched to 32-multiple)
imgs = F.interpolate(imgs, size=ns, mode='bilinear', align_corners=False)
# Forward
imgs = fencemask(imgs)
# imgs = gridmask(imgs)
targets = targets.to(device)
pred, feature_s = model(imgs)
# Loss
loss, loss_items = compute_loss(pred, targets, model)
if not torch.isfinite(loss):
print('WARNING: non-finite loss, ending training ', loss_items)
return results
attention_loss = 0
if t_cfg and opt.AT_str != -1:
if mixed_precision:
with torch.no_grad():
output_t, feature_t = t_model(imgs)
else:
_, output_t, feature_t = t_model(imgs)
if opt.AT_str == 1:
attention_loss = compute_lost_AT(model, targets, pred, output_t, feature_s, feature_t,
imgs.size(0))
elif opt.AT_str == 2:
attention_loss = compute_lost_group_AT(model, targets, pred, output_t, feature_s, feature_t,
imgs.size(0))
elif opt.AT_str == 3:
attention_loss = compute_lost_group_AT_KD(model, targets, pred, output_t, feature_s, feature_t,
imgs.size(0))
elif opt.AT_str == 4:
attention_loss = compute_lost_fine_grained_group_AT_KD(model, targets, pred, output_t, feature_s, feature_t,
batch_size,img_size)
else:
print("please select attention transfer strategy!")
loss = loss + attention_loss
if not torch.isfinite(loss):
print('WARNING: non-finite attention transfer loss, ending training ', )
return results
soft_target = 0
if t_cfg and opt.KDstr != -1:
if mixed_precision:
with torch.no_grad():
output_t, feature_t = t_model(imgs)
else:
_, output_t, feature_t = t_model(imgs)
if opt.KDstr == 1:
soft_target = compute_lost_KD(pred, output_t, model.nc, imgs.size(0))
elif opt.KDstr == 2:
soft_target, reg_ratio = compute_lost_KD2(model, targets, pred, output_t)
elif opt.KDstr == 3:
soft_target = compute_lost_KD3(model, targets, pred, output_t)
elif opt.KDstr == 4:
soft_target = compute_lost_KD4(model, targets, pred, output_t, feature_s, feature_t,
imgs.size(0))
elif opt.KDstr == 5:
soft_target = compute_lost_KD5(model, targets, pred, output_t, feature_s, feature_t,
imgs.size(0),
img_size)
elif opt.KDstr == 6:
soft_target = compute_lost_KD6(model, targets, pred, output_t, imgs.size(0))
else:
print("please select KD strategy!")
loss = loss + soft_target
# Backward
loss *= batch_size / 64 # scale loss
if mixed_precision:
with amp.scale_loss(loss, optimizer) as scaled_loss:
scaled_loss.backward()
else:
loss.backward()
# 对要剪枝层的γ参数稀疏化
if hasattr(model, 'module'):
if opt.prune != -1:
BNOptimizer.updateBN(sr_flag, model.module.module_list, opt.s, prune_idx)
else:
idx2mask = None
if opt.prune == 1 and epoch > opt.epochs * 0.4:
idx2mask = get_mask2(model, prune_idx, 0.70)
# BNOptimizer.updateBN(sr_flag, model.module_list, opt.s, prune_idx)
BNOptimizer.updateBN(sr_flag, model.module_list, opt.s, prune_idx, epoch, idx2mask, opt)
elif opt.prune == 1 and epoch <= opt.epochs * 0.4:
BNOptimizer.updateBN(sr_flag, model.module_list, opt.s, prune_idx, epoch, idx2mask, opt)
# Optimize
if ni % accumulate == 0:
optimizer.step()
optimizer.zero_grad()
if opt.ema:
ema.update(model)
# Print
mloss = (mloss * i + loss_items) / (i + 1) # update mean losses
mem = '%.3gG' % (torch.cuda.memory_reserved() / 1E9 if torch.cuda.is_available() else 0) # (GB)
s = ('%10s' * 2 + '%10.3g' * 6) % ('%g/%g' % (epoch, epochs - 1), mem, *mloss, len(targets), img_size)
pbar.set_description(s)
# Plot
if ni < 1:
