-
Notifications
You must be signed in to change notification settings - Fork 0
/
losses.py
123 lines (105 loc) · 4.35 KB
/
losses.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
import math
import torch
import torch.nn as nn
import torch.nn.functional as F
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.autograd import Variable
import torch.optim as optim
def trades_loss(model,
x_natural,
y,
optimizer,
step_size=0.003,
epsilon=0.031,
perturb_steps=10,
beta=1.0,
distance='l_inf'):
# define KL-loss
criterion_kl = nn.KLDivLoss(size_average=False)
model.eval()
batch_size = len(x_natural)
# generate adversarial example
x_adv = x_natural.detach() + 0.001 * torch.randn(x_natural.shape).cuda().detach()
if distance == 'l_inf':
for _ in range(perturb_steps):
x_adv.requires_grad_()
with torch.enable_grad():
loss_kl = criterion_kl(F.log_softmax(model(x_adv), dim=1),
F.softmax(model(x_natural), dim=1))
grad = torch.autograd.grad(loss_kl, [x_adv])[0]
x_adv = x_adv.detach() + step_size * torch.sign(grad.detach())
x_adv = torch.min(torch.max(x_adv, x_natural - epsilon), x_natural + epsilon)
x_adv = torch.clamp(x_adv, 0.0, 1.0)
elif distance == 'l_2':
delta = 0.001 * torch.randn(x_natural.shape).cuda().detach()
delta = Variable(delta.data, requires_grad=True)
# Setup optimizers
optimizer_delta = optim.SGD([delta], lr=epsilon / perturb_steps * 2)
for _ in range(perturb_steps):
adv = x_natural + delta
# optimize
optimizer_delta.zero_grad()
with torch.enable_grad():
loss = (-1) * criterion_kl(F.log_softmax(model(adv), dim=1),
F.softmax(model(x_natural), dim=1))
loss.backward()
# renorming gradient
grad_norms = delta.grad.view(batch_size, -1).norm(p=2, dim=1)
delta.grad.div_(grad_norms.view(-1, 1, 1, 1))
# avoid nan or inf if gradient is 0
if (grad_norms == 0).any():
delta.grad[grad_norms == 0] = torch.randn_like(delta.grad[grad_norms == 0])
optimizer_delta.step()
# projection
delta.data.add_(x_natural)
delta.data.clamp_(0, 1).sub_(x_natural)
delta.data.renorm_(p=2, dim=0, maxnorm=epsilon)
x_adv = Variable(x_natural + delta, requires_grad=False)
else:
x_adv = torch.clamp(x_adv, 0.0, 1.0)
model.train()
x_adv = Variable(torch.clamp(x_adv, 0.0, 1.0), requires_grad=False)
# zero gradient
optimizer.zero_grad()
# calculate robust loss
logits = model(x_natural)
loss_natural = F.cross_entropy(logits, y)
loss_robust = (1.0 / batch_size) * criterion_kl(F.log_softmax(model(x_adv), dim=1),
F.softmax(model(x_natural), dim=1))
loss = loss_natural + beta * loss_robust
return loss
def focal_loss(input_values, gamma):
"""Computes the focal loss"""
p = torch.exp(-input_values)
loss = (1 - p) ** gamma * input_values
return loss.mean()
class FocalLoss(nn.Module):
def __init__(self, weight=None, gamma=0.):
super(FocalLoss, self).__init__()
assert gamma >= 0
self.gamma = gamma
self.weight = weight
def forward(self, input, target):
return focal_loss(F.cross_entropy(input, target, reduction='none', weight=self.weight), self.gamma)
class LDAMLoss(nn.Module):
def __init__(self, cls_num_list, max_m=0.5, weight=None, s=30):
super(LDAMLoss, self).__init__()
m_list = 1.0 / np.sqrt(np.sqrt(cls_num_list))
m_list = m_list * (max_m / np.max(m_list))
m_list = torch.cuda.FloatTensor(m_list)
self.m_list = m_list
assert s > 0
self.s = s
self.weight = weight
def forward(self, x, target):
index = torch.zeros_like(x, dtype=torch.uint8)
index.scatter_(1, target.data.view(-1, 1), 1)
index_float = index.type(torch.cuda.FloatTensor)
batch_m = torch.matmul(self.m_list[None, :], index_float.transpose(0, 1))
batch_m = batch_m.view((-1, 1))
x_m = x - batch_m
output = torch.where(index, x_m, x)
return F.cross_entropy(self.s * output, target, weight=self.weight)