-
Notifications
You must be signed in to change notification settings - Fork 0
/
train_base.py
222 lines (196 loc) · 7.56 KB
/
train_base.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
import os
import time
import argparse
import numpy as np
import cv2
import torch
from torch.backends import cudnn
from torch.nn import DataParallel
from torch.utils.data import DataLoader
from tqdm import tqdm
from torch import nn
from dataloader.generator import SUN,NYU, RGBD
from dataloader.citys_generate import CityscapesLoader
from dataloader.imagereader import imagefile
from utils.lr_scheduler import LRScheduler
from utils import helper
import torchvision.models as models
#================ load your net here ===================
from net.refinenet import RefineNet4CascadePoolingImproved as network
#from net.refinenet import RefineNet4Cascade as network
#from net.lw_mobilenet import mbv2 as network
#from net.lw_resnet import rf_lw50 as network
# from net.lw_resnet import rf_lw101 as network
#from net.lw_resnet import rf_lw152 as network
#from net.lw_shufflenetv2 import snv2 as network
def print_log(epoch,
lr,
train_metrics,
train_time,
val_metrics=None,
val_time=None,
val_iou=None,
val_acc = None,
val_mean_acc = None,
save_dir=None,
log_mode=None):
if epoch > 1:
log_mode = 'a'
train_metrics = np.mean(train_metrics, axis=0)
str0 = 'Epoch %03d (lr %.7f)' % (epoch, lr)
str0 += ', Train: time %3.2f loss: %2.4f' \
% (train_time, train_metrics)
f = open(save_dir + 'train_log.txt', log_mode)
if val_time is not None:
val_metrics = np.mean(val_metrics, axis=0)
str0 += ', Validation: time %3.2f loss: %2.4f iou: %.4f acc: %2.4f mean_acc: %2.4f' \
% (val_time, val_metrics, val_iou, val_acc, val_mean_acc)
print(str0)
f.write(str0)
f.write('\n')
f.close()
def train(data_loader, net, loss, optimizer, lr):
start_time = time.time()
net.train()
for param_group in optimizer.param_groups:
param_group['lr'] = lr
net.module.freeze_bn()
metrics = []
for i, batch in enumerate(tqdm(data_loader)):
data, heatmaps = batch['image'], batch['label']
data = data.cuda(async=True)
heatmaps = heatmaps.cuda(async=True)
prediction = net(data)
loss_output = loss(prediction, heatmaps)
optimizer.zero_grad()
loss_output.backward()
optimizer.step()
metrics.append(loss_output.item())
end_time = time.time()
metrics = np.asarray(metrics, np.float32)
return metrics, end_time - start_time
def validate(data_loader, net, loss, epoch, num_class):
start_time = time.time()
net.eval()
metrics = []
iou = 0
acc = 0
mean_acc = 0
for i, batch in enumerate(tqdm(data_loader)):
data, heatmaps = batch['image'], batch['label']
data = data.cuda(async=True)
heatmaps = heatmaps.cuda(async=True)
prediction = net(data)
loss_output = loss(prediction, heatmaps)
iou_, acc_, mean_acc_ = helper.cal_metric(heatmaps, prediction, num_class)
iou += iou_
acc += acc_
mean_acc += mean_acc_
metrics.append(loss_output.item())
iou /= len(data_loader)
acc /= len(data_loader)
mean_acc /= len(data_loader)
#img = helper.make_validation_img(batch['image'].numpy(),
# batch['label'].numpy(),
# prediction.cpu().numpy())
#cv2.imwrite('%s/validate_%d_%.4f.png'%(save_dir, epoch, iou),
# img[:, :, ::-1])
end_time = time.time()
metrics = np.asarray(metrics, np.float32)
return metrics, end_time - start_time, iou, acc, mean_acc