f = 'train_batch%g.jpg' % i # filename
res = plot_images(images=imgs, targets=targets, paths=paths, fname=f)
if tb_writer:
tb_writer.add_image(f, res, dataformats='HWC', global_step=epoch)
# tb_writer.add_graph(model, imgs) # add model to tensorboard
# end batch ------------------------------------------------------------------------------------------------
# Update scheduler
scheduler.step()
# Process epoch results
if opt.ema:
ema.update_attr(model)
if hasattr(model, 'module'):
module_defs, module_list = ema.eam.module.module_defs, ema.eam.module.module_list
else:
module_defs, module_list = ema.eam.module_defs, ema.eam.module_list
for i, (mdef, module) in enumerate(zip(module_defs, module_list)):
if mdef['type'] == 'yolo':
yolo_layer = module
yolo_layer.nx, yolo_layer.ny = 0, 0
if hasattr(model, 'module'):
module_defs, module_list = model.module.module_defs, model.module.module_list
else:
module_defs, module_list = model.module_defs, model.module_list
for i, (mdef, module) in enumerate(zip(module_defs, module_list)):
if mdef['type'] == 'yolo':
yolo_layer = module
yolo_layer.nx, yolo_layer.ny = 0, 0
final_epoch = epoch + 1 == epochs
if not opt.notest or final_epoch: # Calculate mAP
is_coco = any([x in data for x in ['coco.data', 'coco2014.data', 'coco2017.data']]) and model.nc == 80
results, maps = test.test(cfg,
data,
batch_size=batch_size,
imgsz=imgsz_test,
model=ema.ema if opt.ema else model,
save_json=final_epoch and is_coco,
single_cls=opt.single_cls,
dataloader=testloader,
multi_label=ni > n_burn,
quantized=opt.quantized,
a_bit=opt.a_bit,
w_bit=opt.w_bit,
BN_Fold=opt.BN_Fold,
FPGA=opt.FPGA)
# Write
with open(results_file, 'a') as f:
f.write(s + '%10.3g' * 7 % results + '\n') # P, R, mAP, F1, test_losses=(GIoU, obj, cls)
if len(opt.name) and opt.bucket:
os.system('gsutil cp results.txt gs://%s/results/results%s.txt' % (opt.bucket, opt.name))
# Tensorboard
if tb_writer:
tags = ['train/giou_loss', 'train/obj_loss', 'train/cls_loss',
'metrics/precision', 'metrics/recall', 'metrics/mAP_0.5', 'metrics/F1',
'val/giou_loss', 'val/obj_loss', 'val/cls_loss']
for x, tag in zip(list(mloss[:-1]) + list(results), tags):
tb_writer.add_scalar(tag, x, epoch)
if opt.sr != -1:
for idx in prune_idx:
if hasattr(model, 'module'):
bn_weights = gather_bn_weights(model.module.module_list, [idx])
else:
# bn_weights = gather_bn_weights(model.module_list, [idx])
bn_weights = gather_bn_weights(model.module_list, [idx])
tb_writer.add_histogram('after sparse train bn_weights/hist', bn_weights.numpy(), epoch, bins='doane')
# Update best mAP
fi = fitness(np.array(results).reshape(1, -1)) # fitness_i = weighted combination of [P, R, mAP, F1]
if fi > best_fitness:
best_fitness = fi
# Save model
save = (not opt.nosave) or (final_epoch and not opt.evolve)
if opt.ema:
if hasattr(model, 'module'):
model_temp = ema.ema.module.state_dict()
else:
model_temp = ema.ema.state_dict()
else:
if hasattr(model, 'module'):
model_temp = model.module.state_dict()
else:
model_temp = model.state_dict()
if save:
with open(results_file, 'r') as f: # create checkpoint
chkpt = {'epoch': epoch,
'best_fitness': best_fitness,
'training_results': f.read(),
'model': model_temp,
'optimizer': None if final_epoch else optimizer.state_dict()}
