-
Notifications
You must be signed in to change notification settings - Fork 147
/
update_spatial_components.m
197 lines (176 loc) · 7.56 KB
/
update_spatial_components.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
189
190
191
192
193
194
195
196
197
function [A,b,C,P] = update_spatial_components(Y,C,f,A_,P,options)
% update spatial footprints and background. Two methods are available for
% estimating the components:
% options.spatial_method = 'constrained'
% [A(i,:),b(i)] = argmin sum(A(i,:))
% subject to || Y(i,:) - A(i,:)*C + b(i)*f || <= sn(i)*sqrt(T);
% options.spatial_method = 'regularized'
% [A(i,:),b(i)] = argmin 0.5*|| Y(i,:) - A(i,:)*C + b(i)*f ||^2 + \Lambda^TA(:)
% the support set of each component is determined a-priori
% INPUTS:
% Y: raw data
% C: temporal components
% f: temporal background
% A_: current estimate of spatial footprints (used for determining search locations only)
% P: dataset parameters (used for noise values and interpolated entries)
% options parameter struct (for noise values and other parameters)
% OUTPUTS:
% A: new estimate of spatial footprints
% b: new estimate of spatial background
% C: temporal components (updated only when spatial components are completely removed)
% Written by:
% Eftychios A. Pnevmatikakis, Simons Foundation, 2015
warning('off', 'MATLAB:maxNumCompThreads:Deprecated');
memmaped = isobject(Y);
if memmaped
sizY = size(Y,'Y');
d = prod(sizY(1:end-1));
T = sizY(end);
else
[d,T] = size(Y);
end
if nargin < 6 || isempty(options); options = []; end
if ~isfield(options,'d1') || isempty(options.d1); d1 = input('What is the total number of rows? \n'); options.d1 = d1; else d1 = options.d1; end % # of rows
if ~isfield(options,'d2') || isempty(options.d2); d2 = input('What is the total number of columns? \n'); options.d2 = d2; else d2 = options.d2; end % # of columns
if ~isfield(options,'d3') || isempty(options.d3); d3 = input('What is the total number of z-planes? \n'); options.d3 = d3; else d3 = options.d3; end % # of columns
if ~isfield(options,'interp'); Y_interp = sparse(d,T); else Y_interp = options.interp; end % identify missing data
if ~isfield(options,'spatial_parallel'); spatial_parallel = ~isempty(which('parpool')); else spatial_parallel = options.spatial_parallel; end % use parallel toolbox if present
if ~isfield(options,'search_method'); method = []; else method = options.search_method; end % search method for determining footprint of spatial components
if ~isfield(options,'tsub') || isempty(options.tsub); tsub = 1; else tsub = options.tsub; end % downsample temporally to estimate A and b
if nargin < 2 || (isempty(A_) && isempty(C)) % at least either spatial or temporal components should be provided
error('Not enough input arguments')
else
if ~isempty(C); K = size(C,1); else K = size(A_,2) - options.nb; end
end
if K == 0
A = [];
b = full(A_(:,K+1:end));
C = [];
return
end
if nargin < 5 || isempty(P); P = preprocess_data(Y,1); end % etsimate noise values if not present
if nargin < 4 || isempty(A_)
IND = ones(d,size(C,1));
else
if islogical(A_) % check if search locations have been provided, otherwise estimate them
IND = A_(:,1:K);
if isempty(C)
INDav = double(IND)/diag(sum(double(IND)));
px = (sum(IND,2)>0);
[b,f] = fast_nmf(double(Y(~px,:)),[],options.nb,50);
b = max(Y*f',0)/norm(f)^2;
C = max(INDav'*Y - (INDav'*b)*f,0);
end
if strcmpi(options.spatial_method,'regularized')
