forked from thuml/Separate_to_Adapt
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathstep_1_cifar_alexnet.py
228 lines (178 loc) · 9.13 KB
/
step_1_cifar_alexnet.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
from data import *
from utilities import *
from networks import *
import matplotlib.pyplot as plt
import numpy as np
from torchvision.datasets import CIFAR10
from torchvision.datasets import STL10
from torchvision import transforms
def skip(data, label, is_train):
return False
batch_size = 32
learning_rate = 1e-4 #1e-3
def get_dataset():
train_dataset = CIFAR10('../data', train=True, download=True,
transform=transforms.Compose([
transforms.Resize(256),
transforms.RandomHorizontalFlip(p=0.5),
transforms.RandomCrop(224),
transforms.ToTensor(),
transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
]))
test_dataset = STL10('../data', split='train', download=True,
transform=transforms.Compose([
transforms.Resize(256),
transforms.RandomHorizontalFlip(p=0.5),
transforms.RandomCrop(224),
transforms.ToTensor(),
transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
]))
def map_labels_stl_to_cifar10(labels_stl):
# map labels...
# 2 -> 1, 1 -> 2
# 6 -> 7, 7 -> 6
new_labels = np.zeros(len(labels_stl))
for idx, l in enumerate(labels_stl):
if l == 1:
new_labels[idx] = 2
elif l == 2:
new_labels[idx] = 1
elif l >= 5:
new_labels[idx] = 5
else:
new_labels[idx] = l
return np.asarray(new_labels)
def apply_one_hot(labels):
new_labels = np.zeros(len(labels))
for idx, l in enumerate(labels):
if l >= 5:
new_labels[idx] = 5
else:
new_labels[idx] = l
return np.asarray(new_labels)
test_dataset.labels = map_labels_stl_to_cifar10(test_dataset.labels)
train_dataset.train_labels = apply_one_hot(train_dataset.train_labels)
return train_dataset, test_dataset
source_dataset, target_dataset = get_dataset()
source_loader = torch.utils.data.DataLoader(source_dataset,
batch_size=batch_size, shuffle=True, num_workers=0)
target_loader = torch.utils.data.DataLoader(target_dataset,
batch_size=batch_size, shuffle=True, num_workers=0)
setGPU('0')
log = Logger('log/step_1', clear=True)
#discriminator_t = CLS_0(2048,2,bottle_neck_dim = 256).cuda()
discriminator_t = CLS_0(4096, 2, bottle_neck_dim=256).cuda()
discriminator_p = Discriminator(n = 5).cuda()
#feature_extractor = ResNetFc(model_name='resnet50',model_path='/home/liuhong/data/pytorchModels/resnet50.pth')
feature_extractor = AlexNetFc()
cls = CLS(feature_extractor.output_num(), 6, bottle_neck_dim=256)
net = nn.Sequential(feature_extractor, cls).cuda()
scheduler = lambda step, initial_lr : inverseDecaySheduler(step, initial_lr, gamma=10, power=0.75, max_iter=10000)
optimizer_discriminator_t = OptimWithSheduler(optim.SGD(discriminator_t.parameters(), lr=learning_rate, weight_decay=5e-4, momentum=0.9, nesterov=True),
scheduler)
optimizer_discriminator_p = OptimWithSheduler(optim.SGD(discriminator_p.parameters(), lr=learning_rate, weight_decay=5e-4, momentum=0.9, nesterov=True),
scheduler)
optimizer_cls = OptimWithSheduler(optim.SGD(cls.parameters(), lr=learning_rate, weight_decay=5e-4, momentum=0.9, nesterov=True),
scheduler)
# =========================train the multi-binary classifier
k=0
while k < 500:
for i, (batch_source, batch_target) in enumerate(zip(source_loader, target_loader)):
im_source, label_source = batch_source
im_target, label_target = batch_target
def one_hot_encoding(y):
y_onehot = y.numpy()
y_onehot = (np.arange(6) == y_onehot[:,None]).astype(np.float32)
y_onehot = torch.from_numpy(y_onehot)
return y_onehot
label_source = one_hot_encoding(label_source)
label_target = one_hot_encoding(label_target)
im_source, label_source = im_source.cuda(), label_source.cuda(non_blocking=True)
im_target, label_target = im_target.cuda(), label_target.cuda(non_blocking=True)
fs1, feature_source, __, predict_prob_source = net.forward(im_source)
ft1, feature_target, __, predict_prob_target = net.forward(im_target)
p0 = discriminator_p.forward(fs1)