# Save last, best and delete
torch.save(chkpt, last)
if (best_fitness == fi) and not final_epoch:
torch.save(chkpt, best)
del chkpt
# end epoch ----------------------------------------------------------------------------------------------------
# end training
n = opt.name
if len(n):
n = '_' + n if not n.isnumeric() else n
fresults, flast, fbest = 'results%s.txt' % n, wdir + 'last%s.pt' % n, wdir + 'best%s.pt' % n
for f1, f2 in zip([wdir + 'v4_tiny_416_map61.pt', wdir + 'best.pt', 'results.txt'], [flast, fbest, fresults]):
if os.path.exists(f1):
os.rename(f1, f2) # rename
ispt = f2.endswith('.pt') # is *.pt
strip_optimizer(f2) if ispt else None # strip optimizer
os.system('gsutil cp %s gs://%s/weights' % (f2, opt.bucket)) if opt.bucket and ispt else None # upload
if not opt.evolve:
plot_results() # save as results.png
print('%g epochs completed in %.3f hours.\n' % (epoch - start_epoch + 1, (time.time() - t0) / 3600))
dist.destroy_process_group() if torch.cuda.device_count() > 1 else None
torch.cuda.empty_cache()
return results
def WarmupForQ(hyp, step, a_bit, w_bit):
cfg = opt.cfg
data = opt.data
epochs = 5 + step * 5
batch_size = opt.batch_size
accumulate = max(round(64 / batch_size), 1) # accumulate n times before optimizer update (bs 64)
if step > 0:
weights = 'weights/v4_tiny_416_map61.pt'
else:
weights = opt.weights # initial training weights
imgsz_min, imgsz_max, imgsz_test = opt.img_size # img sizes (min, max, test)
# Image Sizes
gs = 32 # (pixels) grid size
assert math.fmod(imgsz_min, gs) == 0, '--img-size %g must be a %g-multiple' % (imgsz_min, gs)
opt.multi_scale |= imgsz_min != imgsz_max # multi if different (min, max)
if opt.multi_scale:
if imgsz_min == imgsz_max:
imgsz_min //= 1.5
imgsz_max //= 0.667
grid_min, grid_max = imgsz_min // gs, imgsz_max // gs
imgsz_min, imgsz_max = int(grid_min * gs), int(grid_max * gs)
img_size = imgsz_max # initialize with max size
# Configure run
init_seeds()
data_dict = parse_data_cfg(data)
train_path = data_dict['train']
test_path = data_dict['valid']
nc = 1 if opt.single_cls else int(data_dict['classes']) # number of classes
hyp['cls'] *= nc / 80 # update coco-tuned hyp['cls'] to current dataset
# Remove previous results
for f in glob.glob('*_batch*.jpg') + glob.glob(results_file):
os.remove(f)
# Initialize model
model = Darknet(cfg, quantized=opt.quantized, a_bit=a_bit, w_bit=w_bit, BN_Fold=opt.BN_Fold,
FPGA=opt.FPGA).to(device)
# Optimizer
pg0, pg1, pg2 = [], [], [] # optimizer parameter groups
for k, v in dict(model.named_parameters()).items():
if '.bias' in k:
pg2 += [v] # biases
elif 'Conv2d.weight' in k:
pg1 += [v] # apply weight_decay
else:
pg0 += [v] # all else
if opt.adam:
# hyp['lr0'] *= 0.1 # reduce lr (i.e. SGD=5E-3, Adam=5E-4)
optimizer = optim.Adam(pg0, lr=hyp['lr0'])
# optimizer = AdaBound(pg0, lr=hyp['lr0'], final_lr=0.1)
else:
optimizer = optim.SGD(pg0, lr=hyp['lr0'], momentum=hyp['momentum'], nesterov=True)
optimizer.add_param_group({'params': pg1, 'weight_decay': hyp['weight_decay']}) # add pg1 with weight_decay
optimizer.add_param_group({'params': pg2}) # add pg2 (biases)
print('Optimizer groups: %g .bias, %g Conv2d.weight, %g other' % (len(pg2), len(pg1), len(pg0)))
del pg0, pg1, pg2
start_epoch = 0
best_fitness = 0.0
if weights != 'None':
attempt_download(weights)
