-
Notifications
You must be signed in to change notification settings - Fork 52
/
UtilsNetwork.lua
123 lines (96 loc) · 3.97 KB
/
UtilsNetwork.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
--require 'Probe'
function GaussianSampler(M, D)
local q = - nn.Identity()
local eps = - nn.Identity()
local mean = q - nn.SelectTable(1)
- nn.Replicate(M)
local logVar = q - nn.SelectTable(2)
local std = logVar - nn.MulConstant(0.5)
- nn.Exp()
- nn.Replicate(M)
local noise = {std, eps} - nn.CMulTable()
local sample = {mean, noise}
- nn.CAddTable() -- [MxNxD]
- nn.View(-1, D) -- [(MxN)xD]
return nn.gModule({q, eps}, {sample})
end
-- TEST PASS
function KLDivergence(D, M)
-- KL = 1/2( logvar2 - logvar1 + (var1 + (m1-m2)^2)/var2 - 1 )
local mean1_in = - nn.Identity()
local logVar1_in = - nn.Identity()
local mean2_in = - nn.Identity() -- [(MxN)xD]
local logVar2_in = - nn.Identity() -- [(MxN)xD]
local mean1 = mean1_in - nn.Replicate(M)
local logVar1 = logVar1_in - nn.Replicate(M)
local mean2 = mean2_in - nn.View(M, -1, D)
local logVar2 = logVar2_in - nn.View(M, -1, D)
local var1 = logVar1 - nn.Exp()
local var2 = logVar2 - nn.Exp()
local dm2 = {mean1, mean2}
- nn.CSubTable()
- nn.Power(2)
local dm2_v1 = {dm2, var1} - nn.CAddTable()
local dm2_v1_v2 = {dm2_v1, var2} - nn.CDivTable() - nn.AddConstant(-1)
local total = {dm2_v1_v2, logVar2} - nn.CAddTable()
local totals = {total, logVar1}
- nn.CSubTable()
- nn.MulConstant(0.5) -- [MxNxD]
- nn.Sum(1)
- nn.MulConstant(1/M)
- nn.View(-1, D, 1)
return nn.gModule({mean1_in, logVar1_in, mean2_in, logVar2_in}, {totals})
end
function KL_Table(K, D, M)
local KL_table = nn.ConcatTable()
for k=1, K do
local mean = - nn.Identity() -- [NxD]
local logVar = - nn.Identity() -- [NxD]
local mean_Mixture = - nn.Identity() -- {[NxD]}k
local logVar_Mixture = - nn.Identity() -- {[NxD]}k
local meanK = mean_Mixture - nn.SelectTable(k)
local logVarK = logVar_Mixture - nn.SelectTable(k)
local KL = {mean, logVar, meanK, logVarK} - KLDivergence(D, M)
local KL_module = nn.gModule({mean, logVar, mean_Mixture, logVar_Mixture}, {KL})
KL_table:add(KL_module)
end
return KL_table
end
function ExpectedKLDivergence(K, D, M)
local q_z = - nn.Identity() -- [NxK]
local mean = - nn.Identity() -- [NxD]
local logVar = - nn.Identity() -- [NxD]
local mean_Mixture = - nn.Identity() -- {[NxD]}k
local logVar_Mixture = - nn.Identity() -- {[NxD]}k
local KL_List = {mean, logVar, mean_Mixture, logVar_Mixture}
- KL_Table(K, D, M) -- {[NxDx1]}k
- nn.JoinTable(3) -- [NxDxK]
local weighted_KL = {KL_List, q_z}
- nn.MV() -- [NxDxK]x[NxK] = [NxD]
return nn.gModule({q_z, mean,logVar, mean_Mixture, logVar_Mixture},{weighted_KL})
end
require 'layers/GaussianLogLikelihood'
function Likelihood(K, D, M)
local x_sample = - nn.Identity() -- [(MxN)xD]
local mean = - nn.Identity() -- {[(MxN)xD]}k
local logVar = - nn.Identity() -- {[(MxN)xD]}k
local llh_table = nn.ConcatTable()
for k =1, K do
local x = - nn.Identity()
local mean_k_in = - nn.Identity()
local logVar_k_in = - nn.Identity()
local mean_k = mean_k_in
- nn.SelectTable(k)
local logVar_k = logVar_k_in
- nn.SelectTable(k)
local llh = {x, mean_k, logVar_k}
- nn.GaussianLogLikelihood()
local llh_module = nn.gModule({x, mean_k_in, logVar_k_in}, {llh})
llh_table:add(llh_module)
end
local out = {x_sample, mean, logVar}
- llh_table -- {[MxN,1]}k
- nn.JoinTable(2) -- [MxN,K] -- log unNorm P
- nn.SoftMax()
return nn.gModule({x_sample, mean, logVar},{out})
end