/
teacher_noise.py
226 lines (195 loc) · 8.81 KB
/
teacher_noise.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
import numpy as np
import torch
import torch.nn as nn
from torch.autograd import Variable
import torch.optim
import torch.optim.lr_scheduler as lr_scheduler
import time
import os
import glob
import model
import backbone
from io_utils import model_dict, parse_args, get_resume_file, get_assigned_file
import torchvision
import torchvision.transforms as transforms
import attack
import matplotlib.pyplot as plt
import warnings
from TinyImagenet import *
import torchvision.models as models
warnings.filterwarnings("ignore")
SAVE_DIR = '/DATA1/puneet/interpretable/checkpoints'
def train(trainloader, model, optimization, start_epoch, stop_epoch, params,config):
if optimization == 'Adam':
optimizer = torch.optim.Adam(model.parameters())
else:
raise ValueError('Unknown optimization, please define by yourself')
loss_fn = nn.CrossEntropyLoss()
pgd = attack.AttackPGD(config)
for epoch in range(start_epoch,stop_epoch):
model.train()
print_freq = 50
avg_loss=0
correct ,total =0,0
for i, (x,y,_) in enumerate(trainloader):
noise = torch.zeros_like(x).uniform_(-8/255., 8/255.)
x = torch.clamp(x + noise,0.,1.)
x,y = x.cuda(), y.cuda()
optimizer.zero_grad()
scores,_ = model.forward(x)
predicted = torch.argmax(scores,1)
correct += (predicted==y).sum().item()
total += predicted.size(0)
loss = loss_fn(scores,y)
loss.backward()
optimizer.step()
avg_loss = avg_loss+loss.data.item()
if i % print_freq==0:
print('Epoch {:d} | Batch {:d}/{:d} | Loss {:f} | Train Acc {:f}'.format(epoch, i, len(trainloader), avg_loss/float(i+1),100.*correct/total ))
if not os.path.isdir(params.checkpoint_dir):
os.makedirs(params.checkpoint_dir)
if (epoch % params.save_freq==0) or (epoch==stop_epoch-1):
outfile = os.path.join(params.checkpoint_dir, '{:d}.tar'.format(epoch))
state_dict = {}
state_dict['epoch'] = epoch
state_dict['feature'] = model.feature.state_dict()
state_dict['classifier'] = model.classifier.state_dict()
torch.save(state_dict, outfile)
return model
def test(testloader, model,params,config):
pgd = attack.AttackPGD(config)
model.eval()
correct ,total =0,0
for i, (x,y) in enumerate(testloader):
x,y = x.cuda(), y.cuda()
# print(y)
x = pgd.attack(model,x,y)
scores,_ = model.forward(x)
predicted = torch.argmax(scores,1)
correct += (predicted==y).sum().item()
total += predicted.size(0)
print('Accuracy {:f}'.format(100.*correct/total))
if __name__=='__main__':
np.random.seed(10)
params = parse_args()
print(params)
if params.dataset == 'cifar10':
params.num_classes = 10
transform_train = transforms.Compose([
transforms.RandomCrop(32, padding=4),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
# transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)),
])
transform_test = transforms.Compose([
transforms.ToTensor(),
# transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)),
])
trainset = torchvision.datasets.CIFAR10(root='./../root_cifar', train=True, download=True, transform=transform_train)
trainloader = torch.utils.data.DataLoader(trainset, batch_size=params.bs, shuffle=True, num_workers=12)
testset = torchvision.datasets.CIFAR10(root='./../root_cifar', train=False, download=True, transform=transform_test)
testloader = torch.utils.data.DataLoader(testset, batch_size=params.bs, shuffle=True, num_workers=12)
config = {
'epsilon': 2.0 / 255,
'num_steps': 5,
'step_size': 2.0 / 255,
'random_start': True,
'loss_func': 'xent',
}
elif params.dataset == 'cifar100':
params.num_classes = 100
transform_train = transforms.Compose([
transforms.RandomCrop(32, padding=4),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
# transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)),
])
