-
Notifications
You must be signed in to change notification settings - Fork 2
/
attention_train.py
197 lines (152 loc) · 6.78 KB
/
attention_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
from src.train_resnet import *
import torch.nn.functional as F
'''
Get middle output from network
'''
#def hook(module, input, output):
class BasicAttentionBlock(nn.Module):
def __init__(self, input_dim=2048, n_filters=2048, kernel_size=1, padding=0, stride=1, nclasses=18):
super(BasicAttentionBlock, self).__init__()
self.feature = []
self.att_feat = []
self.attention_score = []
self.conv1 = nn.Conv2d(input_dim, 512, kernel_size=kernel_size, padding=padding, stride=stride)
self.activate = nn.ReLU()
self.conv2 = nn.Conv2d(512,1, kernel_size=kernel_size, padding=padding, stride=stride)
#self.atten_prob = F.softplus()
def forward(self ,x ):
self.feature.append(x.view(x.size(0), 2048, -1))
feat = x
#att_feat = F.normalize(feat, p=2,dim=0)
att_feat = feat
out = self.conv1(x)
out = self.activate(out)
out = self.conv2(out)
score = out
score = out.view(out.size(0), -1)
self.attention_score.append(out.view(out.size(0), -1))
prob = nn.Softplus(out)
prob = F.softplus(out)
att_feat = att_feat.view(att_feat.size(0), 2048, -1)
self.att_feat.append(att_feat)
prob = prob.view(prob.size(0), 1, -1)
prob = torch.transpose(prob, 1, 2)
#print(prob.shape, score.shape, att_feat.shape, torch.t(prob).shape)
#prob = self.atten_prob(out)
att_feat = torch.matmul(att_feat, (prob))
#print(att_feat.shape)
out = att_feat
return out
def getFeature(self):
result = ( self.feature, self.attention_score, self.att_feat)
self.feature.clear()
self.attention_score.clear()
self.att_feat.clear()
return result
def make_model_with_attention(nclasses, model_addr):
model = make_model(nclasses)
state_dict = torch.load(model_addr)
model.load_state_dict(state_dict)
for para in model.parameters():
para.requires_grad = False
model.avgpool = BasicAttentionBlock(input_dim=2048)
model.fc = nn.Linear(2048 ,nclasses)
for para in model.avgpool.parameters():
para.requires_grad = True
for para in model.fc.parameters():
para.requires_grad = True
return model
if __name__ == '__main__':
arg = get_args()
init_logger(log_addr=arg.log, name='attention')
logger('start train attention, batch_size :%d ' %(arg.batch_size))
logger('Set image resize size')
resize_size = (900, 900)
#crop_size = (720, 720)
crop_size = (224, 224)
logger('loading train data, data size:' )
train_data = list(load_train_img(arg.train_data, arg.offset, resize_size))
logger('Loading test data, data size:')
test_data = list(load_test_img(arg.test_data, arg.offset, resize_size))
logger('Test data size: %d, Train Data size: %d'%(len(train_data), len(test_data)))
logger('Croping data...')
train_data = Crop_data(train_data, crop_size, 3, resize_size)
test_data = Crop_data(test_data, crop_size, 3, resize_size)
if arg.shuffle:
logger('Shuffling data...' )
shuffle_data(train_data)
shuffle_data(test_data)
logger('Loading model ')
model = make_model_with_attention(nclasses=arg.n_classes,model_addr=arg.model)
if torch.cuda.device_count() > 1:
nn.DataParallel(model).cuda()
logger('gpu num: %d'%(torch.cuda.device_count()))
else:
model.cuda()
criterion = nn.CrossEntropyLoss()
#optimizer = torch.optim.SGD(model.parameters(), lr=arg.lr, momentum=0.8)
optimizer = torch.optim.SGD(filter(lambda p: p.requires_grad, model.parameters()),lr=arg.lr, momentum=0.8)
a = filter(lambda p:p.requires_grad, model.parameters())
getIgnore(a)
model.train()
logger("Start training ")
n_train_id = 0
n_test_id = 0
for iters in range(arg.iters+1):
if n_train_id + arg.batch_size < len(train_data):
zip_train_data = train_data[n_train_id:n_train_id+arg.batch_size]
else:
shuffle_data(train_data)
n_train_id = 0
zip_train_data = train_data[n_train_id:n_train_id+arg.batch_size]
n_train_id += arg.batch_size
image_train, label_train = zip(*zip_train_data)
zip_train_data.clear()
label_train = np.asarray(label_train)
image_train = Image2numpy(image_train)
image_train = np.array(image_train)
image_train = np.transpose(image_train, (0,3,1,2))
t_image = Variable(torch.from_numpy(image_train).float()).cuda()
t_label = Variable(torch.from_numpy(label_train).long()).cuda()
optimizer.zero_grad()
pre_label = model(t_image)
model.avgpool.getFeature()
loss = criterion(pre_label, t_label)
train_loss = loss.data[0]
loss.backward()
optimizer.step()
pre_label = pre_label.cpu().data.numpy()
if iters > 0 and iters %arg.display_iter == 0:
acc = get_acc(pre_label, label_train)
logger('Show Iter [%d / %d] -loss: %.10f, -acc:%.10f'%(iters, arg.iters, train_loss, acc))
if iters % arg.test_iters == 0:
if n_test_id + arg.test_batch_size < len(test_data):
zip_test_data = test_data[n_test_id:n_test_id+arg.test_batch_size]
else:
shuffle_data(test_data)
n_test_id = 0
zip_test_data = test_data[n_test_id:n_test_id+arg.test_batch_size]
n_test_id += arg.test_batch_size
image_test, label_test = zip(*zip_test_data)
zip_test_data.clear()
image_test = np.array(Image2numpy(image_test))
image_test = np.transpose(image_test, (0, 3, 1, 2))
label_test = np.asarray(label_test)
t_image = Variable(torch.from_numpy(image_test).float()).cuda()
model.eval()
test_pre_label = model(t_image)
model.avgpool.getFeature()
test_pre_label = test_pre_label.cpu().data.numpy()
acc = get_acc(test_pre_label, label_test)
logger('Test Iter [%d / %d] -acc: %.10f'%(iters, arg.iters, acc))
model.train()
if iters % arg.save_iter == 0 and iters > 0:
save_file_name = os.path.join(arg.output_dir, 'attention_iter_%d.pkl'%(iters))
torch.save(model.state_dict(), save_file_name)
logger('Save Iter [%d / %d] attention_iter_%d.pkl'%(iters, arg.iters, iters))
if iters % arg.lr_step == 0 and iters > 0:
lr = optimizer.state_dict()['param_groups'][0]['lr']
lr *= arg.gamma
logger("Update Iter [%d / %d] lr -> %.10f" %(iters, arg.iters, lr))
for param in optimizer.param_groups:
param['lr'] = lr