Man*_*non 3 matlab matrix vectorization
我有一个矩阵:
A = [1 1 1
2 2 2
3 3 3]
Run Code Online (Sandbox Code Playgroud)
是否有矢量化的获取方式:
B = [1 0 0
0 1 0
0 0 1
2 0 0
0 2 0
0 0 2
3 0 0
0 3 0
0 0 3]
Run Code Online (Sandbox Code Playgroud)
%// example data
data = reshape(1:9,3,3).' %'
n = 3; %// assumed to be known
data =
1 2 3
4 5 6
7 8 9
Run Code Online (Sandbox Code Playgroud)
%// row indices
rows = 1:numel(data);
%// column indices
cols = mod(rows-1,n) + 1;
%// pre-allocation
out = zeros(n*n,n);
%// linear indices
linIdx = sub2ind(size(out),rows,cols);
%// assigning
out(linIdx) = data.'
Run Code Online (Sandbox Code Playgroud)
out =
1 0 0
0 2 0
0 0 3
4 0 0
0 5 0
0 0 6
7 0 0
0 8 0
0 0 9
Run Code Online (Sandbox Code Playgroud)
或者,如果您希望保存代码行,而不是可读性:
out = zeros(n*n,n);
out(sub2ind(size(out),1:numel(data),mod((1:numel(data))-1,n) + 1)) = data.'
Run Code Online (Sandbox Code Playgroud)
另外两种快速解决方案,但速度不快于其他解决方案:
%// #1
Z = blockproc(A,[1 size(A,2)],@(x) diag(x.data));
%// #2
n = size(A,2);
Z = zeros(n*n,n);
Z( repmat(logical(eye(n)),n,1) ) = A;
Run Code Online (Sandbox Code Playgroud)
function [t] = bench()
A = magic(200);
% functions to compare
fcns = {
@() thewaywewalk(A);
@() lhcgeneva(A);
@() rayryeng(A);
@() rlbond(A);
};
% timeit
t = zeros(4,1);
for ii = 1:10;
t = t + cellfun(@timeit, fcns);
end
format long
end
function Z = thewaywewalk(A)
n = size(A,2);
rows = 1:numel(A);
cols = mod(rows-1,n) + 1;
Z = zeros(n*n,n);
linIdx = sub2ind(size(Z),rows,cols);
Z(linIdx) = A.';
end
function Z = lhcgeneva(A)
sz = size(A);
Z = zeros(sz(1)*sz(2), sz(2));
for i = 1 : sz(1)
Z((i-1)*sz(2)+1:i*sz(2), :) = diag(A(i, :));
end
end
function Z = rayryeng(A)
A = A.';
Z = full(sparse(1:numel(A), repmat(1:size(A,2),1,size(A,1)), A(:)));
end
function Z = rlbond(A)
D = cellfun(@diag,mat2cell(A, ones(size(A,1), 1), size(A,2)), 'UniformOutput', false);
Z = vertcat(D{:});
end
Run Code Online (Sandbox Code Playgroud)
ans =
0.322633905428601 %// thewaywewalk
0.550931853207228 %// lhcgeneva
0.254718792359946 %// rayryeng - Winner!
0.898236688657039 %// rlbond
Run Code Online (Sandbox Code Playgroud)
A = [1 2 3; 4 5 6; 7 8 9];
A = A.';
B = full(sparse(1:numel(A), repmat(1:size(A,1),1,size(A,2)), A(:)));
Run Code Online (Sandbox Code Playgroud)
原始矩阵在A,我将其转置,以便我可以正确地展开每个矩阵的行以进行下一步.我sparse用来声明矩阵中的非零值.具体来说,我们看到每行只有一个条目,因此行索引的范围应为1到最多A.列从1到最后一列波动并重复. mod肯定是通过thewaywewalk解决方案的方式,但我想使用repmat所以这是他的方法的独立解决方案.因此,我们创建了一个用于访问列的向量,这些列从1到尽可能多的列,并且我们重复这个行以获得尽可能多的行.这些行和列索引向量将决定非零位置的出现位置.最后,进入每个非零位置的是按A行主列顺序展开的元素,遵循行和列索引向量指示的顺序.
请注意,在repmat调用中,size由于转置操作,调用时的行和列会反转.
结果如下,我们得到:
>> B
B =
1 0 0
0 2 0
0 0 3
4 0 0
0 5 0
0 0 6
7 0 0
0 8 0
0 0 9
Run Code Online (Sandbox Code Playgroud)
鉴于上述问题的稀疏性,将矩阵保留在sparse表格中并且仅full在必要时进行转换可能更快.将花费时间在两种格式之间进行转换,因此如果您决定进行基准测试,请考虑这一点.