-
Notifications
You must be signed in to change notification settings - Fork 3
/
train.py
301 lines (236 loc) · 10.8 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
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
from model.MSNet import *
from model.Aconvnet import *
from model.Densenet121 import *
import torch
from torch import nn, optim
import argparse
from torch.utils.data import DataLoader
from torchvision import transforms
import warnings
import os
from tensorboardX import SummaryWriter
from Dataset import*
from tqdm import tqdm
from early_stop import EarlyStopping
from torchsummary import summary
warnings.simplefilter(action='ignore', category=FutureWarning)
def accuracy(output, target, topk=(1,)):
with torch.no_grad():
maxk = max(topk)
batch_size = target.size(0)
_, pred = output.topk(maxk, 1, True, True)
pred = pred.t()
correct = pred.eq(target[None])
return correct
def validate(val_loader, model):
F = nn.CrossEntropyLoss()
model.eval()
sum = 0
val_loss = 0
with torch.no_grad():
print('Validate:')
for i, (images, ASC_part, target,_) in enumerate(tqdm(val_loader)):
images = images.float().to(device)
target = target.to(device)
ASC_part = ASC_part.float().to(device)
output = model(images, ASC_part)
val_loss += F(output,target)
result = accuracy(output, target)
sum += result.sum()
acc = sum/len(val_loader.dataset)
val_loss = val_loss/len(val_loader)
return acc, val_loss
def parameter_setting(args):
config = {}
config['arch'] = args.arch
config['datatxt_train'] = args.datatxt_train
config['datatxt_val'] = args.datatxt_val
config['datatxt_OFA1'] = args.datatxt_OFA1
config['datatxt_OFA2'] = args.datatxt_OFA2
config['datatxt_OFA3'] = args.datatxt_OFA3
config['cate_num'] = args.cate_num
config['batch_size'] = args.batch_size
config['num_epochs'] = args.num_epochs
config['save_path'] = args.save_path
config['pretrain'] = args.pretrain
config['part_num'] = args.part_num
config['patience'] = args.patience
config['attention_setting'] = args.attention_setting
config['device'] = args.device
config['train_num'] = args.train_num
return config
def get_dataloader(config, data_transforms):
dataset_train = Mstar_Components(config['datatxt_train'], transform=data_transforms)
dataloader = {}
dataloader['train'] = DataLoader(dataset_train,
batch_size=config['batch_size'],
shuffle=True,
drop_last=False,
num_workers=4,
)
dataset_test = Mstar_Components(config['datatxt_val'], transform=data_transforms)
dataloader['val'] = DataLoader(dataset_test,
batch_size=config['batch_size'],
shuffle=False,
drop_last=False,
num_workers=4,
)
dataset_test = Mstar_Components(config['datatxt_OFA1'], transform=data_transforms)
dataloader['OFA1'] = DataLoader(dataset_test,
batch_size=config['batch_size'],
shuffle=False,
drop_last=False,
num_workers=4,
)
dataset_test = Mstar_Components(config['datatxt_OFA2'], transform=data_transforms)
dataloader['OFA2'] = DataLoader(dataset_test,
batch_size=config['batch_size'],
shuffle=False,
drop_last=False,
num_workers=4,
)
dataset_test = Mstar_Components(config['datatxt_OFA3'], transform=data_transforms)
dataloader['OFA3'] = DataLoader(dataset_test,
batch_size=config['batch_size'],
shuffle=False,
drop_last=False,
num_workers=1,
)
assert(next(iter(dataloader['train']))[1].size()[1]==config['part_num'])
assert(next(iter(dataloader['val']))[1].size()[1]==config['part_num'])
assert(next(iter(dataloader['OFA1']))[1].size()[1]==config['part_num'])
assert(next(iter(dataloader['OFA2']))[1].size()[1]==config['part_num'])
assert(next(iter(dataloader['OFA3']))[1].size()[1]==config['part_num'])
return dataloader
def load_pretrained_model(path, model):
pretrain_dict = torch.load(path)
model_dict = {}
state_dict = model.state_dict()
print('already loaded')
for k in pretrain_dict.keys():
if k in model.state_dict().keys():
model_dict[k] = pretrain_dict[k]
print(k)
state_dict.update(model_dict)
model.load_state_dict(state_dict)
return model
def train(config, i):
if not os.path.exists(config['save_path']):
os.makedirs(config['save_path'])
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
early_stopping = EarlyStopping(os.path.join(config['save_path'], '{}.pth'.format(i)), config['patience'])
save_dir = config['save_path']
if not os.path.exists(save_dir):
os.mkdir(save_dir)
log_path = os.path.join(save_dir, 'log{}'.format(i))
if not os.path.exists(log_path):
os.makedirs(log_path)
print(torch.cuda.get_device_name(),device)
if config['arch'] == 'MSNet_PIHA':
init_lr = 0.0005
