-
Notifications
You must be signed in to change notification settings - Fork 5
/
Reparametrize.lua
82 lines (60 loc) · 2.29 KB
/
Reparametrize.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
-- Based on JoinTable module
require 'nn'
local Reparametrize, parent = torch.class('nn.Reparametrize', 'nn.Module')
function Reparametrize:__init(dimension)
parent.__init(self)
self.size = torch.LongStorage()
self.dimension = dimension
self.gradInput = {}
end
function Reparametrize:updateOutput(input)
--Different eps for whole batch, or one and broadcast?
self.eps = torch.randn(input[2]:size(1),self.dimension)
self.output = torch.mul(input[2],0.5):exp():cmul(self.eps)
-- Add the mean_
self.output:add(input[1])
return self.output
end
function Reparametrize:updateGradInput(input, gradOutput)
-- Derivative with respect to mean is 1
self.gradInput[1] = gradOutput:clone()
--Not sure if this gradient is right
self.gradInput[2] = torch.mul(input[2],0.5):exp():mul(0.5):cmul(self.eps)
self.gradInput[2]:cmul(gradOutput)
return self.gradInput
end
-- -- Based on JoinTable module
-- require 'nn'
-- local Reparametrize, parent = torch.class('nn.Reparametrize', 'nn.Module')
-- function Reparametrize:__init(dimension)
-- parent.__init(self)
-- self.size = torch.LongStorage()
-- self.dimension = dimension
-- self.gradInput = {}
-- end
-- function Reparametrize:updateOutput(input)
-- self.eps = torch.randn(input[2]:size(1),self.dimension)
-- if torch.typename(input[1]) == 'torch.CudaTensor' then
-- self.eps = self.eps:cuda()
-- self.output = torch.CudaTensor():resizeAs(input[2]):fill(0.5)
-- else
-- self.output = torch.Tensor():resizeAs(input[2]):fill(0.5)
-- end
-- self.output:cmul(input[2]):exp():cmul(self.eps)
-- -- Add the mean
-- self.output:add(input[1])
-- return self.output
-- end
-- function Reparametrize:updateGradInput(input, gradOutput)
-- -- Derivative with respect to mean is 1
-- self.gradInput[1] = gradOutput:clone()
-- --test gradient with Jacobian
-- if torch.typename(input[1]) == 'torch.CudaTensor' then
-- self.gradInput[2] = torch.CudaTensor():resizeAs(input[2]):fill(0.5)
-- else
-- self.gradInput[2] = torch.Tensor():resizeAs(input[2]):fill(0.5)
-- end
-- self.gradInput[2]:cmul(input[2]):exp():mul(0.5):cmul(self.eps)
-- self.gradInput[2]:cmul(gradOutput)
-- return self.gradInput
-- end