-
Notifications
You must be signed in to change notification settings - Fork 0
/
train_val_3d.py
390 lines (329 loc) · 15.7 KB
/
train_val_3d.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
import argparse
import os
import time
import json
import numpy as np
import yaml
import torch
import torch.backends.cudnn as cudnn
import torch.distributed as dist
import torch.nn.functional as F
from torch.utils.tensorboard import SummaryWriter
from torch.nn.parallel import DistributedDataParallel
from torchvision import transforms
from util import clip_transforms
from util.clip_augmentations import ClipRandAugment
from util.util import ClipGaussianBlur, AverageMeter, merge_scores, accuracy, reduce_tensor
from util.lr_scheduler import get_scheduler
from util.logger import setup_logger
from layer.LSR import *
from dataset.video_merge_dataset import VideoMergeDataset
from dataset.video_dataset import VideoRGBTrainDataset, VideoRGBTestDataset
import model as model_factory
from layer.pooling_factory import get_pooling_by_name
from torch.cuda.amp import GradScaler
def add_config(args, name, config):
if isinstance(config, dict):
for key in config.keys():
add_config(args, key, config[key])
else:
setattr(args, name, config)
def merge_config(conf1, conf2):
if isinstance(conf1, dict) and isinstance(conf2, dict):
new_config = {}
key_list = list(set(conf1.keys()).union(set(conf2.keys())))
for key in key_list:
if (key in conf1) and (key in conf2): # union of c1 & c2
new_config[key] = merge_config(conf1.get(key), conf2.get(key))
else:
new_config[key] = conf1.get(key) if key in conf1 else conf2.get(key)
return new_config
else:
return conf1 if conf2 is None else conf2
def parse_option():
parser = argparse.ArgumentParser('training')
parser.add_argument('--config_file', type=str, required=True, help='path of config file (yaml)')
parser.add_argument('--local_rank', type=int, help='local rank for DistributedDataParallel')
args = parser.parse_args()
# load config file, default + base + exp
config_default = yaml.load(open('./base_config/default.yml', 'r'))
config_exp = yaml.load(open(args.config_file, 'r'))
if 'base' in config_exp:
config_base = yaml.load(open(config_exp['base'], 'r'))
else:
config_base = None
config = merge_config(merge_config(config_default, config_base), config_exp)
args.C = config
add_config(args, 'root', config)
return args
def get_loader(args):
if args.rand_augment:
train_transform = transforms.Compose([
clip_transforms.ClipRandomResizedCrop(args.crop_size, scale=(0.2, 1.), ratio=(0.75, 1.3333333333333333)),
ClipRandAugment(n=args.ra_n, m=args.ra_m), # N = [1, 2, 3], M = [5, 7, 9, 11, 13, 15]
clip_transforms.ClipRandomHorizontalFlip(p=0.0 if args.no_horizontal_flip else 0.5),
clip_transforms.ToClipTensor(),
clip_transforms.ClipNormalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
transforms.Lambda(lambda clip: torch.stack(clip, dim=1)) if args.time_dim == "T" else transforms.Lambda(lambda clip: torch.cat(clip, dim=0))
])
else:
train_transform = transforms.Compose([
clip_transforms.ClipRandomResizedCrop(args.crop_size, scale=(0.2, 1.), ratio=(0.75, 1.3333333333333333)),
transforms.RandomApply([
clip_transforms.ClipColorJitter(0.4, 0.4, 0.4, 0.1) # not strengthened
], p=0.8),
clip_transforms.ClipRandomGrayscale(p=0.2),
transforms.RandomApply([ClipGaussianBlur([.1, 2.])], p=0.5),
clip_transforms.ClipRandomHorizontalFlip(p=0.0 if args.no_horizontal_flip else 0.5),
