-
Notifications
You must be signed in to change notification settings - Fork 2
/
train.py
402 lines (339 loc) · 18.5 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
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
# -*- coding： utf-8 -*-
'''
@Time: 2022/5/12 16:18
@Author:YilanZhang
@Filename:train.py
@Software:PyCharm
@Email:[email protected]
'''
import argparse
import torch
import torch.optim as optim
from torch.utils.data import DataLoader
import numpy as np
from torch.autograd import Variable
import os
import time
import torch.nn.functional as F
from src.eval_metrics import ConfusionMatrix
from models.TFormer import TFormer
from src.dataloader import load_dataset,dataset,train_data_transformation,test_data_transformation
'''function for saving model'''
def modelSnapShot(model,newModelPath,oldModelPath=None,onlyBestModel=False):
if onlyBestModel and oldModelPath:
os.remove(oldModelPath)
torch.save(model.state_dict(),newModelPath)
def main(options):
# parse the input args
epochs = options['epochs']
modal_path = options['modal_path']
if not os.path.exists(modal_path):
os.makedirs(modal_path)
log_path = options['log_path']
if not os.path.exists(log_path):
os.makedirs(log_path)
label_t = options['labels']
class_num = options['class_num']
patience = options['patience']
batch_size = options['batch_size']
weight_decay = options['weight_decay']
learning_rate = options['learning_rate']
dir_release = options['dir_release']
log_file = open(log_path + 'train_log.txt', 'w')
#load dataset
derm_data_group = load_dataset(dir_release=dir_release)
# load model
model = TFormer(class_num)
# parallel training
if options['cuda']:
model = model.cuda()
print("Model initialized")
optimizer = optim.Adam(model.parameters(), lr=learning_rate, betas=(0.9, 0.999),
weight_decay=weight_decay) # 更新网络参数，使用不同的更新规则
# 余弦衰减
lr_scheduler = optim.lr_scheduler.CosineAnnealingLR(optimizer, T_max=epochs)
# build a logger of training process and show it
print('===========Training Params===============')
log_file.write('===========Training Params===============' + '\n')
for name, param in options.items():
print('{}: {}'.format(name, param))
log_file.write('{}: {}'.format(name, param) + "\n")
print('========================================')
log_file.write('========================================' + '\n')
# setup training
complete = True
best_accuracy = 0
train_iterator = DataLoader(dataset(derm=derm_data_group,shape=(224,224),mode='train'),
batch_size=batch_size,shuffle=True,num_workers=4)
valid_iterator = DataLoader(dataset(derm= derm_data_group,shape=(224,224),mode='valid'),
batch_size=1,shuffle=False,num_workers=2)
test_iterator = DataLoader(dataset(derm=derm_data_group,shape=(224,224),mode='test'),
batch_size=1,shuffle=False,num_workers=2)
print("Start training...")
log_file.write("Start training...")
log_file.flush() # 清空
old_model_path = None
start_time = time.time()
try:
for e in range(epochs):
model.train()
train_loss = 0.0
print("Training epoch:{}".format(e))
log_file.write("Training epoch:{}".format(e)+'\n')
start_time_epoch = time.time()
for batchIndex, (der_data,cli_data,meta_data,target) in enumerate(train_iterator):
# Diagostic label
diagnosis_label = target[0].squeeze(1).cuda()
# Seven-Point Checklikst labels
pn_label = target[1].squeeze(1).cuda()
bmv_label = target[2].squeeze(1).cuda()
vs_label = target[3].squeeze(1).cuda()
pig_label = target[4].squeeze(1).cuda()
str_label = target[5].squeeze(1).cuda()
dag_label = target[6].squeeze(1).cuda()
rs_label = target[7].squeeze(1).cuda()
if options['cuda']:
der_data,cli_data,meta_data = der_data.cuda(), cli_data.cuda(),meta_data.cuda().float()
