-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy path2_train.py
355 lines (255 loc) · 10.1 KB
/
2_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
import os
import time
import torch
import argparse
from torch import optim
import torch.utils.data
import torchvision
from model.ConvRRN import ST_EncDec as Model
from model.CNN import CNN
class Dataset(torch.utils.data.Dataset):
def __init__(self, train=True, aug=False):
super(Dataset, self).__init__()
if train==True:
data_name = 'pt_dataset/train_data.pt'
label_name = 'pt_dataset/train_binary.pt'
else:
data_name = 'pt_dataset/valid_data.pt'
label_name = 'pt_dataset/valid_binary.pt'
# normalization
self.inputs = (torch.load(data_name)-0.5)/0.5
self.targets = torch.load(label_name).long()
counts = []
counts = [(self.targets==0).sum().item(),
(self.targets==1).sum().item()]
counts = [counts[1]/sum(counts), counts[0]/sum(counts)]
counts = torch.tensor(counts)
self.weight = counts
if train is True and aug is True:
# 1. shift 20
new_inputs = [self.inputs]
new_targets = [self.targets]
while True:
new_input = torch.zeros(self.inputs.size())
new_target = torch.zeros(self.targets.size()).long()
# self.inputs: n by 10 by 3 by h by w
new_input[:,:,:,:-20] = new_inputs[-1][:,:,:,20:]
new_input[:,:,:,:20] = new_inputs[-1][:,:,:,-20:]
new_target[:,:,:-20] = new_targets[-1][:,:,20:]
new_target[:,:,:20] = new_targets[-1][:,:,-20:]
new_inputs.append(new_input)
new_targets.append(new_target)
if len(new_inputs)==5:
break
self.inputs = torch.cat(new_inputs, 0)
self.targets = torch.cat(new_targets, 0).long()
# 2. flip vertically, flip horizontally, rotate 180
# 2-1) flip vertically
new_inputs = self.inputs.chunk(self.inputs.size()[-2], 3)[::-1]
new_inputs = torch.cat(new_inputs, 3)
new_targets = self.targets.chunk(self.targets.size()[-2], 2)[::-1]
new_targets = torch.cat(new_targets, 2).long()
self.inputs = torch.cat([self.inputs, new_inputs], 0)
self.targets = torch.cat([self.targets, new_targets], 0)
# 2-2) flip horizontally
new_inputs = self.inputs.chunk(self.inputs.size()[-1], 4)[::-1]
new_inputs = torch.cat(new_inputs, 4)
new_targets = self.targets.chunk(self.targets.size()[-1], 3)[::-1]
new_targets = torch.cat(new_targets, 3)
self.inputs = torch.cat([self.inputs, new_inputs], 0)
self.targets = torch.cat([self.targets, new_targets], 0)
# 3. reverse order
new_inputs = self.inputs.chunk(self.inputs.size()[1], 1)[::-1]
new_inputs = torch.cat(new_inputs, 1)
new_targets = self.targets.chunk(self.targets.size()[1], 1)[::-1]
new_targets = torch.cat(new_targets, 1)
self.inputs = torch.cat([self.inputs, new_inputs], 0)
self.targets = torch.cat([self.targets, new_targets], 0)
if train is True:
print('Trainset size:')
else:
print('Validset size:')
print(self.inputs.size(), self.targets.size())
def __getitem__(self, index):
return self.inputs[index], self.targets[index]
def __len__(self):
return len(self.inputs)
if __name__ == '__main__':
parser = argparse.ArgumentParser(description='PCB anomaly detection')
parser.add_argument('--batch_size', type=int, default=10)
parser.add_argument('--num_epochs', type=int, default=100)
parser.add_argument('--n_layers', type=int, default=2)