if __name__ == '__main__':
workers = 8
batch_size = 4
base_lr = 1e-3
parser = argparse.ArgumentParser()
parser.add_argument('-r', '--resume', default='', help='resume training from checkpoint ...', type=str)
parser.add_argument('-d', '--dataset', default='NYU', help='NYU or SUN or CITY', type=str)
args = parser.parse_args()
save_dir = './%s_base101/'%args.dataset
if not os.path.exists(save_dir):
os.mkdir(save_dir)
epochs = {'SUN':60,'NYU':300,'CITY':200}[args.dataset]
if args.dataset == 'CITY':
train_dataset = CityscapesLoader(split = 'train')
val_dataset = CityscapesLoader(split = 'val')
else:
train_dataset = RGBD(args.dataset,'train')
val_dataset = RGBD(args.dataset,'val')
train_loader = DataLoader(
train_dataset,
batch_size=batch_size,
shuffle=True,
num_workers=workers)
#val_dataset = CityscapesLoader(split = 'val')
val_loader = DataLoader(
val_dataset, batch_size=batch_size, shuffle=False, num_workers=workers)
num_class = {'SUN':37,'NYU':40,'CITY':19}[args.dataset]
ignore_label = {'SUN':255,'NYU':255,'CITY':250}[args.dataset]
loss = nn.CrossEntropyLoss(ignore_index=ignore_label)
patience = {'SUN':15,'NYU':60,'CITY':40}[args.dataset]
print('Train sample number: %d' % len(train_dataset))
print('Val sample number: %d' % len(val_dataset))
############################################################
net = network((3,640),num_classes = num_class,resnet_factory = models.resnet101, freeze_resnet=False)
start_epoch = 1
lr = base_lr
best_val_loss = float('inf')
log_mode = 'w'
if os.path.exists(args.resume):
print('loading checkpoint %s'%(args.resume))
checkpoint = torch.load(args.resume)
start_epoch = checkpoint['epoch'] + 1
lr = checkpoint['lr']
best_val_loss = checkpoint['best_val_loss']
net.load_state_dict(checkpoint['state_dict'])
log_mode = 'a'
net = net.cuda()
loss = loss.cuda()
cudnn.benchmark = True
#print(net.named_parameter())
net = DataParallel(net)
#for p in module.parameters():
#p.requires_grad = True
#optimizer = torch.optim.Adam(net.parameters(), lr)
optimizer = torch.optim.SGD(
net.parameters(), lr, momentum=0.9, weight_decay=5e-4)
lrs = LRScheduler(
lr, patience=patience, factor=0.5, min_lr=0.5*0.5*0.5 * lr, best_loss=best_val_loss)
#with torch.no_grad():
# val_metrics, val_time, val_iou, val_acc, val_mean_acc = validate(val_loader, net, loss, 0, num_class)
#print (val_iou, val_acc, val_mean_acc)
for epoch in range(start_epoch, epochs + 1):
train_metrics, train_time = train(train_loader, net, loss, optimizer,
lr)
with torch.no_grad():
val_metrics, val_time, val_iou, val_acc, val_mean_acc = validate(val_loader, net, loss, epoch, num_class)
print_log(
epoch,
lr,
train_metrics,
train_time,
val_metrics,
val_time,
val_iou,
val_acc,
val_mean_acc,
save_dir=save_dir,
log_mode=log_mode)
val_loss = np.mean(val_metrics)
lr = lrs.update_by_rule(val_loss)
if val_loss < best_val_loss or epoch % 10 == 0 or lr is None:
if val_loss < best_val_loss:
best_val_loss = val_loss
state_dict = net.state_dict()
for key in state_dict.keys():
state_dict[key] = state_dict[key].cpu()
torch.save({
'epoch': epoch,
'save_dir': save_dir,
'state_dict': state_dict,
'lr': lr,
'best_val_loss': best_val_loss
}, os.path.join(save_dir, 'ckpt_best.ckpt'))
if lr is None:
print('Training is early-stopped')
break