-
Notifications
You must be signed in to change notification settings - Fork 7
/
MBCECriterion_balance.lua
68 lines (53 loc) · 1.69 KB
/
MBCECriterion_balance.lua
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
require 'nn'
require 'torch'
local MBCECriterion, parent = torch.class('nn.MBCECriterion', 'nn.Criterion')
function MBCECriterion:__init()
self.nAttr = 18
parent.__init(self)
self.input = torch.Tensor()
self.target = torch.Tensor()
self.gradInput = torch.Tensor()
self.output = torch.Tensor()
self.w = {}
for i = 1, self.nAttr do
table.insert(self.w, 1)
end
self.w[18] = 10 -- young
self.w[15] = 10 -- no_beard
self.weights = torch.Tensor{self.w}
end
function MBCECriterion:updateOutput(input, target)
self.input:resizeAs(input):copy(input)
self.target:resizeAs(target):copy(target)
local input_log1 = torch.log(self.input)
local loss1 = torch.cmul(input_log1, self.target)
local input_log2 = torch.log(1-self.input)
local loss2 = torch.cmul(input_log2, (1-self.target))
local loss = -loss1 - loss2
--local output = torch.mean(torch.mean(loss,2),1)
--self.output = output:squeeze()
self.output:resizeAs(loss):copy(loss)
return self.output
end
function MBCECriterion:updateGradInput(input, target)
local num = input:size(1)
local num_attr = input:size(2)
local weights = torch.expand(self.weights, target:size())
local nominator = input:cmul(weights) + input:cmul(target):cmul(1-weights) - target
local denominator = torch.cmul(input, 1-input)
local mask = torch.lt(denominator, 1e-5)
denominator[mask] = 1e-3
local grad = nominator:cdiv(denominator)
grad:div(num)
for i = 1, num_attr do
if self.w[i] ~= 1 then
grad:select(2, i):clamp(-300, 300)
else
grad:select(2, i):clamp(-30, 30)
end
end
self.gradInput:resizeAs(input):copy(grad)
-- self.gradInput:resizeAs(input):copy(nominator/num)
-- print(self.gradInput:norm())
return self.gradInput
end