-
Notifications
You must be signed in to change notification settings - Fork 1
/
mitliag_pm.m
130 lines (114 loc) · 3.55 KB
/
mitliag_pm.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
function [T, ErrFro, U, YrHat, t] = mitliag_pm(Y, r, blkSize, floor_mul, no_err)
%%MITLIAG_PM This is an implementation of the Power Method as described in
% "Memory Limited, Streaming PCA" (https://arxiv.org/pdf/1307.0032.pdf)
%
% Author: Andreas Grammenos ([email protected])
%
% Last touched date: 30/12/2018
%
% License: GPLv3
%
fprintf('\n ** Running Mitliagkas Power Method...\n')
% scope in global variables
global use_blk_err
% ambient dim defined by our input matrix Y
[n, Ti] = size(Y);
% check we calculate the error (disabled for speed runs)
if nargin < 5
no_err = 0;
end
% check if we have a floor multiplier as an argument
if nargin < 4
floor_mul = 2;
end
% check if n < r or n == 1
if n == 1 || n < r
error(" ** ERR: Ambient dimension must be > 1 and r < n **");
end
% Random initialization
SrHat = orth(randn(n));
SrHatTemp = SrHat(:, 1:r);
% power method configuration
% set the block, depending on argument
if nargin < 3
B = 2*n;
else
if blkSize < 2*n
fprintf("\n !! WARN: Block size must be at least 2*n !!\n");
B = 2*n;
else
B = blkSize;
end
end
% check if Ti < B, in which case we cannot run it
if Ti < B
error("\n Block size must be lower than the number of columns");
end
K = floor(Ti/B); % Number of blocks
cnt = 1; % index for error block align
% preallocate based on no error run and block error
if no_err == 0
if use_blk_err == 1
T = nan(1, K); % T steps for error log
ErrFro = nan(1, K); % Fro normalised error with T
else
T = 1:Ti; % T steps for error log
ErrFro = nan(1, size(T, 2)); % Fro normalised error with T
end
else
T = 0;
ErrFro = 0;
end
% output the number of blocks and their size
fprintf([' ** Total number of blocks (k): %d ', ...
'with Block size of: %d\n'], K, B);
% start timing
ts = tic;
for k = 1:K
% update the previous subspace estimation
SrHat_old = SrHatTemp;
% indices for current block
min_t = ((k-1)*B)+1;
max_t = k*B;
% initialise the temporary place-holder
temp = zeros(n,r);
for i = 1:B
index = (k-1)*B+i;
temp = temp+1/B*(Y(:,index)*(Y(:,index))')*SrHat(:, 1:r);
end
[SrHat, ~, ~] = svds(temp, floor(floor_mul * r));
SrHatTemp = SrHat(:, 1:r);
% Mitliagkas Power Method Errors start
if no_err == 0
% use the block error
if use_blk_err == 1
% generate current Yr estimation
YrHat_c = (SrHat*SrHat')*Y(:, 1:k*B);
% Frobenius norm incremental error, per block at
% at (k-1)B:kB normalised with current T.
temp = sum(sum((Y(:, 1:max_t)-YrHat_c).^2, 1));
ErrFro(cnt) = temp/max_t;
T(cnt) = max_t;
cnt = cnt + 1;
else
% generate current Yr estimation
YrHat_c = (SrHat_old*SrHat_old')*Y(:, 1:k*B);
% Frobenius norm incremental error, per column in the block
% at (k-1)B:kB normalised with current T.
temp = cumsum(sum((Y(:, 1:max_t)-YrHat_c).^2, 1));
ErrFro(cnt:cnt+B-1) = temp(min_t:max_t)./(min_t:max_t);
cnt = cnt + B;
end
end
% Mitliagkas Power Method errors end
end
% check what to do in case of error generation
if no_err == 0
YrHat = YrHat_c; % current YrHat estimation
else
YrHat = (SrHatTemp*SrHatTemp')*Y(:, 1:k*B);
end
% update the other estimates
U = SrHatTemp; % current U estimation
t = my_toc(ts); % current elapsed time
end