forked from ankurhanda/gvnn
-
Notifications
You must be signed in to change notification settings - Fork 0
/
AffineTransformMatrixGenerator.lua
244 lines (178 loc) · 7.95 KB
/
AffineTransformMatrixGenerator.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
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
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
local ATMG, parent = torch.class('nn.AffineTransformMatrixGenerator', 'nn.Module')
--[[
AffineTransformMatrixGenerator(useRotation, useScale, useTranslation) :
AffineTransformMatrixGenerator:updateOutput(transformParams)
AffineTransformMatrixGenerator:updateGradInput(transformParams, gradParams)
This module can be used in between the localisation network (that outputs the
parameters of the transformation) and the AffineGridGeneratorBHWD (that expects
an affine transform matrix as input).
The goal is to be able to use only specific transformations or a combination of them.
If no specific transformation is specified, it uses a fully parametrized
linear transformation and thus expects 6 parameters as input. In this case
the module is equivalent to nn.View(2,3):setNumInputDims(2).
Any combination of the 3 transformations (rotation, scale and/or translation)
can be used. The transform parameters must be supplied in the following order:
rotation (1 param), scale (1 param) then translation (2 params).
Example:
AffineTransformMatrixGenerator(true,false,true) expects as input a tensor of
if size (B, 3) containing (rotationAngle, translationX, translationY).
]]
function ATMG:__init(useRotation, useScale, useTranslation)
parent.__init(self)
-- if no specific transformation, use fully parametrized version
self.fullMode = not(useRotation or useScale or useTranslation)
if not self.fullMode then
self.useRotation = useRotation
self.useScale = useScale
self.useTranslation = useTranslation
end
end
function ATMG:check(input)
if self.fullMode then
assert(input:size(2)==6, 'Expected 6 parameters, got ' .. input:size(2))
else
local numberParameters = 0
if self.useRotation then
numberParameters = numberParameters + 1
end
if self.useScale then
numberParameters = numberParameters + 1
end
if self.useTranslation then
numberParameters = numberParameters + 2
end
assert(input:size(2)==numberParameters, 'Expected '..numberParameters..
' parameters, got ' .. input:size(2))
end
end
local function addOuterDim(t)
local sizes = t:size()
local newsizes = torch.LongStorage(sizes:size()+1)
newsizes[1]=1
for i=1,sizes:size() do
newsizes[i+1]=sizes[i]
end
return t:view(newsizes)
end
function ATMG:updateOutput(_tranformParams)
local transformParams
if _tranformParams:nDimension()==1 then
transformParams = addOuterDim(_tranformParams)
else
transformParams = _tranformParams
end
self:check(transformParams)
local batchSize = transformParams:size(1)
if self.fullMode then
self.output = transformParams:view(batchSize, 2, 3)
else
local completeTransformation = torch.zeros(batchSize,3,3):typeAs(transformParams)
completeTransformation:select(3,1):select(2,1):add(1)
completeTransformation:select(3,2):select(2,2):add(1)
completeTransformation:select(3,3):select(2,3):add(1)
local transformationBuffer = torch.Tensor(batchSize,3,3):typeAs(transformParams)
local paramIndex = 1
if self.useRotation then
local alphas = transformParams:select(2, paramIndex)
paramIndex = paramIndex + 1
transformationBuffer:zero()
transformationBuffer:select(3,3):select(2,3):add(1)
local cosines = torch.cos(alphas)
local sinuses = torch.sin(alphas)
transformationBuffer:select(3,1):select(2,1):copy(cosines)
transformationBuffer:select(3,2):select(2,2):copy(cosines)
transformationBuffer:select(3,1):select(2,2):copy(sinuses)
transformationBuffer:select(3,2):select(2,1):copy(-sinuses)
completeTransformation = torch.bmm(completeTransformation, transformationBuffer)
end