der_data,cli_data,meta_data= Variable(der_data),Variable(cli_data),Variable(meta_data)
optimizer.zero_grad()
output = model(meta_data,cli_data,der_data)
#multi label loss
loss = torch.true_divide(
model.criterion(output[0],diagnosis_label)
+ model.criterion(output[1],pn_label)
+ model.criterion(output[2],bmv_label)
+ model.criterion(output[3],vs_label)
+ model.criterion(output[4],pig_label)
+ model.criterion(output[5],str_label)
+ model.criterion(output[6],dag_label)
+ model.criterion(output[7],rs_label),8
)
loss.backward()
avg_loss = loss.item()
train_loss += avg_loss
optimizer.step()
if batchIndex % 50 == 0 and batchIndex > 0:
predicted_results = output[0].data.max(1)[1]
correctResultsNum = predicted_results.cpu().eq(diagnosis_label.cpu()).sum()
accuracy = correctResultsNum.item() * 1.0 / len(der_data) # .item()得到一个元素张量里面的元素值
print('Training epoch: {} [{}/{}], Loss: {:.4f}, Accuracy: {:.4f}, Learning rate: {}'.format(e,
batchIndex * len(der_data),len(train_iterator.dataset),loss.item(), accuracy,optimizer.param_groups[0]['lr']))
print("Epoch {} complete! Average Training loss: {:.4f}".format(e, train_loss / len(train_iterator)))
log_file.write("Epoch {} complete! Average Training loss: {:.4f}".format(e, train_loss / len(train_iterator)) + '\n')
log_file.flush()
# Terminate the training process if run into NaN
if np.isnan(train_loss):
print("Training got into NaN values...\n\n")
complete = False
break
# 余弦衰减
lr_scheduler.step()
'''Validation'''
model.eval()
avg_valid_loss = 0
total_correct_results_num = 0
val_confusion_diag = ConfusionMatrix(num_classes=class_num, labels=label_t)
val_confusion_pn = ConfusionMatrix(num_classes=class_num, labels=label_t)
val_confusion_bmv = ConfusionMatrix(num_classes=class_num, labels=label_t)
val_confusion_vs = ConfusionMatrix(num_classes=class_num, labels=label_t)
val_confusion_pig = ConfusionMatrix(num_classes=class_num, labels=label_t)
val_confusion_str = ConfusionMatrix(num_classes=class_num, labels=label_t)
val_confusion_dag = ConfusionMatrix(num_classes=class_num, labels=label_t)
val_confusion_rs = ConfusionMatrix(num_classes=class_num, labels=label_t)
for der_data, cli_data, meta_data, target in valid_iterator:
# target = torch.squeeze(target, 1) # torch.squeeze()对数据维数进行压缩，去掉target中维数为1的维度
# Diagostic label
diagnosis_label = target[0].squeeze(1).cuda()
# Seven-Point Checklikst labels
pn_label = target[1].squeeze(1).cuda()
bmv_label = target[2].squeeze(1).cuda()
vs_label = target[3].squeeze(1).cuda()
pig_label = target[4].squeeze(1).cuda()
str_label = target[5].squeeze(1).cuda()
dag_label = target[6].squeeze(1).cuda()
rs_label = target[7].squeeze(1).cuda()
if options['cuda']:
der_data, cli_data, meta_data = der_data.cuda(), cli_data.cuda(), meta_data.cuda().float()
der_data, cli_data, meta_data = Variable(der_data), Variable(cli_data), Variable(meta_data)
output = model(meta_data, cli_data, der_data)
valid_loss = torch.true_divide(
model.criterion(output[0],diagnosis_label)
+ model.criterion(output[1],pn_label)
+ model.criterion(output[2],bmv_label)
+ model.criterion(output[3],vs_label)
+ model.criterion(output[4],pig_label)
+ model.criterion(output[5],str_label)
+ model.criterion(output[6],dag_label)
+ model.criterion(output[7],rs_label),8
)
avg_valid_loss += valid_loss.item()
predicted_result = output[0].data.max(1)[1]
predictions_pn = output[1].data.max(1)[1]
predictions_bmv = output[2].data.max(1)[1]
predictions_vs = output[3].data.max(1)[1]
predictions_pig = output[4].data.max(1)[1]
predictions_str = output[5].data.max(1)[1]
predictions_dag = output[6].data.max(1)[1]
predictions_rs = output[7].data.max(1)[1]
total_correct_results_num += predicted_result.cpu().eq(diagnosis_label.cpu()).sum()