parser.add_argument('--lr', type=float, default=1e-4)
# parser.add_argument('--init_tr', type=float, default=0.25)
# parser.add_argument('--final_tr', type=float, default=0.0)
parser.add_argument('--gpu_ids', nargs='+', type=int, default=[0])
parser.add_argument('--is_attn', action='store_true')
parser.add_argument('--aug', action='store_true')
parser.add_argument('--mode', type=str, default='ConvLSTM', help='ConvLSTM, ConvConjLSTM, CNN')
args = parser.parse_args()
if not os.path.isdir('visualization'):
os.mkdir('visualization')
if args.mode=='CNN':
vis_dir = 'visualization/CNN'
else:
vis_dir = 'visualization/CRRN'
if not os.path.isdir(vis_dir):
os.mkdir(vis_dir)
# root_dir = 'result'
# if not os.path.isdir(root_dir):
# os.mkdir(root_dir)
# model_dir = root_dir + '/' + ((args.mode+'_a') if args.is_attn else args.mode)
# if not os.path.isdir(model_dir):
# os.mkdir(model_dir)
#
# model_file = '%s/%s'%(model_dir, 'model_dictionary.pt')
train_dataset = Dataset(train=True, aug=args.aug)
train_dataloader = torch.utils.data.DataLoader(train_dataset, batch_size=args.batch_size, shuffle=True, num_workers=1)
valid_dataset = Dataset(train=False)
valid_dataloader = torch.utils.data.DataLoader(valid_dataset, batch_size=args.batch_size, shuffle=False, num_workers=1)
if args.mode=='CNN':
model = CNN([3, 64, 64], [5,5])
else:
model = Model(args.mode, [3, 64, 64], [5,5], args.n_layers, args.is_attn)
optimizer = optim.Adam(model.parameters(), lr=args.lr, weight_decay=1e-5)
if torch.cuda.device_count()>1:
if args.gpu_ids==None:
print("Let's use", torch.cuda.device_count(), "GPUs!")
device = torch.device('cuda:0')
else:
print("Let's use", len(args.gpu_ids), "GPUs!")
device = torch.device('cuda:' + str(args.gpu_ids[0]))
else:
device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
print('args.gpu_ids', args.gpu_ids)
model = torch.nn.DataParallel(model, device_ids=args.gpu_ids)
model = model.to(device)
# criterion = torch.nn.BCELoss()
criterion = torch.nn.CrossEntropyLoss(weight=train_dataset.weight.to(device))
#
# test_loss = []
#
for epoch in range(args.num_epochs):
#
start_time = time.time()
each_train_loss = []
model.train()
origin_inputs = []
origin_outputs = []
out_prob = []
with torch.set_grad_enabled(True):
for inputs, targets in train_dataloader:
inputs = inputs.to(device)
targets = targets.to(device)
if (len(inputs)!=args.batch_size):
break
optimizer.zero_grad()
outputs = model(inputs)
# out_prob.append(torch.nn.Sigmoid()(outputs[:,:,1]))
# origin_outputs.append(targets) # batch by 10 by h by w
# print("targets", targets.size())
# print("outputs", outputs.size())
if args.mode!='CNN':
outputs = torch.nn.Sigmoid()(outputs.view(-1, *outputs.size()[2:]))
# outputs = outputs.view(outputs.size(0), -1)
# targets = targets.view(targets.size(0), -1)
targets = targets.view(-1, *targets.size()[2:])
err = criterion(outputs, targets.long())
err.backward()
torch.nn.utils.clip_grad_norm_(model.parameters(), 0.5)
optimizer.step()
each_train_loss.append(err.item())
epoch_train_loss = sum(each_train_loss)/len(each_train_loss)
each_valid_loss = []
model.eval()
# origin_inputs = []
# origin_outputs = []
# out_prob = []
threshold = torch.linspace(0,1,100).tolist()
