-
Notifications
You must be signed in to change notification settings - Fork 0
/
Training.py
202 lines (184 loc) · 8.94 KB
/
Training.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
from utils import Early_stopping, read_yaml
from Compute_loss import ComputeLoss
from Dataset_Generator import data_loader
from torch.optim import lr_scheduler
from Yolov4_architecture.model import Yolov4
from eval import eval_
from tqdm import tqdm
import torch
import argparse
import os
import warnings
warnings.filterwarnings("ignore")
try:
from apex import amp
except:
amp = None
class Training:
def __init__(self, cfg, aug):
self.device = 'cuda' if torch.cuda.is_available() else 'cpu'
self.best_map = 0
self.batch_size = cfg['batch_size']
self.anchors = cfg['anchors']
print('loading the model')
self.model = Yolov4(num_classes = cfg['num_classes'],
anchors = cfg['anchors'],
device = self.device,
).to(device = self.device)
self.optimizer = self.optimizer_(
cfg['optimizer'],
cfg['lr'] ,
cfg['momentum'],
cfg['weight_decay'])
self.eatly_stopping= Early_stopping(patience = cfg['patience'])
if cfg['pretrained'] and cfg['weights'][-3:]=='.pt':
self.load_model(cfg['weights'])
print('loading the data')
self.train_loader,_= data_loader(img_dir =cfg['train_image_path'],
ann_dir = cfg['train_ann_path'],
anchors = cfg['anchors'],
img_size = cfg['img_size'],
data_aug = True,
cfg_aug = aug,
max_boxes=cfg['max_boxes'],
num_classes = cfg['num_classes'],
anchor_thresh = cfg['anchor_thresh'] ,
batch_size = cfg['batch_size'],
strides = cfg['strides'],
train = True
)
self.val_loader ,self.count_classes = data_loader(img_dir = cfg['val_image_path'],
ann_dir = cfg['val_ann_path'],
anchors = cfg['anchors'],
img_size = cfg['img_size'],
num_classes= cfg['num_classes'],
batch_size = cfg['batch_size'],
)
self.notest = False
try:
self.test_loader ,self.count_classes_test = data_loader(img_dir = cfg['test_image_path'],
ann_dir = cfg['test_ann_path'],
anchors = cfg['anchors'],
img_size = cfg['img_size'],
num_classes= cfg['num_classes'],
batch_size = cfg['batch_size'],
)
except:
self.notest = True
self.epochs = cfg['epochs']
self.classes = dict(enumerate(cfg['classes']))
self.weights_path = cfg['weights']
self.img_size = cfg['img_size']
self.ComputeLoss = ComputeLoss(self.device, cfg['num_classes'],cfg['img_size'])
self.fp16 = cfg['fp16']
lf = lambda x: (1 - x / 100) * (1.0 - cfg['lr']) + cfg['lr']
self.scheduler = lr_scheduler.LambdaLR(self.optimizer, lr_lambda = lf)
def train(self):
print('################ Start Training ################')
if self.fp16 and amp:
self.model, self.optimizer = amp.initialize(self.yolov4, self.optimizer, opt_level="O1")
for epoch in range(self.epochs):
self.model.train()
box_loss = 0
obj_loss = 0
cls_loss = 0
for i,(imgs, targets) in enumerate(tqdm(self.train_loader, ascii = True, desc ="Training")):
imgs = imgs.to(device=self.device, non_blocking=True).float() / 255
preds = self.model(imgs)
loss, loss_item = self.ComputeLoss(preds, targets)
box_loss += loss_item[0].cpu().numpy()
obj_loss += loss_item[1].cpu().numpy()
cls_loss += loss_item[2].cpu().numpy()
self.optimizer.zero_grad()
if self.fp16 and amp:
with amp.scale_loss(loss, self.optimizer) as scaled_loss:
scaled_loss.backward()
else:
loss.backward()
self.optimizer.step()
self.optimizer.zero_grad()
self.scheduler.step()
self.scheduler.get_last_lr()
