-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathmain.py
221 lines (185 loc) · 8.38 KB
/
main.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
import torch
from torch.autograd import Variable
import torch.nn as nn
import torch.optim as optim
import numpy as np
import pickle
import argparse
import os
from model import *
from dataloader import *
#torch.cuda.set_device(1)
def train(model,train_data,verification_data,valid_data,EPOCH,path) :
#optimizer = optim.SGD([
# {'params': model.layer1.parameters(), 'lr': 0.1, 'weight_decay': 0.0005},
# {'params': model.layer2.parameters(), 'lr': 0.1, 'weight_decay': 0.0005},
# {'params': model.layer3.parameters(), 'lr': 0.1, 'weight_decay': 0.0005},
# {'params': model.layer4.parameters(), 'weight_decay': 0.0005},
# {'params': model.linear.parameters(), 'weight_decay': 0.0005},
# {'params': model.final.parameters()}
# ], lr=0.05)
#optimizer = optim.Adam(model.parameters(), lr=0.0005)
optimizer = optim.SGD(model.parameters(), lr=0.1, momentum=0.7, weight_decay=0.0005)
scheduler = optim.lr_scheduler.ExponentialLR(optimizer,gamma=0.9)
loss_function = nn.BCELoss()
train_loss_list = []
verification_loss_list = []
verification_acc_list = []
validation_loss_list = []
validation_acc_list = []
train_batch_size = 128
valid_batch_size = 20
for epoch in range(EPOCH) :
print("Epoch %d" %(epoch+1))
# set learning rate & momentum
scheduler.step()
loss_arr = []
model.train()
model.batch_size = train_batch_size
# train
for i, (images,labels) in enumerate(train_data) :
model.batch_size = train_batch_size
image1 = []
image2 = []
for j in range(images.size(0)) :
image1.append(images[j][0].view(1,1,images.size(2),images.size(3)))
image2.append(images[j][1].view(1,1,images.size(2),images.size(3)))
image1 = Variable(torch.cat(image1)).cuda()
image2 = Variable(torch.cat(image2)).cuda()
labels = Variable(labels.type(torch.FloatTensor),requires_grad=False).view(-1,1).cuda()
scores = model(image1,image2)
loss = loss_function(scores,labels)
loss_arr.append(loss.cpu().data[0])
optimizer.zero_grad()
loss.backward()
optimizer.step()
#if i%50 == 0 :
# print("Iter %d Train loss : %.5f" %(i,loss.cpu().data[0]))
print("Train mean loss(train) : %.5f" %(np.array(loss_arr).mean()))
train_loss_list.append(np.array(loss_arr).mean())
# verification
num_sample = 0
num_correct = 0
loss_arr = []
model.eval()
model.batch_size = valid_batch_size
for i, (images,labels) in enumerate(verification_data) :
image1 = []
image2 = []
for j in range(images.size(0)) :
image1.append(images[j][0].view(1,1,images.size(2),images.size(3)))
image2.append(images[j][1].view(1,1,images.size(2),images.size(3)))
image1 = Variable(torch.cat(image1)).cuda()
image2 = Variable(torch.cat(image2)).cuda()
labels = Variable(labels.type(torch.FloatTensor)).view(-1,1).cuda()
scores = model(image1,image2)
loss = loss_function(scores,labels)
loss_arr.append(loss.cpu().data[0])
inference = scores.cpu().data.numpy().reshape(-1)
answer = labels.cpu().data.numpy().reshape(-1)
for k in range(len(inference)) :
num_sample += 1
if inference[k] >= 0.5 and answer[k] == 1 :
num_correct += 1
elif inference[k] < 0.5 and answer[k] == 0 :
num_correct += 1
print("Verification loss : %.5f" %(np.array(loss_arr).mean()))
print("Verification accuracy : %.3f" %(float(num_correct/num_sample)))
verification_loss_list.append(np.array(loss_arr).mean())
verification_acc_list.append(float(num_correct/num_sample))
# validation
num_sample = 0
num_correct = 0
loss_arr = []
model.eval()
model.batch_size = valid_batch_size
for i, (images,labels) in enumerate(valid_data) :