transform_test = transforms.Compose([
transforms.ToTensor(),
# transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)),
])
trainset = torchvision.datasets.CIFAR100(root='./../root_cifar100', train=True, download=True, transform=transform_train)
trainloader = torch.utils.data.DataLoader(trainset, batch_size=params.bs, shuffle=True, num_workers=12)
testset = torchvision.datasets.CIFAR100(root='./../root_cifar100', train=False, download=True, transform=transform_test)
testloader = torch.utils.data.DataLoader(testset, batch_size=params.bs, shuffle=True, num_workers=12)
config = {
'epsilon': 4.0 / 255,
'num_steps': 5,
'step_size': 2.0 / 255,
'random_start': True,
'loss_func': 'xent',
}
elif params.dataset == 'svhn':
params.num_classes = 10
transform_train = transforms.Compose([
transforms.RandomCrop(32, padding=4),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
# transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)),
])
transform_test = transforms.Compose([
transforms.ToTensor(),
# transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)),
])
trainset = torchvision.datasets.SVHN(root='./../root_shvn', split='train', download=True,transform=transform_train)
trainloader = torch.utils.data.DataLoader(trainset, batch_size=params.bs, shuffle=True, num_workers=2)
testset = torchvision.datasets.SVHN(root='./../root_shvn', split='test', download=True,transform=transform_test)
testloader = torch.utils.data.DataLoader(testset, batch_size=params.bs, shuffle=False, num_workers=2)
config = {
'epsilon': 8.0 / 255,
'num_steps': 5,
'step_size': 2.0 / 255,
'random_start': True,
'loss_func': 'xent',
}
elif params.dataset == 'tiny-img':
params.num_classes = 200
transform_train = transforms.Compose([
transforms.Resize(32),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
# transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)),
])
transform_test = transforms.Compose([
transforms.Resize(32),
transforms.ToTensor(),
# transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)),
])
# trainset = torchvision.datasets.ImageFolder('/DATA1/puneet/tiny-imagenet-200/train', transform_train)
trainset = TinyImageNet('/DATA1/tiny-imagenet-200', 'train', transform=transform_train)
trainloader = torch.utils.data.DataLoader(trainset, batch_size=params.bs, shuffle=True, num_workers=2)
# testset = torchvision.datasets.ImageFolder('/DATA1/puneet/tiny-imagenet-200/val', transform_test)
testset = TinyImageNet('/DATA1/tiny-imagenet-200', 'val',transform=transform_test)
testloader = torch.utils.data.DataLoader(testset, batch_size=params.bs, shuffle=False, num_workers=2)
config = {
'epsilon': 3.0 / 255,
'num_steps': 5,
'step_size': 2.0 / 255,
'random_start': True,
'loss_func': 'xent',
}
model = model.Model(net=params.model,num_classes= params.num_classes)
# model = nn.DataParallel(model,device_ids=[0,1])
optimization = 'Adam'
params.checkpoint_dir = '%s/%s/teacher/%s_%s' %( SAVE_DIR,params.dataset, params.model, 'noise')
print('checkpoints dir',params.checkpoint_dir)
if not os.path.isdir(params.checkpoint_dir):
os.makedirs(params.checkpoint_dir)
start_epoch = params.start_epoch
stop_epoch = params.stop_epoch
if params.resume:
if params.iter !=-1:
resume_file = get_assigned_file(params.checkpoint_dir,params.iter)
else:
resume_file = get_resume_file(params.checkpoint_dir)
if resume_file is not None:
print('Resume file is: ', resume_file)
tmp = torch.load(resume_file)
start_epoch = tmp['epoch']+1
model.classifier.load_state_dict(tmp['classifier'])
model.feature.load_state_dict(tmp['feature'])
model.cuda()
if params.test is None:
model = train(trainloader, model, optimization, start_epoch, stop_epoch, params,config)
else:
test(testloader, model, params,config)