-
Notifications
You must be signed in to change notification settings - Fork 4
/
Train.py
242 lines (202 loc) · 10.9 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
240
241
import os
import torch
import torch.nn.functional as F
import numpy as np
from datetime import datetime
from torchvision.utils import make_grid
from lib.Network import Network
from utils.data_val import get_loader, test_dataset
from utils.utils import clip_gradient, adjust_lr, get_coef,cal_ual
from tensorboardX import SummaryWriter
import logging
import torch.backends.cudnn as cudnn
from torch import optim
def structure_loss(pred, mask):
weit = 1 + 5 * torch.abs(F.avg_pool2d(mask, kernel_size=31, stride=1, padding=15) - mask)
wbce = F.binary_cross_entropy_with_logits(pred, mask, reduce='none')
wbce = (weit * wbce).sum(dim=(2, 3)) / weit.sum(dim=(2, 3))
pred = torch.sigmoid(pred)
inter = ((pred * mask) * weit).sum(dim=(2, 3))
union = ((pred + mask) * weit).sum(dim=(2, 3))
wiou = 1 - (inter + 1) / (union - inter + 1)
return (wbce + wiou).mean()
def dice_loss(predict, target):
smooth = 1
p = 2
valid_mask = torch.ones_like(target)
predict = predict.contiguous().view(predict.shape[0], -1)
target = target.contiguous().view(target.shape[0], -1)
valid_mask = valid_mask.contiguous().view(valid_mask.shape[0], -1)
num = torch.sum(torch.mul(predict, target) * valid_mask, dim=1) * 2 + smooth
den = torch.sum((predict.pow(p) + target.pow(p)) * valid_mask, dim=1) + smooth
loss = 1 - num / den
return loss.mean()
def train(train_loader, model, optimizer, epoch, save_path, writer):
global step
model.train()
loss_all = 0
epoch_step = 0
try:
for i, (images, gts, edges) in enumerate(train_loader, start=1):
optimizer.zero_grad()
images = images.cuda(device=device_ids[0])
gts = gts.cuda(device=device_ids[0])
edges = edges.cuda(device=device_ids[0])
preds = model(images)
ual_coef = get_coef(iter_percentage=i/total_step, method='cos')
ual_loss = cal_ual(seg_logits=preds[4], seg_gts=gts)
ual_loss *= ual_coef
loss_init = structure_loss(preds[0], gts)*0.0625 + structure_loss(preds[1], gts)*0.125 + structure_loss(preds[2], gts)*0.25 + \
structure_loss(preds[3], gts)*0.5
loss_final = structure_loss(preds[4], gts)
loss_edge = dice_loss(preds[6], edges)*0.125 + dice_loss(preds[7], edges)*0.25 + \
dice_loss(preds[8], edges)*0.5
loss = loss_init + loss_final + loss_edge + 2*ual_loss
loss.backward()
clip_gradient(optimizer, opt.clip)
optimizer.step()
step += 1
epoch_step += 1
loss_all += loss.data
if i % 20 == 0 or i == total_step or i == 1:
print('{} Epoch [{:03d}/{:03d}], Step [{:04d}/{:04d}], Total_loss: {:.4f} Loss1: {:.4f} Loss2: {:0.4f} Loss3:{:0.4f}'.
format(datetime.now(), epoch, opt.epoch, i, total_step, loss.data, loss_init.data, loss_final.data, loss_edge.data))
logging.info(
'[Train Info]:Epoch [{:03d}/{:03d}], Step [{:04d}/{:04d}], Total_loss: {:.4f} Loss1: {:.4f} Loss2: {:0.4f} Loss3:{:0.4f}'.
format(epoch, opt.epoch, i, total_step, loss.data, loss_init.data, loss_final.data, loss_edge.data))
# TensorboardX-Loss
writer.add_scalars('Loss_Statistics',
{'Loss_init': loss_init.data, 'Loss_final': loss_final.data, 'Loss_edge': loss_edge.data, 'Loss_total': loss.data},
global_step=step)
# TensorboardX-Training Data
grid_image = make_grid(images[0].clone().cpu().data, 1, normalize=True)
writer.add_image('RGB', grid_image, step)
grid_image = make_grid(gts[0].clone().cpu().data, 1, normalize=True)
writer.add_image('GT', grid_image, step)
grid_image = make_grid(edges[0].clone().cpu().data, 1, normalize=True)
writer.add_image('Edge', grid_image, step)
# TensorboardX-Outputs
res = preds[0][0].clone()
res = res.sigmoid().data.cpu().numpy().squeeze()
res = (res - res.min()) / (res.max() - res.min() + 1e-8)
writer.add_image('Pred_init', torch.tensor(res), step, dataformats='HW')
res = preds[4][0].clone()
res = res.sigmoid().data.cpu().numpy().squeeze()
res = (res - res.min()) / (res.max() - res.min() + 1e-8)
writer.add_image('Pred_final', torch.tensor(res), step, dataformats='HW')
res = preds[8][0].clone()
res = res.sigmoid().data.cpu().numpy().squeeze()
res = (res - res.min()) / (res.max() - res.min() + 1e-8)
writer.add_image('Pred_edge', torch.tensor(res), step, dataformats='HW')
loss_all /= epoch_step
logging.info('[Train Info]: Epoch [{:03d}/{:03d}], Loss_AVG: {:.4f}'.format(epoch, opt.epoch, loss_all))
writer.add_scalar('Loss-epoch', loss_all, global_step=epoch)
if epoch % 80 == 0:
torch.save(model.state_dict(), save_path + 'Net_epoch_{}.pth'.format(epoch))
except KeyboardInterrupt:
print('Keyboard Interrupt: save model and exit.')
