-
Notifications
You must be signed in to change notification settings - Fork 966
/
ClassSimplexCriterion.lua
118 lines (99 loc) · 3.73 KB
/
ClassSimplexCriterion.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
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
local ClassSimplexCriterion, parent
= torch.class('nn.ClassSimplexCriterion', 'nn.MSECriterion')
--[[
This file implements a criterion for multi-class classification.
It learns an embedding per class, where each class' embedding
is a point on an (N-1)-dimensional simplex, where N is
the number of classes.
For example usage of this class, look at doc/criterion.md
Reference: http://arxiv.org/abs/1506.08230
]]--
--[[
function regsplex(n):
regsplex returns the coordinates of the vertices of a
regular simplex centered at the origin.
The Euclidean norms of the vectors specifying the vertices are
all equal to 1. The input n is the dimension of the vectors;
the simplex has n+1 vertices.
input:
n -- dimension of the vectors specifying the vertices of the simplex
output:
a -- tensor dimensioned (n+1,n) whose rows are
vectors specifying the vertices
reference:
http://en.wikipedia.org/wiki/Simplex#Cartesian_coordinates_for_regular_n-dimensional_simplex_in_Rn
--]]
local function regsplex(n)
local a = torch.zeros(n+1,n)
for k = 1,n do
-- determine the last nonzero entry in the vector for the k-th vertex
if k==1 then a[k][k] = 1 end
if k>1 then a[k][k] = math.sqrt( 1 - a[{ {k},{1,k-1} }]:norm()^2 ) end
-- fill the k-th coordinates for the vectors of the remaining vertices
local c = (a[k][k]^2 - 1 - 1/n) / a[k][k]
a[{ {k+1,n+1},{k} }]:fill(c)
end
return a
end
function ClassSimplexCriterion:__init(nClasses)
parent.__init(self)
assert(nClasses and nClasses > 1 and nClasses == (nClasses -(nClasses % 1)),
"Required positive integer argument nClasses > 1")
self.nClasses = nClasses
-- embedding the simplex in a space of dimension strictly greater than
-- the minimum possible (nClasses-1) is critical for effective training.
local simp = regsplex(nClasses - 1)
self.simplex = torch.cat(simp,
torch.zeros(simp:size(1), nClasses -simp:size(2)),
2)
self._target = torch.Tensor(nClasses)
end
-- handle target being both 1D tensor, and
-- target being 2D tensor (2D tensor means don't do anything)
local function transformTarget(self, target)
if torch.type(target) == 'number' then
self._target:resize(self.nClasses)
self._target:copy(self.simplex[target])
elseif torch.isTensor(target) then
assert(target:dim() == 1, '1D tensors only!')
local nSamples = target:size(1)
self._target:resize(nSamples, self.nClasses)
for i=1,nSamples do
self._target[i]:copy(self.simplex[target[i]])
end
end
end
function ClassSimplexCriterion:updateOutput(input, target)
transformTarget(self, target)
assert(input:nElement() == self._target:nElement())
self.output_tensor = self.output_tensor or input.new(1)
input.THNN.MSECriterion_updateOutput(
input:cdata(),
self._target:cdata(),
self.output_tensor:cdata(),
self.sizeAverage
)
self.output = self.output_tensor[1]
return self.output
end
function ClassSimplexCriterion:updateGradInput(input, target)
assert(input:nElement() == self._target:nElement())
input.THNN.MSECriterion_updateGradInput(
input:cdata(),
self._target:cdata(),
self.gradInput:cdata(),
self.sizeAverage
)
return self.gradInput
end
function ClassSimplexCriterion:getPredictions(input)
if input:dim() == 1 then
input = input:view(1, -1)
end
return torch.mm(input, self.simplex:t())
end
function ClassSimplexCriterion:getTopPrediction(input)
local prod = self:getPredictions(input)
local _, maxs = prod:max(prod:nDimension())
return maxs:view(-1)
end