data_transforms = transforms.Compose([
transforms.ToTensor(),
transforms.CenterCrop(64),
])
model_CNN = eval(config['arch'])(config['cate_num'], config['part_num'], 100, config['attention_setting'])
elif config['arch'] == 'Aconvnet_PIHA':
init_lr = 0.005
data_transforms = transforms.Compose([
transforms.ToTensor(),
])
model_CNN = eval(config['arch'])(config['cate_num'], config['part_num'], config['attention_setting'])
elif config['arch'] == 'densenet121_PIHA':
init_lr = 0.0005
data_transforms = transforms.Compose([
transforms.ToTensor(),
transforms.CenterCrop(64),
])
model_CNN = eval(config['arch'])(config['cate_num'], part_num=config['part_num'], attention_setting=config['attention_setting'])
if config['pretrain']:
model_CNN = load_pretrained_model(config['pretrain'], model_CNN)
model_CNN.to(device)
optimizer = optim.SGD(filter(lambda p: p.requires_grad, model_CNN.parameters()), init_lr, momentum=0.9, weight_decay=1e-3, nesterov=True)
dataloader = get_dataloader(config, data_transforms)
loss_func = nn.CrossEntropyLoss()
writer = SummaryWriter(log_path)
for epoch in range(config['num_epochs']):
loss_sum = 0
model_CNN.train()
print('Epochs: ', epoch)
for data, ASC_part, labels, _ in tqdm(dataloader['train']):
data = data.float().to(device)
labels = labels.to(device)
# ASC_part = transforms.CenterCrop(64)(ASC_part).float().to(device)
ASC_part = ASC_part.float().to(device)
output = model_CNN(data, ASC_part)
loss = loss_func(output, labels)
loss_sum += loss
optimizer.zero_grad()
loss.backward()
optimizer.step()
val_loss = 0.0
acc, val_loss = validate(dataloader['val'], model_CNN)
writer.add_scalar('accuracy', acc, (epoch+1))
writer.add_scalars('loss', {'cls_train': loss_sum.item()/len(dataloader['train']),
'cls_val': val_loss.item(),}, epoch + 1)
print('{}准确率:{}'.format(epoch+1, acc.item()))
conuter = early_stopping(acc, model_CNN)
writer.add_scalar('conuter', conuter, (epoch+1))
if early_stopping.early_stop:
print("Early stopping")
break
acc_val, _ = validate(dataloader['val'], model_CNN)
acc_OFA1, _ = validate(dataloader['OFA1'], model_CNN)
acc_OFA2, _ = validate(dataloader['OFA2'], model_CNN)
acc_OFA3, _ = validate(dataloader['OFA3'], model_CNN)
print('*******************************************************')
return acc_val, acc_OFA1, acc_OFA2, acc_OFA3, epoch
if __name__ == '__main__':
parser = argparse.ArgumentParser(prog='CNN_training')
# data setting
parser.add_argument('--datatxt_train', default='data/Train/list/train.txt')
parser.add_argument('--datatxt_OFA1', default='data/OFA1_2/list/OFA1.txt')
parser.add_argument('--datatxt_OFA2', default='data/OFA1_2/list/OFA2.txt')
parser.add_argument('--datatxt_OFA3', default='data/OFA3/list/OFA3.txt')
parser.add_argument('--datatxt_val', default='data/Train/list/val.txt')
# training setting
parser.add_argument('--train_num', default=5)
parser.add_argument('--num_epochs', type=int, default=1000)
parser.add_argument('--patience', type=int, default=200)
parser.add_argument('--batch_size', type=int, nargs='+', default=32)
parser.add_argument('--device', default='0')
#Algorithmic hyperparameters
parser.add_argument('--arch', default='MSNet_PIHA')# Optional: MSNet_PIHA Aconvnet_PIHA densenet121_PIHA
parser.add_argument('--cate_num', type=int, default=10)
parser.add_argument('--part_num', type=int, default=4)
parser.add_argument('--attention_setting', default=True)
# other
parser.add_argument('--save_path', default='result/')
parser.add_argument('--pretrain', default=None)
args = parser.parse_args()
config = parameter_setting(args)
os.environ["CUDA_VISIBLE_DEVICES"] = config['device']
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
val_list = []
OFA1_list = []
OFA2_list = []
OFA3_list = []
stop_epoch_list = []
for i in range(config['train_num']):
val_acc, OFA1_acc, OFA2_acc, OFA3_acc, stop_epoch = train(config, i)
val_list.append(str(val_acc.item()))
OFA1_list.append(str(OFA1_acc.item()))
OFA2_list.append(str(OFA2_acc.item()))
OFA3_list.append(str(OFA3_acc.item()))
stop_epoch_list.append(str(stop_epoch))
val_result = 'val:' + '\t'.join(val_list) + '\n'
OFA1_result = 'OFA1:' + '\t'.join(OFA1_list) + '\n'
OFA2_result = 'OFA2:' + '\t'.join(OFA2_list) + '\n'
OFA3_result = 'OFA3:' + '\t'.join(OFA3_list) + '\n'
stop_epoch_result = 'stop_epoch:' + '\t'.join(stop_epoch_list) + '\n'
with open(os.path.join(config['save_path'], 'result.txt'), 'w') as f:
f.write(val_result)
f.write(OFA1_result)
f.write(OFA2_result)
f.write(OFA3_result)
f.write(stop_epoch_result)