if not os.path.exists(save_path):
os.makedirs(save_path)
torch.save(model.state_dict(), save_path + 'Net_epoch_{}.pth'.format(epoch + 1))
print('Save checkpoints successfully!')
raise
def val(test_loader, model, epoch, save_path, writer):
"""
validation function
"""
global best_mae, best_epoch
model.eval()
with torch.no_grad():
mae_sum = 0
mae_sum_edge = 0
for i in range(test_loader.size):
image, gt, name, img_for_post = test_loader.load_data()
gt = np.asarray(gt, np.float32)
gt /= (gt.max() + 1e-8)
image = image.cuda(device=device_ids[0])
result = model(image)
res = F.upsample(result[4], size=gt.shape, mode='bilinear', align_corners=False)
res = res.sigmoid().data.cpu().numpy().squeeze()
res = (res - res.min()) / (res.max() - res.min() + 1e-8)
mae_sum += np.sum(np.abs(res - gt)) * 1.0 / (gt.shape[0] * gt.shape[1])
mae = mae_sum / test_loader.size
writer.add_scalar('MAE', torch.tensor(mae), global_step=epoch)
print('Epoch: {}, MAE: {}, bestMAE: {}, bestEpoch: {}.'.format(epoch, mae, best_mae, best_epoch))
if epoch == 1:
best_mae = mae
best_epoch = 1
else:
if mae < best_mae:
best_mae = mae
best_epoch = epoch
torch.save(model.state_dict(), save_path + 'Net_epoch_best.pth')
print('Save state_dict successfully! Best epoch:{}.'.format(epoch))
logging.info(
'[Val Info]:Epoch:{} MAE:{} bestEpoch:{} bestMAE:{}'.format(epoch, mae, best_epoch, best_mae))
if __name__ == '__main__':
import argparse
parser = argparse.ArgumentParser()
parser.add_argument('--epoch', type=int, default=300, help='epoch number')
parser.add_argument('--lr', type=float, default=1e-4, help='learning rate')
parser.add_argument('--batchsize', type=int, default=36, help='training batch size')
parser.add_argument('--trainsize', type=int, default=384, help='training dataset size')
parser.add_argument('--clip', type=float, default=0.5, help='gradient clipping margin')
parser.add_argument('--decay_rate', type=float, default=0.1, help='decay rate of learning rate')
parser.add_argument('--decay_epoch', type=int, default=80, help='every n epochs decay learning rate')
parser.add_argument('--load', type=str, default=None, help='train from checkpoints')
parser.add_argument('--gpu_id', type=str, default='4,5', help='train use gpu')
parser.add_argument('--train_root', type=str, default='/data0/hcm/dataset/COD/TrainDataset/',
help='the training rgb images root')
parser.add_argument('--val_root', type=str, default='/data0/hcm/dataset/COD/TestDataset/COD10K/',
help='the test rgb images root')
parser.add_argument('--save_path', type=str,default='./snapshot/model_new/',help='the path to save model and log')
opt = parser.parse_args()
os.environ["CUDA_VISIBLE_DEVICES"] = opt.gpu_id
print('USE GPU 2,3')
cudnn.benchmark = True
# build the model
device_ids = [0,1] # if you want to use more gpus than 2, you shoule change it just like when use opt.gpu_id='1,2,6,8' , device_ids = [0,1,2,3]
model = torch.nn.DataParallel(Network(channels=96), device_ids=device_ids)
model = model.cuda(device=device_ids[0])
if opt.load is not None:
model.load_state_dict(torch.load(opt.load))
print('load model from ', opt.load)
optimizer = torch.optim.Adam(model.parameters(), opt.lr)
save_path = opt.save_path
if not os.path.exists(save_path):
os.makedirs(save_path)
# load data
print('load data...')
train_loader = get_loader(image_root=opt.train_root + 'Imgs/',
gt_root=opt.train_root + 'GT/',
edge_root=opt.train_root + 'Edge/',
batchsize=opt.batchsize,
trainsize=opt.trainsize,
num_workers=8)
val_loader = test_dataset(image_root=opt.val_root + 'Imgs/',
gt_root=opt.val_root + 'GT/',
testsize=opt.trainsize)
total_step = len(train_loader)
# logging
logging.basicConfig(filename=save_path + 'log.log',
format='[%(asctime)s-%(filename)s-%(levelname)s:%(message)s]',
level=logging.INFO, filemode='a', datefmt='%Y-%m-%d %I:%M:%S %p')
logging.info("Network-Train")
logging.info('Config: epoch: {}; lr: {}; batchsize: {}; trainsize: {}; clip: {}; decay_rate: {}; load: {}; '
'save_path: {}; decay_epoch: {}'.format(opt.epoch, opt.lr, opt.batchsize, opt.trainsize, opt.clip,
opt.decay_rate, opt.load, save_path, opt.decay_epoch))
step = 0
writer = SummaryWriter(save_path + 'summary')
best_mae = 1
best_epoch = 0
# learning rate schedule
cosine_schedule = optim.lr_scheduler.CosineAnnealingLR(optimizer=optimizer, T_max=20, eta_min=1e-5)
print("Start train...")
for epoch in range(1, opt.epoch):
cur_lr = adjust_lr(optimizer, opt.lr, epoch, opt.decay_rate, opt.decay_epoch)
writer.add_scalar('learning_rate', cur_lr, global_step=epoch)
cosine_schedule.step()
writer.add_scalar('learning_rate', cosine_schedule.get_lr()[0], global_step=epoch)
logging.info('>>> current lr: {}'.format(cosine_schedule.get_lr()[0]))
train(train_loader, model, optimizer, epoch, save_path, writer)
val(val_loader, model, epoch, save_path, writer)