if weights.endswith('.pt'): # pytorch format
# possible weights are '*.pt', 'yolov3-spp.pt', 'yolov3-tiny.pt' etc.
chkpt = torch.load(weights, map_location=device)
# load model
try:
chkpt['model'] = {k: v for k, v in chkpt['model'].items() if model.state_dict()[k].numel() == v.numel()}
model.load_state_dict(chkpt['model'], strict=False)
except KeyError as e:
s = "%s is not compatible with %s. Specify --weights '' or specify a --cfg compatible with %s. " \
"See https://github.com/ultralytics/yolov3/issues/657" % (opt.weights, opt.cfg, opt.weights)
raise KeyError(s) from e
# load optimizer
if chkpt['optimizer'] is not None:
optimizer.load_state_dict(chkpt['optimizer'])
best_fitness = chkpt['best_fitness']
# load results
if chkpt.get('training_results') is not None:
with open(results_file, 'w') as file:
file.write(chkpt['training_results']) # write results.txt
start_epoch = chkpt['epoch'] + 1
del chkpt
elif len(weights) > 0: # darknet format
# possible weights are '*.weights', 'yolov3-tiny.conv.15', 'darknet53.conv.74' etc.
load_darknet_weights(model, weights, pt=opt.pt, BN_Fold=opt.BN_Fold)
# Mixed precision training https://github.com/NVIDIA/apex
if mixed_precision:
model, optimizer = amp.initialize(model, optimizer, opt_level='O1', verbosity=0)
# Scheduler https://arxiv.org/pdf/1812.01187.pdf
lf = lambda x: (((1 + math.cos(x * math.pi / epochs)) / 2) ** 1.0) * 0.95 + 0.05 # cosine
scheduler = lr_scheduler.LambdaLR(optimizer, lr_lambda=lf)
scheduler.last_epoch = start_epoch - 1 # see link below
# https://discuss.pytorch.org/t/a-problem-occured-when-resuming-an-optimizer/28822
# Plot lr schedule
# y = []
# for _ in range(epochs):
# scheduler.step()
# y.append(optimizer.param_groups[0]['lr'])
# plt.plot(y, '.-', label='LambdaLR')
# plt.xlabel('epoch')
# plt.ylabel('LR')
# plt.tight_layout()
# plt.savefig('LR.png', dpi=300)
# Initialize distributed training
if device.type != 'cpu' and torch.cuda.device_count() > 1 and torch.distributed.is_available():
dist.init_process_group(backend='nccl', # 'distributed backend'
init_method='tcp://127.0.0.1:9999', # distributed training init method
world_size=1, # number of nodes for distributed training
rank=0) # distributed training node rank
model = torch.nn.parallel.DistributedDataParallel(model, find_unused_parameters=True)
model.yolo_layers = model.module.yolo_layers # move yolo layer indices to top level
# Dataset
dataset = LoadImagesAndLabels(train_path, img_size, batch_size,
augment=True,
hyp=hyp, # augmentation hyperparameters
rect=opt.rect, # rectangular training
cache_images=opt.cache_images,
single_cls=opt.single_cls)
# Dataloader
batch_size = min(batch_size, len(dataset))
nw = min([os.cpu_count(), batch_size if batch_size > 1 else 0, 8]) # number of workers
dataloader = torch.utils.data.DataLoader(dataset,
batch_size=batch_size,
num_workers=nw,
shuffle=not opt.rect, # Shuffle=True unless rectangular training is used
pin_memory=True,
collate_fn=dataset.collate_fn)
# Testloader
testloader = torch.utils.data.DataLoader(LoadImagesAndLabels(test_path, imgsz_test, batch_size,
hyp=hyp,
rect=True,
cache_images=opt.cache_images,
single_cls=opt.single_cls),
batch_size=batch_size,
num_workers=nw,
pin_memory=True,
collate_fn=dataset.collate_fn)
# Model parameters
model.nc = nc # attach number of classes to model
model.hyp = hyp # attach hyperparameters to model