val_confusion_diag.update(predicted_result.cpu().numpy(),diagnosis_label.cpu().numpy())
val_confusion_pn.update(predictions_pn.cpu().numpy(), pn_label.cpu().numpy())
val_confusion_bmv.update(predictions_bmv.cpu().numpy(), bmv_label.cpu().numpy())
val_confusion_vs.update(predictions_vs.cpu().numpy(), vs_label.cpu().numpy())
val_confusion_pig.update(predictions_pig.cpu().numpy(), pig_label.cpu().numpy())
val_confusion_str.update(predictions_str.cpu().numpy(), str_label.cpu().numpy())
val_confusion_dag.update(predictions_dag.cpu().numpy(), dag_label.cpu().numpy())
val_confusion_rs.update(predictions_rs.cpu().numpy(), rs_label.cpu().numpy())
if np.isnan(avg_valid_loss):
print("Training got into NaN values...\n\n")
complete = False
break
avg_valid_loss = avg_valid_loss / len(valid_iterator)
dia_acc = val_confusion_diag.summary(log_file)
pn_acc = val_confusion_pn.summary(log_file)
bmv_acc = val_confusion_bmv.summary(log_file)
vs_acc = val_confusion_vs.summary(log_file)
pig_acc = val_confusion_pig.summary(log_file)
str_acc = val_confusion_str.summary(log_file)
dag_acc = val_confusion_dag.summary(log_file)
rs_acc = val_confusion_rs.summary(log_file)
accuracy_valid =100.0 * (dia_acc+pn_acc+bmv_acc+vs_acc+pig_acc+str_acc+dag_acc+rs_acc)/8.0
print("Valid loss is:{:.4f},average accuracy:{:.4f}%".format(avg_valid_loss, accuracy_valid))
log_file.write("Valid loss is:{:.4f},average accuracy:{:.4f}%".format(avg_valid_loss, accuracy_valid) + '\n')
if (accuracy_valid > best_accuracy):
curr_patience = patience
best_accuracy = accuracy_valid
new_model_path = os.path.join(modal_path, 'bestacc_model_{}.pth'.format(e))
modelSnapShot(model, new_model_path, oldModelPath=old_model_path, onlyBestModel=True)
old_model_path = new_model_path
print("Found new best model, saving to disk...")
else:
curr_patience -= 1
if (e % 10 == 0 and e >= 30) or e==epochs:
modelSnapShot(model, os.path.join(modal_path, 'model-' + str(e) + '.pth'))
if curr_patience <= 0:
break
end_time_epoch = time.time()
training_time_epoch = end_time_epoch - start_time_epoch
total_training_time = time.time() - start_time
remaining_time = training_time_epoch * epochs - total_training_time
print("Total training time: {:.4f}s, {:.4f} s/epoch, Estimated remaining time: {:.4f}s".format(
total_training_time, training_time_epoch, remaining_time))
if complete:
model.load_state_dict(torch.load(old_model_path), strict=True)
model.eval()
avg_test_loss = 0
confusion_diag =ConfusionMatrix(num_classes=class_num,labels =label_t)
confusion_pn = ConfusionMatrix(num_classes=class_num, labels=label_t)
confusion_bmv = ConfusionMatrix(num_classes=class_num, labels=label_t)
confusion_vs = ConfusionMatrix(num_classes=class_num, labels=label_t)
confusion_pig = ConfusionMatrix(num_classes=class_num, labels=label_t)
confusion_str = ConfusionMatrix(num_classes=class_num, labels=label_t)
confusion_dag = ConfusionMatrix(num_classes=class_num, labels=label_t)
confusion_rs = ConfusionMatrix(num_classes=class_num, labels=label_t)
for der_data,cli_data,meta_data,target in test_iterator:
# target=torch.squeeze(target,1) #torch.squeeze()对数据维数进行压缩，去掉target中维数为1的维度
# Diagostic label
diagnosis_label = target[0].squeeze(1).cuda()
# Seven-Point Checklikst labels
pn_label = target[1].squeeze(1).cuda()
bmv_label = target[2].squeeze(1).cuda()
vs_label = target[3].squeeze(1).cuda()
pig_label = target[4].squeeze(1).cuda()
str_label = target[5].squeeze(1).cuda()
dag_label = target[6].squeeze(1).cuda()
rs_label = target[7].squeeze(1).cuda()
if options['cuda']:
der_data,cli_data,meta_data= der_data.cuda(), cli_data.cuda(),meta_data.cuda().float()
der_data,cli_data,meta_data= Variable(der_data),Variable(cli_data),Variable(meta_data)
