-
Notifications
You must be signed in to change notification settings - Fork 4
/
train.py
216 lines (184 loc) · 8.63 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
import datetime
import os
import torch
from torch import nn
from torch import optim
from torch.backends import cudnn
from torch.utils.data import DataLoader
from torchvision import transforms
from tqdm import tqdm
import joint_transforms
from config import ViSha_training_root
from dataset.VShadow_crosspairwise import CrossPairwiseImg
from misc import AvgMeter, check_mkdir
from networks.TVSD import TVSD
from torch.optim.lr_scheduler import StepLR
import math
from losses import lovasz_hinge, binary_xloss
import random
import torch.nn.functional as F
import numpy as np
from apex import amp
import time
os.environ["CUDA_VISIBLE_DEVICES"] = "0"
cudnn.deterministic = True
cudnn.benchmark = False
ckpt_path = './models'
exp_name = 'TVSD'
args = {
'max_epoch': 12,
'train_batch_size': 5,
'last_iter': 0,
'finetune_lr': 5e-5,
'scratch_lr': 5e-4,
'weight_decay': 5e-4,
'momentum': 0.9,
'snapshot': '',
'scale': 416,
'multi-scale': None,
'gpu': '0,1',
'multi-GPUs': False,
'fp16': True,
'warm_up_epochs': 3,
'seed': 2020
}
# fix random seed
np.random.seed(args['seed'])
torch.manual_seed(args['seed'])
torch.cuda.manual_seed(args['seed'])
# multi-GPUs training
if args['multi-GPUs']:
os.environ['CUDA_VISIBLE_DEVICES'] = args['gpu']
batch_size = args['train_batch_size'] * len(args['gpu'].split(','))
# single-GPU training
else:
torch.cuda.set_device(0)
batch_size = args['train_batch_size']
joint_transform = joint_transforms.Compose([
joint_transforms.RandomHorizontallyFlip(),
joint_transforms.Resize((args['scale'], args['scale']))
])
val_joint_transform = joint_transforms.Compose([
joint_transforms.Resize((args['scale'], args['scale']))
])
img_transform = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])
target_transform = transforms.ToTensor()
to_pil = transforms.ToPILImage()
print('=====>Dataset loading<======')
training_root = [ViSha_training_root] # training_root should be a list form, like [datasetA, datasetB, datasetC], here we use only one dataset.
train_set = CrossPairwiseImg(training_root, joint_transform, img_transform, target_transform)
train_loader = DataLoader(train_set, batch_size=batch_size, num_workers=8, shuffle=True)
print("max epoch:{}".format(args['max_epoch']))
ce_loss = nn.CrossEntropyLoss()
log_path = os.path.join(ckpt_path, exp_name, str(datetime.datetime.now()) + '.txt')
def main():
print('=====>Prepare Network {}<======'.format(exp_name))
# multi-GPUs training
if args['multi-GPUs']:
net = torch.nn.DataParallel(TVSD()).cuda().train()
params = [
{"params": net.module.encoder.parameters(), "lr": args['finetune_lr']},
{"params": net.module.co_attention.parameters(), "lr": args['scratch_lr']},
{"params": net.module.encoder.final_pre.parameters(), "lr": args['scratch_lr']},
{"params": net.module.co_attention.parameters(), "lr": args['scratch_lr']},
{"params": net.module.project.parameters(), "lr": args['scratch_lr']},
{"params": net.module.final_pre.parameters(), "lr": args['scratch_lr']}
]
# single-GPU training
else:
net = TVSD().cuda().train()
params = [
{"params": net.encoder.backbone.parameters(), "lr": args['finetune_lr']},
{"params": net.encoder.aspp.parameters(), "lr": args['scratch_lr']},
{"params": net.encoder.final_pre.parameters(), "lr": args['scratch_lr']},
{"params": net.co_attention.parameters(), "lr": args['scratch_lr']},
{"params": net.project.parameters(), "lr": args['scratch_lr']},
{"params": net.final_pre.parameters(), "lr": args['scratch_lr']}
]
# optimizer = optim.SGD(params, momentum=args['momentum'], weight_decay=args['weight_decay'])
optimizer = optim.Adam(params, betas=(0.9, 0.99), eps=6e-8, weight_decay=args['weight_decay'])