model.gr = 1.0 # giou loss ratio (obj_loss = 1.0 or giou)
model.class_weights = labels_to_class_weights(dataset.labels, nc).to(device) # attach class weights
# Model EMA
if opt.ema:
ema = torch_utils.ModelEMA(model)
# Start training
nb = len(dataloader) # number of batches
n_burn = max(3 * nb, 500) # burn-in iterations, max(3 epochs, 500 iterations)
maps = np.zeros(nc) # mAP per class
# torch.autograd.set_detect_anomaly(True)
results = (0, 0, 0, 0, 0, 0, 0) # 'P', 'R', 'mAP', 'F1', 'val GIoU', 'val Objectness', 'val Classification'
t0 = time.time()
print('Image sizes %g - %g train, %g test' % (imgsz_min, imgsz_max, imgsz_test))
print('Using %g dataloader workers' % nw)
print('Starting training for %g epochs...' % epochs)
for epoch in range(start_epoch, epochs): # epoch ------------------------------------------------------------------
model.train()
# Update image weights (optional)
if dataset.image_weights:
w = model.class_weights.cpu().numpy() * (1 - maps) ** 2 # class weights
image_weights = labels_to_image_weights(dataset.labels, nc=nc, class_weights=w)
dataset.indices = random.choices(range(dataset.n), weights=image_weights, k=dataset.n) # rand weighted idx
mloss = torch.zeros(4).to(device) # mean losses
print(('\n' + '%10s' * 8) % ('Epoch', 'gpu_mem', 'GIoU', 'obj', 'cls', 'total', 'targets', 'img_size'))
pbar = tqdm(enumerate(dataloader), total=nb) # progress bar
for i, (imgs, targets, paths, _) in pbar: # batch -------------------------------------------------------------
ni = i + nb * epoch # number integrated batches (since train start)
imgs = imgs.to(device).float() / 255.0 # uint8 to float32, 0 - 255 to 0.0 - 1.0
targets = targets.to(device)
# Burn-in
if ni <= n_burn:
xi = [0, n_burn] # x interp
model.gr = np.interp(ni, xi, [0.0, 1.0]) # giou loss ratio (obj_loss = 1.0 or giou)
accumulate = max(1, np.interp(ni, xi, [1, 64 / batch_size]).round())
for j, x in enumerate(optimizer.param_groups):
# bias lr falls from 0.1 to lr0, all other lrs rise from 0.0 to lr0
x['lr'] = np.interp(ni, xi, [0.1 if j == 2 else 0.0, x['initial_lr'] * lf(epoch)])
x['weight_decay'] = np.interp(ni, xi, [0.0, hyp['weight_decay'] if j == 1 else 0.0])
if 'momentum' in x:
x['momentum'] = np.interp(ni, xi, [0.9, hyp['momentum']])
# Multi-Scale
if opt.multi_scale:
if ni / accumulate % 1 == 0: # adjust img_size (67% - 150%) every 1 batch
img_size = random.randrange(grid_min, grid_max + 1) * gs
sf = img_size / max(imgs.shape[2:]) # scale factor
if sf != 1:
ns = [math.ceil(x * sf / gs) * gs for x in imgs.shape[2:]] # new shape (stretched to 32-multiple)
imgs = F.interpolate(imgs, size=ns, mode='bilinear', align_corners=False)
# Forward
pred, _ = model(imgs)
# Loss
loss, loss_items = compute_loss(pred, targets, model)
if not torch.isfinite(loss):
print('WARNING: non-finite loss, ending training ', loss_items)
return results
# Backward
loss *= batch_size / 64 # scale loss
if mixed_precision:
with amp.scale_loss(loss, optimizer) as scaled_loss:
scaled_loss.backward()
else:
loss.backward()
# Optimize
if ni % accumulate == 0:
optimizer.step()
optimizer.zero_grad()
if opt.ema:
ema.update(model)
# Print
mloss = (mloss * i + loss_items) / (i + 1) # update mean losses
mem = '%.3gG' % (torch.cuda.memory_reserved() / 1E9 if torch.cuda.is_available() else 0) # (GB)
s = ('%10s' * 2 + '%10.3g' * 6) % ('%g/%g' % (epoch, epochs - 1), mem, *mloss, len(targets), img_size)