self.rotationOutput = completeTransformation:narrow(2,1,2):narrow(3,1,2):clone()
if self.useScale then
local scaleFactors = transformParams:select(2,paramIndex)
paramIndex = paramIndex + 1
transformationBuffer:zero()
transformationBuffer:select(3,1):select(2,1):copy(scaleFactors)
transformationBuffer:select(3,2):select(2,2):copy(scaleFactors)
transformationBuffer:select(3,3):select(2,3):add(1)
completeTransformation = torch.bmm(completeTransformation, transformationBuffer)
end
self.scaleOutput = completeTransformation:narrow(2,1,2):narrow(3,1,2):clone()
if self.useTranslation then
local txs = transformParams:select(2,paramIndex)
local tys = transformParams:select(2,paramIndex+1)
transformationBuffer:zero()
transformationBuffer:select(3,1):select(2,1):add(1)
transformationBuffer:select(3,2):select(2,2):add(1)
transformationBuffer:select(3,3):select(2,3):add(1)
transformationBuffer:select(3,3):select(2,1):copy(txs)
transformationBuffer:select(3,3):select(2,2):copy(tys)
completeTransformation = torch.bmm(completeTransformation, transformationBuffer)
end
self.output=completeTransformation:narrow(2,1,2)
end
if _tranformParams:nDimension()==1 then
self.output = self.output:select(1,1)
end
return self.output
end
function ATMG:updateGradInput(_tranformParams, _gradParams)
local transformParams, gradParams
if _tranformParams:nDimension()==1 then
transformParams = addOuterDim(_tranformParams)
gradParams = addOuterDim(_gradParams):clone()
else
transformParams = _tranformParams
gradParams = _gradParams:clone()
end
-- print (gradParams:size())
local batchSize = transformParams:size(1)
if self.fullMode then
self.gradInput = gradParams:view(batchSize, 6)
else
local paramIndex = transformParams:size(2)
self.gradInput:resizeAs(transformParams)
if self.useTranslation then
local gradInputTranslationParams = self.gradInput:narrow(2,paramIndex-1,2)
local tParams = torch.Tensor(batchSize, 1, 2):typeAs(transformParams)
tParams:select(3,1):copy(transformParams:select(2,paramIndex-1))
tParams:select(3,2):copy(transformParams:select(2,paramIndex))
paramIndex = paramIndex-2
local selectedOutput = self.scaleOutput
local selectedGradParams = gradParams:narrow(2,1,2):narrow(3,3,1):transpose(2,3)
gradInputTranslationParams:copy(torch.bmm(selectedGradParams, selectedOutput))
print ( gradParams )
print ( selectedOutput )
print ( selectedGradParams )
print ( self.output )
print ( gradInputTranslationParams )
local gradientCorrection = torch.bmm(selectedGradParams:transpose(2,3), tParams)
gradParams:narrow(3,1,2):narrow(2,1,2):add(1,gradientCorrection)
end
if self.useScale then
local gradInputScaleparams = self.gradInput:narrow(2,paramIndex,1)
local sParams = transformParams:select(2,paramIndex)
paramIndex = paramIndex-1
local selectedOutput = self.rotationOutput
local selectedGradParams = gradParams:narrow(2,1,2):narrow(3,1,2)
gradInputScaleparams:copy(torch.cmul(selectedOutput, selectedGradParams):sum(2):sum(3))
gradParams:select(3,1):select(2,1):cmul(sParams)
gradParams:select(3,2):select(2,1):cmul(sParams)
gradParams:select(3,1):select(2,2):cmul(sParams)
gradParams:select(3,2):select(2,2):cmul(sParams)
end
if self.useRotation then
local gradInputRotationParams = self.gradInput:narrow(2,paramIndex,1)
local rParams = transformParams:select(2,paramIndex)
local rotationDerivative = torch.zeros(batchSize, 2, 2):typeAs(rParams)
torch.sin(rotationDerivative:select(3,1):select(2,1),-rParams)
torch.sin(rotationDerivative:select(3,2):select(2,2),-rParams)
torch.cos(rotationDerivative:select(3,1):select(2,2),rParams)
torch.cos(rotationDerivative:select(3,2):select(2,1),rParams):mul(-1)
local selectedGradParams = gradParams:narrow(2,1,2):narrow(3,1,2)
gradInputRotationParams:copy(torch.cmul(rotationDerivative,selectedGradParams):sum(2):sum(3))
end
end
if _tranformParams:nDimension()==1 then
self.gradInput = self.gradInput:select(1,1)
end
return self.gradInput
end