model.batch_size = valid_batch_size
model.eval()
image1 = []
image2 = []
for j in range(images.size(0)) :
image1.append(images[j][0].view(1,1,images.size(2),images.size(3)))
image2.append(images[j][1].view(1,1,images.size(2),images.size(3)))
image1 = Variable(torch.cat(image1)).cuda()
image2 = Variable(torch.cat(image2)).cuda()
labels = Variable(labels.type(torch.FloatTensor)).view(-1,1).cuda()
#targets = Variable(labels.type(torch.FloatTensor)).view(-1,1).cuda()
scores = model(image1,image2)
loss = loss_function(scores,labels)
#loss = loss_function(targets,scores)
loss_arr.append(loss.cpu().data[0])
num_sample += 1
inference = scores.cpu().data.numpy()
answer = labels.cpu().data.numpy()
if np.argmax(inference) == np.argmax(answer) :
num_correct += 1
print("Validation loss : %.5f" %(np.array(loss_arr).mean()))
print("Validation accuracy : %.3f" %(float(num_correct/num_sample)))
validation_loss_list.append(np.array(loss_arr).mean())
validation_acc_list.append(float(num_correct/num_sample))
#if min(validation_loss_list) >= np.array(loss_arr).mean() :
param_path = os.path.join(path,'epoch'+str(epoch)+'param.param')
torch.save(model.state_dict(),param_path)
print("")
print("Learning Finished")
print("Final Train Loss : %.5f" %(train_loss_list[len(train_loss_list)-1]))
print("Final Validation Loss : %.5f" %(validation_loss_list[len(validation_loss_list)-1]))
print("Final Validation Accuracy : %.3f\n" %(validation_acc_list[len(validation_acc_list)-1]))
return model, train_loss_list, validation_loss_list, validation_acc_list
def test(model,test_data) :
model.eval()
model.batch_size = 20
num_sample = 0
num_correct = 0
for i, (images,labels) in enumerate(test_data) :
model.batch_size = 20
model.eval()
image1 = []
image2 = []
for i in range(images.size(0)) :
image1.append(images[i][0].view(1,1,images.size(2),images.size(3)))
image2.append(images[i][1].view(1,1,images.size(2),images.size(3)))
image1 = Variable(torch.cat(image1)).cuda()
image2 = Variable(torch.cat(image2)).cuda()
labels = Variable(labels.type(torch.FloatTensor)).view(-1,1).cuda()
scores = model(image1,image2)
num_sample += 1
inference = scores.cpu().data.numpy()
answer = labels.cpu().data.numpy()
if np.argmax(inference) == np.argmax(answer) :
num_correct += 1
test_acc = float(num_correct/num_sample)
print("Test Accuracy : %.3f\n" %(test_acc))
return test_acc
def main(image_path,batch_size,EPOCH,seed) :
torch.manual_seed(seed)
model = Model(batch_size)
model.cuda()
# Call train_data, test_data
with open(os.path.join(image_path,'train_loader.pkl'),'rb') as f :
train_data = pickle.load(f)
with open(os.path.join(image_path,'verification_loader.pkl'),'rb') as f :
verification_data = pickle.load(f)
with open(os.path.join(image_path,'valid_loader.pkl'),'rb') as f :
valid_data = pickle.load(f)
with open(os.path.join(image_path,'test_loader.pkl'),'rb') as f :
test_data = pickle.load(f)
model, train_loss, validation_loss, validation_acc = train(model,train_data,verification_data,valid_data,EPOCH,image_path)
test_acc = test(model,test_data)
if __name__ == "__main__" :
parser = argparse.ArgumentParser()
parser.add_argument('--image_path', type=str, default='./data',
help='Image path')
parser.add_argument('--batch_size', type=int, default=128,
help='batch size')
parser.add_argument('--EPOCH', type=int, default=200,
help='max epoch')
parser.add_argument('--seed', type=int, default=0,
help='torch seed')
args = parser.parse_args()
image_path = args.image_path
batch_size = args.batch_size
EPOCH = args.EPOCH
seed = args.seed
main(image_path,batch_size,EPOCH,seed)