pbar.set_description(s)
# Plot
if ni < 1:
f = 'train_batch%g.jpg' % i # filename
res = plot_images(images=imgs, targets=targets, paths=paths, fname=f)
if tb_writer:
tb_writer.add_image(f, res, dataformats='HWC', global_step=epoch)
# tb_writer.add_graph(model, imgs) # add model to tensorboard
# end batch ------------------------------------------------------------------------------------------------
# Update scheduler
scheduler.step()
# Process epoch results
if opt.ema:
ema.update_attr(model)
if hasattr(model, 'module'):
module_defs, module_list = ema.eam.module.module_defs, ema.eam.module.module_list
else:
module_defs, module_list = ema.eam.module_defs, ema.eam.module_list
for i, (mdef, module) in enumerate(zip(module_defs, module_list)):
if mdef['type'] == 'yolo':
yolo_layer = module
yolo_layer.nx, yolo_layer.ny = 0, 0
if hasattr(model, 'module'):
module_defs, module_list = model.module.module_defs, model.module.module_list
else:
module_defs, module_list = model.module_defs, model.module_list
for i, (mdef, module) in enumerate(zip(module_defs, module_list)):
if mdef['type'] == 'yolo':
yolo_layer = module
yolo_layer.nx, yolo_layer.ny = 0, 0
final_epoch = epoch + 1 == epochs
if not opt.notest or final_epoch: # Calculate mAP
is_coco = any([x in data for x in ['coco.data', 'coco2014.data', 'coco2017.data']]) and model.nc == 80
results, maps = test.test(cfg,
data,
batch_size=batch_size,
imgsz=imgsz_test,
model=ema.ema if opt.ema else model,
save_json=final_epoch and is_coco,
single_cls=opt.single_cls,
dataloader=testloader,
multi_label=ni > n_burn,
quantized=opt.quantized,
a_bit=opt.a_bit,
w_bit=opt.w_bit,
BN_Fold=opt.BN_Fold,
FPGA=opt.FPGA)
# Write
with open(results_file, 'a') as f:
f.write(s + '%10.3g' * 7 % results + '\n') # P, R, mAP, F1, test_losses=(GIoU, obj, cls)
if len(opt.name) and opt.bucket:
os.system('gsutil cp results.txt gs://%s/results/results%s.txt' % (opt.bucket, opt.name))
# Tensorboard
if tb_writer:
tags = ['train/giou_loss', 'train/obj_loss', 'train/cls_loss',
'metrics/precision', 'metrics/recall', 'metrics/mAP_0.5', 'metrics/F1',
'val/giou_loss', 'val/obj_loss', 'val/cls_loss']
for x, tag in zip(list(mloss[:-1]) + list(results), tags):
tb_writer.add_scalar(tag, x, epoch)
if opt.sr:
for idx in prune_idx:
if hasattr(model, 'module'):
bn_weights = gather_bn_weights(model.module.module_list, [idx])
else:
bn_weights = gather_bn_weights(model.module_list, [idx])
tb_writer.add_histogram('bn_weights/hist', bn_weights.numpy(), epoch, bins='doane')
# Update best mAP
fi = fitness(np.array(results).reshape(1, -1)) # fitness_i = weighted combination of [P, R, mAP, F1]
if fi > best_fitness:
best_fitness = fi
# Save model
save = (not opt.nosave) or (final_epoch and not opt.evolve)
if opt.ema:
if hasattr(model, 'module'):
model_temp = ema.ema.module.state_dict()
else:
model_temp = ema.ema.state_dict()
else:
if hasattr(model, 'module'):
model_temp = model.module.state_dict()
else:
model_temp = model.state_dict()
if save:
with open(results_file, 'r') as f: # create checkpoint
chkpt = {'epoch': epoch,
'best_fitness': best_fitness,
'training_results': f.read(),
'model': model_temp,
'optimizer': None if final_epoch else optimizer.state_dict()}
# Save last, best and delete
torch.save(chkpt, last)
if (best_fitness == fi) and not final_epoch:
torch.save(chkpt, best)