cm = torch.zeros(2,2, dtype=torch.float)
tp, fp, fn, tn = {}, {}, {}, {}
for th in threshold:
tp[th] = 0
fp[th] = 0
fn[th] = 0
tn[th] = 0
# Validation
with torch.set_grad_enabled(False):
for inputs, targets in valid_dataloader:
inputs = inputs.to(device)
targets = targets.to(device)
if (len(inputs)!=args.batch_size):
break
# origin_inputs.append(inputs)
origin_inputs.append(inputs*0.5+0.5)
outputs = model(inputs) # batch by 10 by 2 by h by w
outputs = torch.nn.Sigmoid()(outputs)
if args.mode=='CNN':
temp_outputs = outputs.view(-1, 10, *temp_outputs.size()[1:])
out_prob.append(temp_outputs)
else:
outputs = outputs.squeeze()
out_prob.append(outputs)
origin_outputs.append(targets) # batch by 10 by h by w
if args.mode!='CNN':
outputs = outputs.view(-1, *outputs.size()[2:])
targets = targets.view(-1, *targets.size()[2:])
err = criterion(outputs, targets.long())
each_valid_loss.append(err.item())
# arg_outputs = torch.argmax(outputs, dim=1)
# for i in range(2):
# for j in range(2):
# pred = (arg_outputs==j)
# real = (targets==i)
# cm[i][j] += (pred&real).cpu().float().sum()
for th in threshold:
th_code = targets.int()*2 + (outputs[:,1]>th).int()
tp[th] += (th_code==3).sum().item()
fp[th] += (th_code==1).sum().item()
fn[th] += (th_code==2).sum().item()
tn[th] += (th_code==0).sum().item()
precision, recall, f1 = [], [], []
for th in threshold:
p = tp[th]/(tp[th]+fp[th]+1e-7)
r = tp[th]/(tp[th]+fn[th]+1e-7)
if p!=0 and r!=0:
precision.append(p)
recall.append(r)
f1_each = 2/(1/p + 1/r)
f1.append(f1_each)
print(max(f1))
epoch_valid_loss = sum(each_valid_loss)/len(each_valid_loss)
# save out_prob
origin_inputs = torch.cat(origin_inputs, 0) # N by 10 by 3 by h by w
origin_outputs = torch.cat(origin_outputs, 0)
out_prob = torch.cat(out_prob, 0) # 10*k by 10 by h by w
for i in range(len(out_prob)):
if not os.path.isdir('%s/iter_%03d'%(vis_dir,epoch)):
os.mkdir('%s/iter_%03d'%(vis_dir,epoch))
scaled_file = '%s/iter_%03d/%04d.png'%(vis_dir,epoch, i)
scaled_input = torchvision.utils.make_grid(origin_inputs[i], nrow=10, padding=2, pad_value=1)
# output_and_target = torch.cat([out_prob[i].unsqueeze(1), origin_outputs[i].unsqueeze(1).float()], 0)
output_and_target = torch.cat([origin_outputs[i].float(), out_prob[i][:,1]], 0).unsqueeze(1)
# scaled_result = torchvision.utils.make_grid(out_prob[i].unsqueeze(1), nrow=10, padding=2, pad_value=1)
scaled_result = torchvision.utils.make_grid(output_and_target, nrow=10, padding=2, pad_value=1)
scaled = torch.cat([scaled_input, scaled_result], 1)
torchvision.utils.save_image(scaled, scaled_file)
print('Epoch %03d:'%(epoch))
print('-> Train loss: %f'%(epoch_train_loss))
print('-> Valid loss: %f'%(epoch_valid_loss))
print(' -> Confusion matrix')
#
# print(epoch, 'train: ', epoch_train_loss, ', test: ', epoch_valid_loss, 'correct: ', correct/all_count, 'elapsed: %4.2f'%(time.time()-start_time))
#
# cm_sum = torch.sum(cm, dim=1).unsqueeze(1).expand_as(cm)
# cm/=cm_sum
# print(cm)
#
# model_dictionary = {'epoch': epoch,
# 'test_loss': test_loss,
# 'state_dict': list(model.children())[0].state_dict(),
# 'optimizer': optimizer.state_dict()
# }
#
# torch.save(model_dictionary, model_file)
#