print(f'epoch {epoch}/{self.epochs} box loss = ',box_loss.round(decimals=3),' conf loss = ',obj_loss.round(decimals=3),' cls loss = ',cls_loss.round(decimals=3),'\n')
map, maps =eval_(device = self.device,
batch_size = self.batch_size,
img_size = self.img_size,
model = self.model,
val_loader = self.val_loader,
classes = self.classes,
count_classes = self.count_classes)
self.save_model(epoch,map[0],map[1])
### early stopping
if self.eatly_stopping(self.weighted_map(map[0],map[1])):
print('################ early_stopping ################')
break
### end epoch
print('################ Training Finished ################')
print('loading the best weights')
self.load_model(os.path.join(self.weights_path, 'best.pt'))
if not self.notest:
self.val_loader = self.test_loader
self.count_classes = self.count_classes_test
map, maps = eval_(device = self.device,
batch_size = self.batch_size,
img_size = self.img_size,
model = self.model,
val_loader = self.val_loader,
classes = self.classes,
count_classes = self.count_classes)
def weighted_map(self,map50,map50_95,Wmap50=0.1,Wmap50_95=0.9):
return map50 * Wmap50 + map50_95 * Wmap50_95
def load_model(self,weights_path):
ckpt = torch.load(weights_path, map_location=self.device)
self.model.load_state_dict(ckpt['weights'])
self.optimizer.load_state_dict(ckpt['optimizer'])
#self.best_map50_95=ckpt(['mAP50-95'])
#self.best_map50 = ckpt(['mAP50'])
self.best_map = 0
del ckpt
def save_model(self, epoch,map50,map50_95):
map_ = self.weighted_map(map50, map50_95)
ckpt ={'weights':self.model.state_dict(),
'mAP50':map50,
'mAP50-95' :map50_95,
'optimizer':self.optimizer.state_dict(),
'epoch':epoch,
'anchors':self.anchors,
'classes': list(self.classes.values())
}
torch.save(ckpt, os.path.join(self.weights_path,'last.pt'))
if map_>self.best_map:
self.best_map=map_
torch.save(ckpt, os.path.join(self.weights_path, 'best.pt'))
del ckpt
def optimizer_(self,opt_name='RMSProp',lr=0.001,momentum=0.9,weight_decay=1e-5):
if opt_name=='SGD':
optimizer = torch.optim.SGD(self.model.parameters(), lr=lr, momentum=momentum,weight_decay=weight_decay)
elif opt_name=='Adam':
optimizer = torch.optim.Adam(self.model.parameters(), lr=lr, betas=(momentum,0.999),weight_decay=weight_decay)
elif opt_name=='RMSProp':
optimizer = torch.optim.RMSprop(self.model.parameters(), lr=lr, momentum=momentum,weight_decay=weight_decay)
else:
raise NotImplementedError(f'optimizer {opt_name} not implemented')
return optimizer
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument("--data", type = str, default = "data/BBCD.yaml", help = "data.yaml")
parser.add_argument("--hyp", type = str, default = "data/hyp.yaml", help = "hyp.yaml")
parser.add_argument("--aug", type = str, default = "data/aug.yaml", help = "augmentation file")
parser.add_argument("--pretrained", type = bool, default = False, help = "Resume trainin")
parser.add_argument("--weights", type = str, default = "", help = "weights path")
parser.add_argument("--epochs", type = int, default = 100, help = "number of epochs")
parser.add_argument("--fp16", type = bool, default = False, help = "FP16 Training")
opt = parser.parse_args()
data, hyp, aug = read_yaml(opt.data), read_yaml(opt.hyp), read_yaml(opt.aug)
cfg = dict()
cfg.update(data)
cfg.update(hyp)
cfg['pretrained'] = opt.pretrained
cfg['weights'] = opt.weights
cfg['epochs'] = opt.epochs
cfg['anchors'] = torch.tensor(cfg['anchors']).view(3,3,2)
cfg['fp16'] = opt.fp16
train = Training(cfg, aug)
train.train()