A_ = max((Y - b*f)*C'/(C*C'),0);
A_ = A_.*IND;
A_ = [A_,b];
end
else
IND = determine_search_location(A_(:,1:K),method,options);
end
end
K = size(C,1);
if strcmpi(options.spatial_method,'constrained'); A_ = A_(:,1:K); end
Cf = [C;f];
if size(Cf,1) > size(A_,2) && strcmpi(options.spatial_method,'regularized');
error('When using options.spatial_method = regularized pass [A,b] as an input and not just A');
end
if tsub ~= 1 % downsample data
Ts = floor(T/tsub);
Cf_ds = squeeze(mean(reshape(Cf(:,1:Ts*tsub),[],tsub,Ts),2));
T = Ts;
if memmaped
Y_ds = matfile('Y_ds.mat','Writable',true);
Y_ds.Yr(d,T) = double(0);
options.sn = zeros(d,1);
step_size = 2e4;
for i = 1:step_size:d
Ytemp = double(Y.Yr(i:min(i+step_size-1,d),:));
Yt_ds = mean(reshape(Ytemp(:,1:Ts*tsub),[],tsub,Ts),2);
Y_ds.Yr(i:min(i+step_size-1,d),:) = squeeze(Yt_ds);
options.sn(i:min(i+step_size-1,d)) = get_noise_fft(Yt_ds);
end
else
Y_ds = squeeze(mean(reshape(Y(:,1:Ts*tsub),[],tsub,Ts),2));
options.sn = get_noise_fft(Y_ds);
end
else
options.sn = P.sn;
Y_ds = Y;
Cf_ds = Cf;
end
f_ds = Cf_ds(end-size(f,1)+1:end,:);
if strcmpi(options.spatial_method,'constrained')
if spatial_parallel % solve BPDN problem for each pixel
Nthr = max(20*maxNumCompThreads,round(d*T/2^24));
Nthr = min(Nthr,round(d/1e3));
siz_row = [floor(d/Nthr)*ones(Nthr-mod(d,Nthr),1);(floor(d/Nthr)+1)*ones(mod(d,Nthr),1)];
indeces = [0;cumsum(siz_row)];
Yf = cell(Nthr,1);
A = spalloc(d,size(Cf,1),nnz(IND)+size(f,1)*d);
for nthr = 1:Nthr
if memmaped
Ytemp = double(Y_ds.Yr(indeces(nthr)+1:indeces(nthr+1),:));
else
Ytemp = Y_ds(indeces(nthr)+1:indeces(nthr+1),:);
end
sn_temp = options.sn(indeces(nthr)+1:indeces(nthr+1));
IND_temp = IND(indeces(nthr)+1:indeces(nthr+1),:);
Atemp = spalloc(siz_row(nthr),size(Cf,1),nnz(IND_temp));
Yf{nthr} = Ytemp*f_ds';
parfor px = 1:siz_row(nthr)
fn = ~isnan(Ytemp(px,:)); % identify missing data
ind = find(IND_temp(px,:));
if ~isempty(ind);
ind2 = [ind,K+(1:size(f,1))];
[~, ~, a, ~] = lars_regression_noise(Ytemp(px,fn)', Cf_ds(ind2,fn)', 1, sn_temp(px)^2*T);
a_sparse = sparse(1,ind2,double(a'));
Atemp(px,:) = a_sparse';
end
end
if mod(nthr,50) == 0
fprintf('%2.1f%% of pixels completed \n', indeces(nthr+1)*100/d);
end
A(indeces(nthr)+1:indeces(nthr+1),:) = Atemp;
end
Yf = cell2mat(Yf);
else
A = [zeros(d,K),zeros(d,size(f,1))];
sA = zeros(d1,d2,d3);
Yf = Y_ds*f_ds';
for px = 1:d % estimate spatial components
fn = ~isnan(Y_ds(px,:)); % identify missing data
ind = find(IND(px,:));
if ~isempty(ind);
ind2 = [ind,K+(1:size(f,1))];
[~, ~, a, ~] = lars_regression_noise(Y_ds(px,fn)', Cf_ds(ind2,fn)', 1, options.sn(px)^2*T);
A(px,ind2) = a';
sA(px) = sum(a);
end
end
end
elseif strcmpi(options.spatial_method,'regularized')
options.nb = size(f,1);
A = update_spatial_lasso(Y_ds, A_, Cf_ds, IND, options.sn, [], [], options);
K = size(A,2)-options.nb;
b = full(A(:,K+1:end));
A = A(:,1:K);
C = C(1:K,:);
end
A(isnan(A))=0;
A = sparse(double(A));
A = threshold_components(A,options); % post-processing of components
fprintf('Updated spatial components \n');
ff = find(sum(A(:,1:K))==0); % remove empty components
if ~isempty(ff)
K = K - length(ff);
A(:,ff) = [];
C(ff,:) = [];
Cf_ds(ff,:) = [];
end
if memmaped; delete('Y_ds.mat'); end
if strcmpi(options.spatial_method,'constrained');
b = double(max((double(Yf) - A(:,1:K)*double(Cf_ds(1:K,:)*f_ds'))/(f_ds*f_ds'),0));
A = A(:,1:K);
end