-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathmyTrain.py
170 lines (153 loc) · 8.93 KB
/
myTrain.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
# -*- coding:utf-8 -*-
# @Time: 2023-11-1 21:58
# @Author: TonyWang-SEU ([email protected])
# @File: myTrain.py
# @ProjectName: PolypNet
import os
# import logging
# import torch
import torch.backends.cudnn as cudnn
import torch.optim as optim
# from lib.base_model_for_train import *
# from lib.full_model_for_train import *
from lib.model_for_train import *
from utils.trainer_for_six_logits import *
from utils.utils import adjust_lr, adjust_lr_v2, adjust_lr_v3, adjust_lr_with_warmup, init_logger
from utils.dataloader import get_loader, test_dataset
from torch.utils.tensorboard import SummaryWriter
if __name__ == '__main__':
import argparse
parser = argparse.ArgumentParser()
parser.add_argument('--epoch', type=int, default=100)
parser.add_argument('--lr', type=float, default=1e-4)
parser.add_argument('--optimizer', type=str, default='Adam', choices=['AdamW', 'Adam', 'SGD'])
parser.add_argument('--batch_size', type=int, default=3) # 本地测试
parser.add_argument('--train_size', type=int, default=352)
parser.add_argument('--gradient_clip', type=float, default=0.5)
parser.add_argument('--lr_decay_rate', type=float, default=0.5)
parser.add_argument('--lr_decay_epoch', type=int, default=50)
parser.add_argument('--load_checkpoint_path', type=str, default=None)
parser.add_argument('--gpu_id', type=str, default='0')
parser.add_argument('--train_data_path', type=str, default=r'E:\data\Polyp-Dataset\PolypPVT-dataset\TrainDataset')
parser.add_argument('--val_data_path', type=str, default=r'E:\data\Polyp-Dataset\PolypPVT-dataset\TestDataset\CVC-300')
parser.add_argument('--train_model_save_path', type=str, default=r'./checkpoint/record_0')
parser.add_argument('--train_model_save_epoch', type=int, default=50)
parser.add_argument('--multi_scale_training', type=bool, default=False)
# # ---- network settings ----
# parser.add_argument('--net_sc_type', type=str, default='sc0', choices=['sc0', 'sc1', 'sc2', 'sc3'])
# parser.add_argument('--net_um_type', type=str, default='um0', choices=['um0', 'um1'])
# ---- 2023.10.12 ----
parser.add_argument('--net_type', type=str, default='train_model_arch',
choices=[
'train_model_arch', #
])
parser.add_argument('--net_backbone', type=str, default='pvt-v2-b2',
choices=['res2net50', 'resnet50', 'pvt-v2-b0', 'pvt-v2-b1', 'pvt-v2-b2', 'pvt-v2-b4', 'efficientnet-b1',
'efficientnet-b4'])
# ---- param for middle channel ----
parser.add_argument('--net_planes', type=int, default=64)
# ---- param for rf ----
# parser.add_argument('--net_rf_shuffles', type=list, default=[4, 8])
parser.add_argument('--net_pgs_reduction_scale', type=int, default=4)
parser.add_argument('--net_pgs_se_channel_reduction', type=int, default=16)
# ---- param for df ----
parser.add_argument('--net_df_channel_reduction', type=list, default=[4, 4])
# ---- param for sc ----
parser.add_argument('--net_sc_middle_channle_scale', type=int, default=2)
parser.add_argument('--net_sc_ksize', default=[3, 3, 3])
parser.add_argument('--net_sc_hidden_scale', type=int, default=32)
parser.add_argument('--net_sc_gas_ksize', type=int, default=1)
parser.add_argument('--net_sc_gas_ksize_scale', default=[15, 15, 15])
# ---- pretrained dict ----
parser.add_argument('--net_pretrained_backbone_path', type=str, default=r'E:/model/backbone/pvt_v2_b2.pth')
opt = parser.parse_args()
# ---- set logger and start training ----
os.makedirs(opt.train_model_save_path, exist_ok=True)
logger = init_logger(save_path=opt.train_model_save_path)
print('=' * 35, ' Start Training ', '=' * 35)
logger.info('>>> Start Training')
logger.info('>>> config:{}'.format(opt))
# ---- set device ----
os.environ['CUDA_VISIBLE_DEVICES'] = opt.gpu_id
print('>>> Using Gpu: ' + opt.gpu_id)
logger.info('>>> Using Gpu: ' + opt.gpu_id)
cudnn.benchmark = True
# ---- set seed ----
torch.manual_seed(0)
torch.cuda.manual_seed(0)
# ---- build the net ----
# ---- GN Slim ----
if opt.net_type == 'train_model_arch':
print('>>> Using Train Model')
model = SEPNet_TrainModel(backbone_name=opt.net_backbone,
pretrained_backbone_path=opt.net_pretrained_backbone_path,
mid_channels=opt.net_planes,
rf_reduction_scale=opt.net_pgs_reduction_scale,
rf_se_channel_reduction=opt.net_pgs_se_channel_reduction,
df_channel_reduction=opt.net_df_channel_reduction,
sc_hidden_scale=opt.net_sc_hidden_scale,
sc_gas_ksize=opt.net_sc_gas_ksize,
sc_middle_channle_scale=opt.net_sc_middle_channle_scale,
sc_ksize=opt.net_sc_ksize,
sc_gas_ksize_scale=opt.net_sc_gas_ksize_scale,
).cuda()
else:
raise Exception('Model not set!')