output = model(meta_data,cli_data,der_data)
test_loss = torch.true_divide(
model.criterion(output[0],diagnosis_label)
+ model.criterion(output[1],pn_label)
+ model.criterion(output[2],bmv_label)
+ model.criterion(output[3],vs_label)
+ model.criterion(output[4],pig_label)
+ model.criterion(output[5],str_label)
+ model.criterion(output[6],dag_label)
+ model.criterion(output[7],rs_label),8
)
avg_test_loss += test_loss.item()
#confusion matrix
ret,predictions_diag = torch.max(output[0].data,1)
ret, predictions_pn = torch.max(output[1].data, 1)
ret, predictions_bmv = torch.max(output[2].data, 1)
ret, predictions_vs = torch.max(output[3].data, 1)
ret, predictions_pig = torch.max(output[4].data, 1)
ret, predictions_str = torch.max(output[5].data, 1)
ret, predictions_dag = torch.max(output[6].data, 1)
ret, predictions_rs = torch.max(output[7].data, 1)
confusion_diag.update(predictions_diag.cpu().numpy(),diagnosis_label.cpu().numpy())
confusion_pn.update(predictions_pn.cpu().numpy(), pn_label.cpu().numpy())
confusion_bmv.update(predictions_bmv.cpu().numpy(), bmv_label.cpu().numpy())
confusion_vs.update(predictions_vs.cpu().numpy(), vs_label.cpu().numpy())
confusion_pig.update(predictions_pig.cpu().numpy(), pig_label.cpu().numpy())
confusion_str.update(predictions_str.cpu().numpy(), str_label.cpu().numpy())
confusion_dag.update(predictions_dag.cpu().numpy(), dag_label.cpu().numpy())
confusion_rs.update(predictions_rs.cpu().numpy(), rs_label.cpu().numpy())
print("Daig:\n")
log_file.write("Daig:\n")
confusion_diag.summary(log_file)
print("PN:\n")
log_file.write("PN:\n")
confusion_pn.summary(log_file)
print("BMV:\n")
log_file.write("BMV:\n")
confusion_bmv.summary(log_file)
print("VS:\n")
log_file.write("VS:\n")
confusion_vs.summary(log_file)
print("PIG:\n")
log_file.write("PIG:\n")
confusion_pig.summary(log_file)
print("STR:\n")
log_file.write("STR:\n")
confusion_str.summary(log_file)
print("DAG:\n")
log_file.write("DAG:\n")
confusion_dag.summary(log_file)
print("RS:\n")
log_file.write("RS:\n")
confusion_rs.summary(log_file)
log_file.flush()
except Exception:
import traceback
traceback.print_exc()
finally:
log_file.close()
OPTIONS = argparse.ArgumentParser()
# # parse the input args
# epochs = options['epochs']
OPTIONS.add_argument('--epochs',dest='epochs',type=int,default=100)
# dir_release = options['dir_release']
OPTIONS.add_argument('--dir_release',dest='dir_release',type=str,default="./data/derm7pt/release_v0/")
# modal_path = options['modal_path']
OPTIONS.add_argument('--modal_path',dest='modal_path',type=str,
default="./result")
# log_path = options['log_path']
OPTIONS.add_argument('--log_path',dest='log_path',type=str,
default="./result")
# labels = options['labels']
OPTIONS.add_argument('--labels',dest='labels',default=[0,1,2,3,4])
# class_num = options['class_num']
OPTIONS.add_argument('--class_num',dest='class_num',type=int,default=5)
# patience = options['patience']
OPTIONS.add_argument('--patience',dest='patience',type=int,default=100)
# batch_size = options['batch_size']
OPTIONS.add_argument('--batch_size',dest='batch_size',type=int,default=32)
# weight_decay = options['weight_decay']
OPTIONS.add_argument('--weight_decay',dest='weight_decay',type=float,default=1e-4)
# learning_rate = options['learning_rate']
OPTIONS.add_argument('--learning_rate',dest='learning_rate',type=float,default=0.0001)
OPTIONS.add_argument('--cuda', dest='cuda', type=bool, default=True)
OPTIONS.add_argument('--pretrained', dest='pretrained', type=bool, default=True)
PARAMS = vars(OPTIONS.parse_args())
if __name__ == "__main__":
os.environ["CUDA_VISIBLE_DEVICES"] = "0"
args = OPTIONS.parse_args()
PARAMS = vars(OPTIONS.parse_args())
main(PARAMS)