clip_transforms.ToClipTensor(),
clip_transforms.ClipNormalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
transforms.Lambda(lambda clip: torch.stack(clip, dim=1)) if args.time_dim == "T" else transforms.Lambda(lambda clip: torch.cat(clip, dim=0))
])
if args.dataset_class == 'video_dataset':
assert (args.list_file != '' and args.root_path != '')
train_dataset = VideoRGBTrainDataset(list_file=args.list_file, root_path=args.root_path,
transform=train_transform, clip_length=args.clip_length,
num_steps=args.num_steps, num_segments=args.num_segments,
format=args.format)
else:
assert (args.lmdb_path != '' and args.video_num != 0 and args.repeat_num != 0)
train_dataset = VideoMergeDataset(args.lmdb_path + '_' + str(args.local_rank), video_num=args.video_num,
repeat_num=args.repeat_num, transform=train_transform,
clip_length=args.clip_length, num_steps=args.num_steps,
num_segments=args.num_segments)
train_sampler = torch.utils.data.distributed.DistributedSampler(train_dataset)
train_loader = torch.utils.data.DataLoader(
train_dataset, batch_size=args.batch_size, shuffle=False,
num_workers=args.num_workers, pin_memory=True,
sampler=train_sampler, drop_last=True)
return train_loader
def get_val_loader(args):
# only crop the center clip for evaluation
crop = clip_transforms.ClipCenterCrop
test_transform = transforms.Compose([
clip_transforms.ClipResize(size=args.crop_size),
crop(size=args.crop_size),
clip_transforms.ToClipTensor(),
clip_transforms.ClipNormalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
transforms.Lambda(lambda clip: torch.stack(clip, dim=1)) if args.time_dim == "T" else transforms.Lambda(lambda clip: torch.cat(clip, dim=0))
])
test_dataset = VideoRGBTestDataset(args.eva_list_file, num_clips=args.num_clips, transform=test_transform, root_path=args.root_path, \
clip_length=args.clip_length, num_steps=args.num_steps, num_segments=args.num_segments, \
format=args.format)
#test_sampler = torch.utils.data.distributed.DistributedSampler(test_dataset, shuffle=False)
test_loader = torch.utils.data.DataLoader(
test_dataset, batch_size=args.batch_size, shuffle=False,
num_workers=args.num_workers, pin_memory=True, drop_last=False)
#sampler=test_sampler, )
return test_loader
def build_model(args):
model = model_factory.get_model_by_name(net_name=args.net_name, pooling_arch=get_pooling_by_name(args.pooling_name),
num_classes=args.num_classes, dropout_ratio=args.dropout_ratio,
clip_length=(args.num_segments*args.clip_length), sifa_kernel=args.sifa_kernel).cuda()
if args.pretrained_model:
load_pretrained(args, model)
return model
def load_pretrained(args, model):
ckpt = torch.load(args.pretrained_model, map_location='cpu')
if 'model' in ckpt:
state_dict = {k.replace("module.", ""): v for k, v in ckpt['model'].items()}
else:
state_dict = ckpt
# convert initial weights
if args.transfer_weights:
state_dict = model_factory.transfer_weights(args.net_name, state_dict)
if args.remove_fc:
state_dict = model_factory.remove_fc(args.net_name, state_dict)
if args.remove_defcor_weight:
state_dict = model_factory.remove_defcor_weight(args.net_name, state_dict)
[misskeys, unexpkeys] = model.load_state_dict(state_dict, strict=False)
logger.info('Missing keys: {}'.format(misskeys))
logger.info('Unexpect keys: {}'.format(unexpkeys))
logger.info("==> loaded checkpoint '{}'".format(args.pretrained_model))
def save_checkpoint(args, epoch, model, optimizer, scheduler):
logger.info('==> Saving...')
