-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathGASSOM_Online.m
189 lines (142 loc) · 6.19 KB
/
GASSOM_Online.m
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
classdef GASSOM_Online < handle
% GASSOM_ONLINE This is the GASSOM oniline algorithm with winner selectionwhich assumes
% slowness in the data. Please refer the example for usage.
% specific queries, tnc<at>connect<dot>ust<dot>hk
%This could be run online, (i.e, X is one d-dimension column vector compricing the input, or things could be made faster by combining
%the inputs into mini batches of ~10 samples, X~[dxN])
properties
dim_patch_single;
topo_subspace;
max_iter;
length_basis;
dim_patch;
n_subspace;
segment_length;
n_basis;
size_subspace;
alpha_A;alpha_C;
sigma_A;sigma_C;
sigmaTrans;
alphaTrans;
bases;
transProb;
nodeProb;
winCoef;
winError;
Proj;
resi;
coef;
iter;
sigma_n;
sigma_w;
updatecount;
winners;
tconst;
tconst_n;
winnerTrack;
h_plot;
end
methods
function obj = GASSOM_Online(PARAM)
obj.dim_patch_single = PARAM{1};
obj.topo_subspace = PARAM{2};
obj.max_iter = PARAM{3};
%default
obj.n_subspace = prod(obj.topo_subspace);
obj.segment_length = prod(obj.dim_patch_single);
obj.dim_patch = [obj.dim_patch_single 2];
obj.length_basis = prod(obj.dim_patch_single);
obj.size_subspace = 2;
obj.n_basis = obj.size_subspace * obj.n_subspace;
obj.alpha_A = 0.08;
obj.alpha_C = 1e-4;
obj.tconst = 8000*10;
obj.sigma_A = 2;
obj.tconst_n = 8000;
obj.sigma_C = 0.5;
obj.sigmaTrans = 1.25;
obj.alphaTrans = 0.4;
obj.updatecount = 1;
obj.sigma_n = 0.2;
obj.sigma_w = 2;
%initialize
%random initial bases
A =randn(obj.length_basis, obj.size_subspace, obj.n_subspace);
A = orthonormalize_subspace (A);
obj.bases{1}= squeeze(A(:,1,:)); obj.bases{2}= squeeze(A(:,2,:));
obj.transProb = genTransProbG(obj.topo_subspace,obj.sigmaTrans, obj.alphaTrans,0);
np = rand(obj.n_subspace,1);
obj.nodeProb = bsxfun(@rdivide,np,sum(np));
obj.iter=1;
%init visualization
obj.h_plot = cell(1,2);
end
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% Encode
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
function [winners] = assomEncode(this,X)
[~ ,batch_size] = size(X);
this.coef{1} = this.bases{1}'*X; %[n_subspace batch_size]
this.coef{2} = this.bases{2}'*X;
this.Proj = this.coef{1}.^2 + this.coef{2}.^2; %P[n_subspace,batch_size]
Perr = ones(size(this.Proj))-this.Proj;
emissProb=exp(-this.Proj/(2*this.sigma_w^2)).*exp(-Perr/(2*this.sigma_n^2));
nodeprobTmp =zeros(this.n_subspace,batch_size);
for i=1:batch_size
nodeprobTmp(:,i) = (this.transProb'*this.nodeProb).* emissProb(:,i);
this.nodeProb = nodeprobTmp(:,i)./sum( nodeprobTmp(:,i));
end
[~,this.winners] = max(nodeprobTmp);
winners = this.winners;
end
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% updateBasis
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
function updateBasis(this,X)
alpha = (this.alpha_A*exp(-this.iter/this.tconst)+this.alpha_C);
sigma_h = (this.sigma_A*exp(-this.iter/this.tconst)+this.sigma_C);
[~, batch_size] = size(X);
[cj,ci] = ind2sub(this.topo_subspace,this.winners);
k = 1:this.n_subspace;
[kj,ki] = ind2sub(this.topo_subspace,k);
kj = repmat(kj',[1,batch_size] );
ki = repmat(ki',[1,batch_size] );
cj = repmat(cj, [this.n_subspace,1]);
ci = repmat(ci, [this.n_subspace,1]);
func_h = exp((-(ki-ci).^2-(kj-cj).^2)/(2*(sigma_h)^2)); %gaussian [n_subspace,batchsize]
n_const = 1./(sqrt(this.Proj)+eps);
weights = func_h.*n_const;
w_c{1} =weights.*this.coef{1};
w_c{2} =weights.*this.coef{2};
winput{1} = X*w_c{1}';
winput{2} = X*w_c{2}';
diff{1} = winput{1}-bsxfun(@times,this.bases{1},sum(w_c{1}.*this.coef{1},2)')-bsxfun(@times,this.bases{2},sum(w_c{1}.*this.coef{2},2)');
diff{2} = winput{2}-bsxfun(@times,this.bases{1},sum(w_c{2}.*this.coef{1},2)')-bsxfun(@times,this.bases{2},sum(w_c{2}.*this.coef{2},2)');
Bases{1} = this.bases{1} +alpha*diff{1};
Bases{2} = this.bases{2} +alpha*diff{2};
this.bases{1} = bsxfun(@rdivide, Bases{1}, sqrt(sum(Bases{1}.^2)));
Bases{2} = Bases{2} - bsxfun(@times,this.bases{1}, sum(this.bases{1}.*Bases{2}));
this.bases{2} = bsxfun(@rdivide, Bases{2}, sqrt(sum(Bases{2}.^2)));
this.iter = this.iter+1;
end
function visualizeBases(this)
A1 =(this.bases{1});
A2 =(this.bases{2});
[this.length_basis ,this.n_subspace] = size(A1);
A =zeros([this.length_basis, 2, this.n_subspace]);
for i=1:this.n_subspace
A(:,1,i) = A1(:,i);
A(:,2,i) = A2(:,i);
end
for h = 1:2
if(isempty(this.h_plot{h}))
this.h_plot{h} = plot_RF(squeeze(A(:,h,:)), ...
this.dim_patch_single, this.topo_subspace, '','', 1, 0, 1);
else
plot_RF(squeeze(A(:,h,:)), ...
this.dim_patch_single, this.topo_subspace, '','', 1, 0, 1,this.h_plot{h});
end
end
end
end
end