del chkpt
# end epoch ----------------------------------------------------------------------------------------------------
# end training
n = opt.name
if len(n):
n = '_' + n if not n.isnumeric() else n
fresults, flast, fbest = 'results%s.txt' % n, wdir + 'last%s.pt' % n, wdir + 'best%s.pt' % n
for f1, f2 in zip([wdir + 'v4_tiny_416_map61.pt', wdir + 'best.pt', 'results.txt'], [flast, fbest, fresults]):
if os.path.exists(f1):
os.rename(f1, f2) # rename
ispt = f2.endswith('.pt') # is *.pt
strip_optimizer(f2) if ispt else None # strip optimizer
os.system('gsutil cp %s gs://%s/weights' % (f2, opt.bucket)) if opt.bucket and ispt else None # upload
if not opt.evolve:
plot_results() # save as results.png
print('%g epochs completed in %.3f hours.\n' % (epoch - start_epoch + 1, (time.time() - t0) / 3600))
dist.destroy_process_group() if torch.cuda.device_count() > 1 else None
torch.cuda.empty_cache()
return results
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('--epochs', type=int, default=300) # 500200 batches at bs 16, 117263 COCO images = 273 epochs
parser.add_argument('--batch-size', type=int, default=16) # effective bs = batch_size * accumulate = 16 * 4 = 64
parser.add_argument('--cfg', type=str, default='cfg/yolov3-spp.cfg', help='*.cfg path')
parser.add_argument('--t_cfg', type=str, default='', help='teacher model cfg file path for knowledge distillation')
parser.add_argument('--data', type=str, default='data/coco2017.data', help='*.data path')
parser.add_argument('--multi-scale', action='store_true', help='adjust (67%% - 150%%) img_size every 10 batches')
parser.add_argument('--img-size', nargs='+', type=int, default=[320, 640], help='[min_train, max-train, test]')
parser.add_argument('--rect', action='store_true', help='rectangular training')
parser.add_argument('--resume', action='store_true', help='resume training from v4_tiny_416_map61.pt')
parser.add_argument('--nosave', action='store_true', help='only save final checkpoint')
parser.add_argument('--notest', action='store_true', help='only test final epoch')
parser.add_argument('--evolve', action='store_true', help='evolve hyperparameters')
parser.add_argument('--bucket', type=str, default='', help='gsutil bucket')
parser.add_argument('--cache-images', action='store_true', help='cache images for faster training')
parser.add_argument('--weights', type=str, default='weights/yolov3-spp-ultralytics.pt', help='initial weights path')
parser.add_argument('--t_weights', type=str, default='', help='teacher model weights')
parser.add_argument('--AT_str', type=int, default=-1, help='Attention tranfer strategy')
parser.add_argument('--KDstr', type=int, default=-1, help='KD strategy')
parser.add_argument('--name', default='', help='renames results.txt to results_name.txt if supplied')
parser.add_argument('--device', default='', help='device id (i.e. 0 or 0,1 or cpu)')
parser.add_argument('--adam', action='store_true', help='use adam optimizer')
parser.add_argument('--ema', action='store_true', help='use ema')
parser.add_argument('--single-cls', action='store_true', help='train as single-class dataset')
parser.add_argument('--sparsity-regularization', '-sr', dest='sr', action='store_true',
help='train with channel sparsity regularization')
parser.add_argument('--pretrain', '-pt', dest='pt', action='store_true',
help='use pretrain model')