state = {
'opt': args,
'model': model.state_dict(),
'optimizer': optimizer.state_dict(),
'scheduler': scheduler.state_dict(),
'epoch': epoch,
}
torch.save(state, os.path.join(args.output_dir, 'current.pth'))
if epoch % args.save_freq == 0:
torch.save(state, os.path.join(args.output_dir, 'ckpt_epoch_{}.pth'.format(epoch)))
def main(args):
train_loader = get_loader(args)
val_loader = get_val_loader(args)
n_data = len(train_loader.dataset)
logger.info("length of training dataset: {}".format(n_data))
model = build_model(args)
if args.pretrained_model:
ckpt = torch.load(args.pretrained_model, map_location='cpu')
# print network architecture
if dist.get_rank() == 0:
logger.info(model)
if args.label_smooth:
criterion = LSR(e=0.1).cuda()
else:
criterion = torch.nn.CrossEntropyLoss().cuda()
# optimizer
optimizer = torch.optim.SGD(filter(lambda p: p.requires_grad, model.parameters()),
lr=args.base_learning_rate,
momentum=args.momentum,
weight_decay=args.weight_decay)
if args.resume:
if args.reverse:
optimizer.load_state_dict(ckpt['scheduler'])
else: optimizer.load_state_dict(ckpt['optimizer'])
# scheduler
scheduler = get_scheduler(optimizer, len(train_loader), args)
if args.resume:
if args.reverse:
scheduler.load_state_dict(ckpt['optimizer'])
else: scheduler.load_state_dict(ckpt['scheduler'])
model = DistributedDataParallel(model, device_ids=[args.local_rank], broadcast_buffers=True,
find_unused_parameters=True)
#model, optimizer = amp.initialize(model, optimizer, opt_level='O1')
#model = DDP(model)
# tensorboard
if dist.get_rank() == 0:
summary_writer = SummaryWriter(log_dir=args.output_dir)
else:
summary_writer = None
# routine
start_epoch = 1
if args.resume: start_epoch = ckpt['epoch']
for epoch in range(start_epoch, args.epochs + 1):
train_loader.sampler.set_epoch(epoch)
tic = time.time()
loss = train(epoch, train_loader, model, criterion, optimizer, scheduler, args)
logger.info('epoch {}, total time {:.2f}'.format(epoch, time.time() - tic))
if summary_writer is not None:
# tensorboard logger
summary_writer.add_scalar('ins_loss', loss, epoch)
summary_writer.add_scalar('learning_rate', optimizer.param_groups[0]['lr'], epoch)
if dist.get_rank() == 0:
# save model
save_checkpoint(args, epoch, model, optimizer, scheduler)
if epoch % args.eva_inter_freq == 0 or epoch == args.epochs:
# evaluation on test data
tic_val = time.time()
eva_accuracy = eval(epoch, val_loader, model, args)
top1_accuracy = eva_accuracy[0].cuda()
top3_accuracy = eva_accuracy[1].cuda()
top5_accuracy = eva_accuracy[2].cuda()
t1 = top1_accuracy.data.cpu().item()
t3 = top3_accuracy.data.cpu().item()
t5 = top5_accuracy.data.cpu().item()
logger.info('val top1 accuracy {:.4f}, top3 accuracy: {:.4f}, top5: {:.4f} val time {:.2f}'.format(t1, t3, t5, time.time() - tic_val))
def frozen_bn(model):
first_bn = True
for name, m in model.named_modules():
if isinstance(m, (torch.nn.BatchNorm2d, torch.nn.BatchNorm3d)):
if first_bn:
first_bn = False
print('Skip frozen first bn layer: ' + name)
continue
m.eval()
m.weight.requires_grad = False
m.bias.requires_grad = False
def train(epoch, train_loader, model, criterion, optimizer, scheduler, args):
model.train()
if args.frozen_bn:
frozen_bn(model)
batch_time = AverageMeter()
loss_meter = AverageMeter()
norm_meter = AverageMeter()