p1 = discriminator_p.forward(ft1)
p2 = torch.sum(p1, dim = -1)
# =========================rank the output of the multi-binary classifiers
__,_,_,dptarget = discriminator_t.forward(ft1.detach())
try:
r = torch.sort(dptarget[:,1].detach(),dim = 0)[1][30:]
feature_otherep = torch.index_select(ft1, 0, r.view(2))
_, _, __, predict_prob_otherep = cls.forward(feature_otherep)
w = torch.sort(p2.detach(),dim = 0)[1][30:]
h = torch.sort(p2.detach(),dim = 0)[1][0:2]
feature_otherep2 = torch.index_select(ft1, 0, w.view(2))
feature_otherep1 = torch.index_select(ft1, 0, h.view(2))
_,_,_,pred00 = discriminator_t.forward(feature_otherep2)
_,_,_,pred01 = discriminator_t.forward(feature_otherep1)
except:
continue
# =========================loss function
ce = CrossEntropyLoss(label_source, predict_prob_source)
d1 = BCELossForMultiClassification(label_source[:,0:5],p0)
with OptimizerManager([optimizer_cls, optimizer_discriminator_p]):
loss = ce + d1
loss.backward()
k += 1
log.step += 1
if log.step % 10 == 1:
counter = AccuracyCounter()
counter.addOntBatch(variable_to_numpy(predict_prob_source), variable_to_numpy(label_source))
acc_train = Variable(torch.from_numpy(np.asarray([counter.reportAccuracy()], dtype=np.float32))).cuda()
track_scalars(log, ['ce', 'acc_train', 'd1'], globals())
if log.step % 100 == 0:
clear_output()
# =========================train the known/unknown discriminator
k=0
while k <400:
for i, (batch_source, batch_target) in enumerate(zip(source_loader, target_loader)):
im_source, label_source = batch_source
im_target, label_target = batch_target
def one_hot_encoding(y):
y_onehot = y.numpy()
y_onehot = (np.arange(6) == y_onehot[:,None]).astype(np.float32)
y_onehot = torch.from_numpy(y_onehot)
return y_onehot
label_source = one_hot_encoding(label_source)
label_target = one_hot_encoding(label_target)
im_source, label_source = im_source.cuda(), label_source.cuda(non_blocking=True)
im_target, label_target = im_target.cuda(), label_target.cuda(non_blocking=True)
fs1, feature_source, __, predict_prob_source = net.forward(im_source)
ft1, feature_target, __, predict_prob_target = net.forward(im_target)
p0 = discriminator_p.forward(fs1)
p1 = discriminator_p.forward(ft1)
p2 = torch.sum(p1, dim = -1)
# =========================rank the output of the multi-binary classifiers
try:
__,_,_,dptarget = discriminator_t.forward(ft1.detach())
r = torch.sort(dptarget[:,1].detach(),dim = 0)[1][30:]
feature_otherep = torch.index_select(ft1, 0, r.view(2))
_, _, __, predict_prob_otherep = cls.forward(feature_otherep)
w = torch.sort(p2.detach(),dim = 0)[1][30:]
h = torch.sort(p2.detach(),dim = 0)[1][0:2]
feature_otherep2 = torch.index_select(ft1, 0, w.view(2))
feature_otherep1 = torch.index_select(ft1, 0, h.view(2))
_,_,_,pred00 = discriminator_t.forward(feature_otherep2)
_,_,_,pred01 = discriminator_t.forward(feature_otherep1)
except:
continue
# =========================loss function
ce = CrossEntropyLoss(label_source, predict_prob_source)
d1 = BCELossForMultiClassification(label_source[:,0:5],p0)
d2 = CrossEntropyLoss(Variable(torch.from_numpy(np.concatenate((np.ones((2,1)), np.zeros((2,1))), axis = -1).astype('float32'))).cuda(),pred00)
d2 += CrossEntropyLoss(Variable(torch.from_numpy(np.concatenate((np.zeros((2,1)), np.ones((2,1))), axis = -1).astype('float32'))).cuda(),pred01)
with OptimizerManager([optimizer_cls, optimizer_discriminator_p, optimizer_discriminator_t]):
loss = ce + d1 +d2
loss.backward()
k += 1
log.step += 1
if log.step % 10 == 1:
counter = AccuracyCounter()
counter.addOntBatch(variable_to_numpy(predict_prob_source), variable_to_numpy(label_source))
acc_train = Variable(torch.from_numpy(np.asarray([counter.reportAccuracy()], dtype=np.float32))).cuda()
track_scalars(log, ['ce', 'acc_train', 'd1', 'd2'], globals())
if log.step % 100 == 0:
clear_output()
# =========================save the parameters of the known/unknown discriminator
torch.save(discriminator_t.state_dict(), 'discriminator_a_cifar_alexnet.pkl')