File size: 4,846 Bytes
381fcd6
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import torch
import torch.nn as nn
import torch.nn.functional as F
from attacks import create_attack
import numpy as np
from torch.autograd import Variable
from contextlib import contextmanager
if torch.cuda.is_available():
    device = torch.device('cuda')
else:
    device = torch.device('cpu')

class ctx_noparamgrad(object):
    def __init__(self, module):
        self.prev_grad_state = get_param_grad_state(module)
        self.module = module
        set_param_grad_off(module)

    def __enter__(self):
        pass

    def __exit__(self, *args):
        set_param_grad_state(self.module, self.prev_grad_state)
        return False


class ctx_eval(object):
    def __init__(self, module):
        self.prev_training_state = get_module_training_state(module)
        self.module = module
        set_module_training_off(module)

    def __enter__(self):
        pass

    def __exit__(self, *args):
        set_module_training_state(self.module, self.prev_training_state)
        return False


@contextmanager
def ctx_noparamgrad_and_eval(module):
    with ctx_noparamgrad(module) as a, ctx_eval(module) as b:
        yield (a, b)


def get_module_training_state(module):
    return {mod: mod.training for mod in module.modules()}


def set_module_training_state(module, training_state):
    for mod in module.modules():
        mod.training = training_state[mod]


def set_module_training_off(module):
    for mod in module.modules():
        mod.training = False


def get_param_grad_state(module):
    return {param: param.requires_grad for param in module.parameters()}


def set_param_grad_state(module, grad_state):
    for param in module.parameters():
        param.requires_grad = grad_state[param]


def set_param_grad_off(module):
    for param in module.parameters():
        param.requires_grad = False
class MadrysLoss(nn.Module):
    def __init__(self, step_size=0.007, epsilon=0.031, perturb_steps=10, beta=6.0,
                 distance='l_inf', cutmix=False, adjust_freeze=True, cutout=False,
                 cutout_length=16):
        super(MadrysLoss, self).__init__()
        self.step_size = step_size
        self.epsilon = epsilon
        self.perturb_steps = perturb_steps
        self.beta = beta
        self.distance = distance
        self.cross_entropy =  torch.nn.CrossEntropyLoss()
        self.adjust_freeze = adjust_freeze
        self.cutout = cutout
        self.cutout_length = cutout_length

    def forward(self, model, x_natural, labels): #optimizer
        model.eval()
        if self.adjust_freeze:
            for param in model.parameters():
                param.requires_grad = False

        # generate adversarial example
        x_adv = x_natural.detach() + self.step_size * torch.randn(x_natural.shape).to(device).detach()
        if self.distance == 'l_inf':
            adv_loss = 0
            for _ in range(self.perturb_steps):
                x_adv.requires_grad_()
                loss_ce = self.cross_entropy(model(x_adv), labels)
                grad = torch.autograd.grad(loss_ce, [x_adv])[0]
                x_adv = x_adv.detach() + self.step_size * torch.sign(grad.detach())
                x_adv = torch.min(torch.max(x_adv, x_natural - self.epsilon), x_natural + self.epsilon)
                x_adv = torch.clamp(x_adv, 0.0, 1.0)
        else:
            x_adv = torch.clamp(x_adv, 0.0, 1.0)

        x_adv = Variable(x_adv, requires_grad=False)

        if self.adjust_freeze:
            for param in model.parameters():
                param.requires_grad = True

        if self.cutout:
            batch_size = x_adv.shape[0]
            c, h, w = x_adv.shape[1], x_adv.shape[2], x_adv.shape[3]
            mask = torch.ones(batch_size, c, h, w).float()
            for j in range(batch_size):
                y = np.random.randint(h)
                x = np.random.randint(w)

                y1 = np.clip(y - self.cutout_length // 2, 0, h)
                y2 = np.clip(y + self.cutout_length // 2, 0, h)
                x1 = np.clip(x - self.cutout_length // 2, 0, w)
                x2 = np.clip(x + self.cutout_length // 2, 0, w)

                mask[j, :, y1: y2, x1: x2] = 0.0
            x_adv = x_adv * mask.to(device)

        model.train()
        # optimizer.zero_grad()


        logits = model(x_adv)
        loss = self.cross_entropy(logits, labels)

        return loss


def sat_loss(model, x, y,optimizer,step_size,epsilon,num_steps,attack_type,beta,criterion= torch.nn.CrossEntropyLoss()):
    """
    Adversarial training (Madry et al, 2017).
    """
    attack = create_attack(model, criterion, 'linf-pgd', epsilon, num_steps, step_size)  
    with ctx_noparamgrad_and_eval(model):
        x_adv, _ = attack.perturb(x, y)
    print(x_adv.shape)
    y_adv = y
    out = model(x_adv)
    loss = criterion(out, y_adv)
    
    return loss