end = time.time()
optimizer.zero_grad()
scaler = GradScaler()
bnorm = 0
for idx, (x, label) in enumerate(train_loader):
bsz = x.size(0)
# forward
x = x.cuda(non_blocking=True) # clip
label = label.cuda(non_blocking=True) # label
# with torch.cuda.amp.autocast():
# forward and get the predict score
score = model(x)
# get crossentropy loss
if isinstance(score, list):
loss = criterion(score[0], label) + criterion(score[1], label)
else:
loss = criterion(score, label)
# backward
scaler.scale(loss / args.iter_size * args.loss_weight).backward()
#with amp.scale_loss(loss, optimizer) as scaled_loss:
# scaled_loss.backward()
if (idx + 1) % args.iter_size == 0:
scaler.unscale_(optimizer)
bnorm = torch.nn.utils.clip_grad_norm_(filter(lambda p: p.requires_grad, model.parameters()),
args.clip_gradient)
scaler.step(optimizer)
scaler.update()
optimizer.zero_grad()
scheduler.step()
# update meters
loss_meter.update(loss.item(), bsz)
norm_meter.update(bnorm, bsz)
batch_time.update(time.time() - end)
end = time.time()
# print info
if idx % args.print_freq == 0:
lr = scheduler.get_lr()[0]
logger.info(
'Train: [{:>3d}]/[{:>4d}/{:>4d}] BT={:>0.3f}/{:>0.3f} Loss={:>0.3f}/{:>0.3f} GradNorm={:>0.3f}/{:>0.3f} Lr={:>0.3f}'.format(
epoch, idx, len(train_loader),
batch_time.val, batch_time.avg,
loss.item(), loss_meter.avg,
bnorm, norm_meter.avg,lr
))
return loss_meter.avg
def eval(epoch, val_loader, model, args):
model.eval()
softmax = torch.nn.Softmax(dim=1)
all_scores = np.zeros([len(val_loader) * args.batch_size, args.num_classes], dtype=np.float)
all_labels = np.zeros([len(val_loader) * args.batch_size], dtype=np.float)
top_idx = 0
with torch.no_grad():
logger.info('==> Validating... num clips {} val video num {}'.format(args.num_clips,args.val_video_num))
for idx, (x, label) in enumerate(val_loader):
if idx % 100 == 0:
logger.info('{}/{}'.format(idx, len(val_loader)))
bsz = x.size(0)
score = model(x)
#score = softmax(score)
if isinstance(score, list):
score_numpy = (softmax(score[0]).data.cpu().numpy() + softmax(score[1]).data.cpu().numpy()) / 2
else:
score_numpy = softmax(score).data.cpu().numpy()
label_numpy = label.data.cpu().numpy()
all_scores[top_idx: top_idx + bsz, :] = score_numpy
all_labels[top_idx: top_idx + bsz] = label_numpy
top_idx += bsz
all_scores = all_scores[:top_idx, :]
# pooling the scores for each video
v_all_scores, v_all_labels = merge_scores(all_scores, all_labels, args)
# compute the accuracy
acc = accuracy(v_all_scores, v_all_labels, topk=(1, 3, 5))
return acc
if __name__ == '__main__':
opt = parse_option()
if 'RANK' in os.environ and 'WORLD_SIZE' in os.environ:
opt.rank = int(os.environ["RANK"])
opt.world_size = int(os.environ['WORLD_SIZE'])
opt.local_rank = int(os.environ['LOCAL_RANK'])
torch.cuda.set_device(opt.local_rank)
else:
torch.cuda.set_device(opt.local_rank)
torch.distributed.init_process_group(backend='nccl', init_method='env://')
cudnn.benchmark = True
os.makedirs(opt.output_dir, exist_ok=True)
logger = setup_logger(output=opt.output_dir, distributed_rank=dist.get_rank(), name="sifa")
if dist.get_rank() == 0:
path = os.path.join(opt.output_dir, "train_val_3d.config.json")
with open(path, 'w') as f:
json.dump(vars(opt), f, indent=2)
logger.info("Full config saved to {}".format(path))
main(opt)