How to avoid a loop that contains if statements
조회 수: 1(최근 30일)
I've got this loop that I'm trying to get rid of, in the hope of saving on computational time.
Here is the basic idea: I have a vector x_offgrid(:,j1,j2) that contains values that are off a grid. E.g. let the grid be x=[1 3 5 7 9], then x_offgrid might have values [1.7 5.5 3 8.5]. The vector w(:,j1,j2) contains weights associated with x_offgrid. Now, I want to find w_new(:,j1,j2) that gives me weights associated with x.
For now, my idea is to follow the steps
- Find the nearest neighbor of every value in x_offgrid in x. Call that vector x_ongrid. In my example, that is x_ongrid=[1 5 3 9]. The corresponding index is xind x_ongrid = x(xind)
- For every value in x_offgrid(i,j1,j2), a fraction of the weight w(i,j1,j2) goes to the nearest neighbor and the other fraction goes to the opposite neighbor (xindmin and xindmax). Those fractions are given by fractiondown and fractionup.
for i = 1:N2
if x_ongrid(i,j1,j2)>x_offgrid(i,j1,j2) % nearest neighbor is above
z(i,xind(i,j1,j2)) = w(i,j1,j2)*(1-fractiondown(i,j1,j2));
z(i,xindmax(i,j1,j2)) = z(i,xindmax(i,j1,j2)) ...
else % nearest neighbor is below
z(i,xind(i,j1,j2)) = w(i,j1,j2)*(1-fractionup(i,j1,j2));
z(i,xindmin(i,j1,j2)) = z(i,xindmin(i,j1,j2)) ...
Does anybody have a general hint what I could try? Or is it not really possible to rewrite this in a more efficient way? I thought about some type of interpolation function, but right now, I don't see how that could work.
Update: Thanks to Guillaume, I've been able to get rid of the if statement, but not of the loop. Now my code looks as follows:
for i = 1:N2
wn(i,j1,xind(i,j1,j2)+1,j2) = w_tmp(i,j1,j2)*weightup(i,j1,j2);
wn(i,j1,xind(i,j1,j2),j2) = w_tmp(i,j1,j2)*weightdown(i,j1,j2);
The problem is that xind might contain the same index multiple times. I'm not sure if I can make use of accumarray because I want to keep wn is a 4-dimensional matrix.
Guillaume 2015년 2월 17일
편집: Guillaume 2015년 2월 17일
Assuming your x is monotically increasing (if not, sort it and save the original indices to reorder the data at the end), and also assuming that x_offgrid is always within [x(1) x(end)], it think this may be what you want:
x = [1 3 5 7 9]; %also ignoring vector vs 3d array, not sure it matters
x_offgrid = [1.7 5.5 3 8.5];
w = [1 10 100 1000];
[~, xind] = histc(x_offgrid, x);
%[~, ~, xind] = histcounts(x_offgrid, x); %using new histcounts in R2014b
%x_ongrid = x(xind); %not used
weightup = w .* (x_offgrid - x(xind)) ./ (x(xind+1) - x(xind));
weightdown = w .* (x(xind+1) - x_offgrid) ./ (x(xind+1) - x(xind));
w_new = accumarray([xind xind+1]', [weightdown weightup])
Note that I've not tried to fully understand your example code (since you've not provided all the details). What I've done is find the lower point corresponding to your x_offgrid using histc. I then calculate a proportion of the weight to redistribute to the upper and lower point according to the distance of x_offgrid to these points (the weightup and weightdown). Finally, the final weight is the sum of the weightup and weightdown corresponding to each grid point. I use accumarray for that.