-
Notifications
You must be signed in to change notification settings - Fork 1
/
train.py
239 lines (205 loc) · 7.8 KB
/
train.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
import sys
import time
from torchvision import transforms
from datasets.augmentation import *
from evaluate import _evaluate_casia_b, evaluate
from losses import SupConLoss
from utils import AverageMeter
from common import *
from models.SpatialTransformerTemporalConv import SpatialTransformerTemporalConv, SpatioTemporalTransformer
from logs import UnifyLog
from pytorch_metric_learning import miners, losses
from sampler import BalancedBatchSampler
from tqdm import tqdm
from datasets.gait import (
CasiaBPose,
CasiaQueryDataset
)
miner = miners.MultiSimilarityMiner()
def train(train_loader, model, loss_func, optimizer, scheduler, scaler, epoch, opt):
"""one epoch training"""
model.train()
batch_time = AverageMeter()
data_time = AverageMeter()
losses = AverageMeter()
end = time.time()
for idx, (points, target) in enumerate(train_loader):
data_time.update(time.time() - end)
if opt.loss_func == 'supcon':
bsz = points[0].shape[0]
points = torch.cat([points[0], points[1]], dim=0)
else:
bsz = points.shape[0]
labels = target[0]
if torch.cuda.is_available():
points = points.cuda(non_blocking=True)
labels = labels.cuda(non_blocking=True)
features = model(points)
if opt.loss_func == 'supcon':
f1, f2 = torch.split(features, [bsz, bsz], dim=0)
features = torch.cat([f1.unsqueeze(1), f2.unsqueeze(1)], dim=1)
loss = loss_func(features, labels)
else:
hard_pairs = miner(features, labels)
loss = loss_func(features, labels, hard_pairs)
# update metric
losses.update(loss.item(), bsz)
loss.backward()
optimizer.step()
optimizer.zero_grad()
scheduler.step()
############################################################################
# measure elapsed time
batch_time.update(time.time() - end)
end = time.time()
# print info
if (idx + 1) % 10 == 0:
print(
f"Train: [{epoch}][{idx + 1}/{len(train_loader)}]\t"
f"BT {batch_time.val:.3f} ({batch_time.avg:.3f})\t"
f"DT {data_time.val:.3f} ({data_time.avg:.3f})\t"
f"loss {losses.val:.3f} ({losses.avg:.3f})"
)
sys.stdout.flush()
return losses.avg
def main(opt):
opt = setup_environment(opt)
# Dataset
transform = transforms.Compose(
[
MirrorPoses(opt.mirror_probability),
FlipSequence(opt.flip_probability),
RandomSelectSequence(opt.sequence_length),
PointNoise(std=opt.point_noise_std),
JointNoise(std=opt.joint_noise_std),
joint_drop if opt.joint_drop == 'single' else lambda x:x,
remove_conf(enable=opt.rm_conf),
normalize_width,
ToTensor()
],
)
if opt.loss_func == 'supcon':
transform = TwoNoiseTransform(transform)
val_transform = transforms.Compose(
[
SelectSequenceCenter(opt.sequence_length),
remove_conf(enable=opt.rm_conf),
normalize_width,
ToTensor()
]
)
if opt.dataset == "casia-b":
dataset = CasiaBPose(
opt.train_data_path,
sequence_length=opt.sequence_length,
duplicate_bgcl=True,
transform=transform,
)
dataset_valid = CasiaBPose(
opt.valid_data_path,
sequence_length=opt.sequence_length,
transform=ThreeCenterSequenceTransform(
transform=val_transform, sequence_length=opt.sequence_length),
)
elif opt.dataset == "casia-b-query":
dataset = CasiaQueryDataset(
opt.train_data_path,
id_range=opt.train_id,
duplicate_bgcl=True,
transform=transform
)
dataset_valid = CasiaQueryDataset(
opt.train_data_path,
id_range=opt.test_id,
transform=ThreeCenterSequenceTransform(
transform=val_transform, sequence_length=opt.sequence_length),
)
if opt.balance_sampler:
if opt.train_id is None:
num_class = 74
else:
num_class = int(opt.train_id.split('-')[1])
print('---- balance sampler ----- class=', num_class)
labels = []
for i in dataset.targets:
labels.append(i[0])
labels = torch.tensor(labels)
_sampler = BalancedBatchSampler(
labels=labels, n_classes=num_class, n_samples=opt.sampler_num_sample)
train_loader = torch.utils.data.DataLoader(
dataset,
num_workers=opt.num_workers,
pin_memory=True,
batch_sampler=_sampler,
)
else:
print('---- No balance sampler -----')
train_loader = torch.utils.data.DataLoader(
dataset,
batch_size=opt.batch_size,
num_workers=opt.num_workers,
pin_memory=True,
shuffle=True,
)
val_loader = torch.utils.data.DataLoader(
dataset_valid,
batch_size=opt.batch_size_validation,
num_workers=opt.num_workers,
pin_memory=True,
)
if opt.model_type == "spatialtransformer_temporalconv":
model = SpatialTransformerTemporalConv(
num_frame=opt.sequence_length, in_chans=2 if opt.rm_conf else 3, spatial_embed_dim=opt.embedding_spatial_size, out_dim=opt.embedding_layer_size, num_joints=17, kernel_frame=opt.kernel_frame)
elif opt.model_type == "spatiotemporal_transformer":
model = SpatioTemporalTransformer(
num_frame=opt.sequence_length, in_chans=2 if opt.rm_conf else 3, spatial_embed_dim=opt.embedding_spatial_size, out_dim=opt.embedding_layer_size, num_joints=17)
elif opt.model_type == "gaitgraph":
model = get_model_resgcn()
else:
raise ValueError("No model type support:", opt.model_type)
unify_log = UnifyLog(opt, model)
print("# parameters: ", count_parameters(model))
# Load checkpoint or weights
load_checkpoint(model, opt)
if torch.cuda.device_count() > 1:
model = torch.nn.DataParallel(model)
if opt.loss_func == 'supcon':
loss_func = SupConLoss(temperature=0.004, base_temperature=0.004)
else:
loss_func = losses.TripletMarginLoss(margin=0.01)
if opt.cuda:
model.cuda()
loss_func.cuda()
# Trainer
optimizer, scheduler, scaler = get_trainer(model, opt, len(train_loader))
best_acc = 0
loss = 0
for epoch in tqdm(range(opt.start_epoch, opt.epochs + 1)):
# train for one epoch
time1 = time.time()
loss = train(
train_loader, model, loss_func, optimizer, scheduler, scaler, epoch, opt
)
time2 = time.time()
print(f"epoch {epoch}, total time {time2 - time1:.2f}")
unify_log.log(
{'loss/train': loss, 'lr': optimizer.param_groups[0]["lr"]}, step=epoch)
if epoch % opt.test_epoch_interval == 0 or epoch == opt.epochs:
# evaluation
accuracy_avg, sub_accuracies = evaluate(
val_loader, model, opt.evaluation_fn)
unify_log.log(
{'accuracy/validation/avg': accuracy_avg}, step=epoch)
for key, sub_accuracy in sub_accuracies.items():
unify_log.log(
{'accuracy/validation/'+key: sub_accuracy}, step=epoch)
print(f"epoch {epoch}, avg accuracy {accuracy_avg:.4f}")
is_best = accuracy_avg > best_acc
if is_best:
best_acc = accuracy_avg
unify_log.save_model(model, 'last.pth')
print(f"best accuracy: {best_acc*100:.2f}")
if __name__ == "__main__":
opt = parse_option()
opt.evaluation_fn = _evaluate_casia_b
main(opt)