warm_up_with_cosine_lr = lambda epoch: epoch / args['warm_up_epochs'] if epoch <= args['warm_up_epochs'] else 0.5 * \
(math.cos((epoch-args['warm_up_epochs'])/(args['max_epoch']-args['warm_up_epochs'])*math.pi)+1)
scheduler = torch.optim.lr_scheduler.LambdaLR(optimizer, lr_lambda=warm_up_with_cosine_lr)
# scheduler = StepLR(optimizer, step_size=10, gamma=0.1) # change learning rate after 20000 iters
check_mkdir(ckpt_path)
check_mkdir(os.path.join(ckpt_path, exp_name))
open(log_path, 'w').write(str(args) + '\n\n')
if args['fp16']:
net, optimizer = amp.initialize(net, optimizer, opt_level="O1")
train(net, optimizer, scheduler)
def train(net, optimizer, scheduler):
curr_epoch = 1
curr_iter = 1
start = 0
print('=====>Start training<======')
while True:
loss_record1, loss_record2, loss_record3, loss_record4 = AvgMeter(), AvgMeter(), AvgMeter(), AvgMeter()
loss_record5, loss_record6, loss_record7, loss_record8 = AvgMeter(), AvgMeter(), AvgMeter(), AvgMeter()
for i, sample in enumerate(tqdm(train_loader, desc=f'Epoch: {curr_epoch}', ncols=100, ascii=' =', bar_format='{l_bar}{bar}|')):
exemplar, exemplar_gt, query, query_gt = sample['exemplar'].cuda(), sample['exemplar_gt'].cuda(), sample['query'].cuda(), sample['query_gt'].cuda()
other, other_gt = sample['other'].cuda(), sample['other_gt'].cuda()
optimizer.zero_grad()
exemplar_pre, query_pre, other_pre, scene_logits = net(exemplar, query, other)
bce_loss1 = binary_xloss(exemplar_pre, exemplar_gt)
bce_loss2 = binary_xloss(query_pre, query_gt)
bce_loss3 = binary_xloss(other_pre, other_gt)
loss_hinge1 = lovasz_hinge(exemplar_pre, exemplar_gt)
loss_hinge2 = lovasz_hinge(query_pre, query_gt)
loss_hinge3 = lovasz_hinge(other_pre, other_gt)
loss_seg = bce_loss1 + bce_loss2 + bce_loss3 + loss_hinge1 + loss_hinge2 + loss_hinge3
# classification loss
scene_labels = torch.zeros(scene_logits.shape[0], dtype=torch.long).cuda()
cla_loss = ce_loss(scene_logits, scene_labels) * 10
loss = loss_seg + cla_loss
if args['fp16']:
with amp.scale_loss(loss, optimizer) as scaled_loss:
scaled_loss.backward()
else:
loss.backward()
torch.nn.utils.clip_grad_norm_(net.parameters(), 12) # gradient clip
optimizer.step() # change gradient
loss_record1.update(bce_loss1.item(), batch_size)
loss_record2.update(bce_loss2.item(), batch_size)
loss_record3.update(bce_loss3.item(), batch_size)
loss_record4.update(loss_hinge1.item(), batch_size)
loss_record5.update(loss_hinge2.item(), batch_size)
loss_record6.update(loss_hinge3.item(), batch_size)
loss_record7.update(cla_loss.item(), batch_size)
curr_iter += 1
log = "iter: %d, bce1: %f5, bce2: %f5, bce3: %f5, hinge1: %f5, hinge2: %f5, hinge3: %f5, cla: %f5, lr: %f8"%\
(curr_iter, loss_record1.avg, loss_record2.avg, loss_record3.avg, loss_record4.avg, loss_record5.avg,
loss_record6.avg, loss_record7.avg, scheduler.get_lr()[0])
if (curr_iter-1) % 20 == 0:
elapsed = (time.clock() - start)
start = time.clock()
log_time = log + ' [time {}]'.format(elapsed)
print(log_time)
open(log_path, 'a').write(log + '\n')
if curr_epoch % 1 == 0:
if args['multi-GPUs']:
# torch.save(net.module.state_dict(), os.path.join(ckpt_path, exp_name, '%d.pth' % curr_epoch))
checkpoint = {
'model': net.module.state_dict(),
'optimizer': optimizer.state_dict(),
'amp': amp.state_dict()
}
torch.save(checkpoint, os.path.join(ckpt_path, exp_name, f'{curr_epoch}.pth'))
else:
# torch.save(net.state_dict(), os.path.join(ckpt_path, exp_name, '%d.pth' % curr_epoch))
checkpoint = {
'model': net.state_dict(),
'optimizer': optimizer.state_dict(),
'amp': amp.state_dict()
}
torch.save(checkpoint, os.path.join(ckpt_path, exp_name, f'{curr_epoch}.pth'))
if curr_epoch > args['max_epoch']:
# torch.save(net.state_dict(), os.path.join(ckpt_path, exp_name, '%d.pth' % curr_iter))
return
curr_epoch += 1
scheduler.step() # change learning rate after epoch
if __name__ == '__main__':
main()