# ---- set optimizer ----
if opt.optimizer == 'AdamW':
my_optimizer = optim.AdamW(params=model.parameters(), lr=opt.lr, weight_decay=1e-4)
print('>>> Using optimizer: AdamW')
logger.info('>>> Using optimizer: AdamW')
elif opt.optimizer == 'Adam':
my_optimizer = optim.Adam(params=model.parameters(), lr=opt.lr, weight_decay=1e-4)
print('>>> Using optimizer: Adam')
logger.info('>>> Using optimizer: Adam')
elif opt.optimizer == 'SGD':
my_optimizer = optim.SGD(params=model.parameters(), lr=opt.lr, weight_decay=1e-4, momentum=0.9)
print('>>> Using optimizer: SGD')
logger.info('>>> Using optimizer: SGD')
else:
raise Exception('Optimizer not set!')
# optim_schedule = optim.lr_scheduler.CosineAnnealingLR(optimizer=my_optimizer, T_max=opt.epoch, eta_min=1e-5) # change the learning rate
# ---- set dataloader ----
# train_image_root = '{}/Imgs/'.format(opt.train_data_path)
# train_gt_root = '{}/GT/'.format(opt.train_data_path)
train_image_root = '{}/images/'.format(opt.train_data_path)
train_gt_root = '{}/masks/'.format(opt.train_data_path)
train_loader = get_loader(img_root=train_image_root, gt_root=train_gt_root, batch_size=opt.batch_size,
train_size=opt.train_size, shuffle=True, num_workers=8, pin_memory=True, prefetch_factor=4)
# val_image_root = '{}/Imgs/'.format(opt.val_data_path)
# val_gt_root = '{}/GT/'.format(opt.val_data_path)
val_image_root = '{}/images/'.format(opt.val_data_path)
val_gt_root = '{}/masks/'.format(opt.val_data_path)
val_loader = test_dataset(image_root=val_image_root, gt_root=val_gt_root, test_size=opt.train_size)
# ---- whether load from checkpoint ----
if opt.load_checkpoint_path is not None:
model.load_state_dict(torch.load(opt.load_checkpoint_path)['model_state_dict'])
my_optimizer.load_state_dict(torch.load(opt.load_checkpoint_path)['optimizer_state_dict'])
start_epoch = torch.load(opt.load_checkpoint_path)['epoch']
print('>>> Successfully load checkpoint model at Epoch: ', start_epoch)
logger.info('>>> Successfully load checkpoint model at Epoch:{}'.format(start_epoch)) # revised log at 2023/8/16
else:
# ---- set start epoch and train from scratch----
start_epoch = 1
# ---- set writer ----
from datetime import datetime
log_dir = opt.train_model_save_path + '/summary_{}'.format(datetime.now().strftime("%Y%m%d_%H%M%S"))
writer = SummaryWriter(log_dir=log_dir)
# ---- Start training ----
for epoch in range(start_epoch, opt.epoch + 1):
adjust_lr_with_warmup(optimizer=my_optimizer, init_lr=opt.lr, epoch=epoch, decay_rate=opt.lr_decay_rate,
decay_epoch=opt.lr_decay_epoch,
min_lr=0.1 * opt.lr, warmup_epoch=30)
train_stage(train_loader=train_loader, model=model, optimizer=my_optimizer, epoch=epoch, opt=opt, logger=logger, writer=writer)
# schedule
# adjust_lr_v3(optimizer=my_optimizer, init_lr=opt.lr, epoch=epoch, decay_rate=opt.lr_decay_rate, decay_epoch=opt.lr_decay_epoch,
# min_lr=0.1 * opt.lr)
# optim_schedule.step()
# if epoch == start_epoch or (epoch % 2 == 1 and epoch > 10) or epoch > 100:
# if epoch == start_epoch or (epoch % 10 == 0) or epoch > 100:
val_stage_with_multiMetrics(test_loader=val_loader, model=model, epoch=epoch, opt=opt, logger=logger, start_epoch=start_epoch,
writer=writer)
print('>>> Training finished!')
logger.info('>>> Training finished!')