-
Notifications
You must be signed in to change notification settings - Fork 1
/
train.py
165 lines (144 loc) · 6.54 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
import torch
from torch import optim
from torch.utils.data import DataLoader
import os
from tqdm import tqdm
from torch.utils.tensorboard import SummaryWriter
from model import Yolo, YoloLoss
from utils.utils import get_class_names, get_anchors
from utils.data import YoloDataset, yolo_dataset_collate
from config import cfg
from utils.train_utils import get_box_from_out, draw_multi_box_in_tensor
def train(**args):
# 1.参数设置
cfg.parse(**args)
class_names = get_class_names(cfg.class_names_path)
anchors = get_anchors(cfg.anchors_path)
num_classes = len(class_names)
num_anchors = len(anchors[0])
record_epoch = 0
print(f'num classes: {num_classes} num anchors: {num_anchors}')
# 2.模型
net = Yolo(num_classes, num_anchors)
if os.path.exists(cfg.pretrain_model):
pth = torch.load(cfg.pretrain_model, map_location=lambda storage, loc: storage)
if 'model' in pth:
record_epoch = pth['epoch']
pretrained_dict = pth['model']
else:
pretrained_dict = pth
# load part weight
model_dict = net.state_dict()
changed_keys = [k for k,v in pretrained_dict.items() if k in model_dict and v.shape != model_dict[k].shape]
if changed_keys:
print('model changed!')
print('--->', changed_keys)
record_epoch = 0
pretrained_dict = {k:v for k,v in pretrained_dict.items() if k in model_dict and v.shape == model_dict[k].shape}
# pretrained_dict.pop('yolo_head1.head.1.weight')
model_dict.update(pretrained_dict)
net.load_state_dict(model_dict)
print(f'load weight: {cfg.pretrain_model} epoch:{record_epoch}')
else:
print("haven't weight!")
if torch.cuda.is_available():
# net = torch.nn.DataParallel(net)
torch.backends.cudnn.benchmark = True
net.cuda()
net.train()
# 3.损失函数、优化器、lr
yolo_losses = YoloLoss( # YoloLoss中的get_target()会通过传入的模型预测的out尺寸选择对应的anchor(2, 3, 2)
anchors.reshape(-1, 2),
num_classes,
(cfg.input_shape[1], cfg.input_shape[0]),
cfg.smooth_label,
torch.cuda.is_available(),
cfg.loss_normalize
)
optimizer = optim.Adam(net.parameters(), cfg.lr)
if cfg.use_cosine_lr:
lr_scheduler = optim.lr_scheduler.CosineAnnealingLR(optimizer, T_max=cfg.T_max, eta_min=cfg.eta_min)
else:
lr_scheduler = optim.lr_scheduler.StepLR(optimizer, step_size=cfg.step_size, gamma=cfg.gamma)
# 4.数据集
train_dataset = YoloDataset(cfg.train_datasets_images_path, cfg.train_datasets_labels_path, (cfg.input_shape[0], cfg.input_shape[1]), is_train=True)
train_dataloader = DataLoader(
train_dataset,
shuffle=True,
batch_size=cfg.batch_size,
num_workers=cfg.cpu_count,
pin_memory=True,
drop_last=True,
collate_fn=yolo_dataset_collate
)
# 5.其它（训练无关）
writer = SummaryWriter(comment=f'-{record_epoch}')
# 6.训练
print('start train...')
for epoch in tqdm(range(cfg.max_epoch), position=1):
for k, (imgs, labels) in enumerate(tqdm(train_dataloader, position=0)):
imgs = torch.from_numpy(imgs).type(torch.FloatTensor)
labels = [torch.from_numpy(label).type(torch.FloatTensor) for label in labels] # (torch.Tensor只允许生成维度一致的，torch.Tensor([[1], [2, 3]])不被允许)
if torch.cuda.is_available():
imgs = imgs.cuda()
def closure():
optimizer.zero_grad()
out = net(imgs)
losses = []
num_pos_all = 0
for i in range(2):
loss_item, num_pos = yolo_losses(out[i], labels)
losses.append(loss_item)
num_pos_all += num_pos
loss = sum(losses) / num_pos_all
loss.backward()
writer.add_scalar('Loss/train/batch', loss.detach().item(), epoch * len(train_dataloader) + k)
return loss
optimizer.step(closure)
# debug
if os.path.exists(cfg.debug):
import ipdb; ipdb.set_trace()
lr_scheduler.step()
with torch.no_grad():
record_epoch += 1
if epoch % cfg.every_save == 0:
if not os.path.exists(cfg.save_folder):
os.mkdir(cfg.save_folder)
torch.save({
'model' : net.state_dict(),
'epoch' : record_epoch,
}, cfg.pretrain_model)
if epoch % cfg.every_valid == 0:
valid(net, writer=writer, record_epoch=record_epoch)
def valid(net, **kwargs):
writer = kwargs.get('writer', None)
record_epoch = kwargs.get('record_epoch', 0)
val_dataset = YoloDataset(cfg.valid_datasets_images_path, cfg.valid_datasets_labels_path, (cfg.input_shape[0], cfg.input_shape[1]), is_train=False)
val_dataloader = DataLoader(
val_dataset,
batch_size=cfg.batch_size,
num_workers=cfg.cpu_count,
pin_memory=True,
drop_last=False,
collate_fn=yolo_dataset_collate,
shuffle=True
)
for imgs, labels in tqdm(val_dataloader):
imgs = torch.from_numpy(imgs).type(torch.FloatTensor)
labels = [torch.from_numpy(label).type(torch.FloatTensor) for label in labels] # (torch.Tensor只允许生成维度一致的，torch.Tensor([[1], [2, 3]])不被允许)
if torch.cuda.is_available():
imgs = imgs.cuda()
out = net(imgs)
batch_boxes = get_box_from_out(out, class_names_path=cfg.class_names_path)
labels = [label.detach().cpu().numpy() for label in labels]
if writer:
mark_imgs = draw_multi_box_in_tensor(imgs.cpu(), labels, class_names=get_class_names(cfg.class_names_path), colors=[[44, 255, 44]], format='cxcywh', resize_box=True)
mark_imgs = draw_multi_box_in_tensor(mark_imgs, batch_boxes, class_names=get_class_names(cfg.class_names_path), colors=[[44, 44, 255]], format='cxcywh', resize_box=True)
writer.add_images('Images/valid', mark_imgs, record_epoch)
writer = None
if __name__ == '__main__':
train(
# train_datasets_labels_path = '/home/data/datasets/coco2017/coco/labels/train2017',
# train_datasets_images_path = '/home/data/datasets/coco2017/coco/images/train2017',
# class_names_path = './cfg/coco.txt',
)