What is the recommended practice for plotting the outputs of histcounts?

I have generated multiple histograms with the histcounts function. The histograms are computed from large data sets that I cannot store in memory all at once. I want to be able to plot the histograms to compare the distributions of these large data sets. The 'histogram' function accepts the raw population and makes a histogram plot, but does not accept the outputs of histcounts and simply plot them up. The 'bar' function does not directly accept the outputs from histcounts either, since the edge array is an N+1 point array and the counts array is a N point array. Do I have to convert my N+1 point edge arrays to N point bin center arrays to plot these histograms? Or is there a another function that directly accepts the outputs from histcounts to make a histogram plot? Not a difficult problem to solve, but it makes these 'new and encouraged' histogram functions seem kind of clunky compared to the 'discouraged' histogram functions.

Answer 1

[Steven Lord]

If you cannot directly call histogram to bin your data while visualizing it but need to visualize the output of one or more calls to histcounts (for example if you call histcounts repeatedly with a common set of bin edges and a subset of your data each time then add the bin counts together), call histogram with the 'BinCounts' and 'BinEdges' properties. Compare the two figures created by this example:

rng default
x = randn(10000, 1);
y = randn(10000, 1) + 5;
[minCombined, maxCombined] = bounds([x; y]);
BE = linspace(minCombined, maxCombined, 100);

figure
h1 = histogram([x; y], BE);

bincounts = histcounts(x, BE);
bincounts2 = histcounts(y, BE);
figure;
h2 = histogram('BinCounts', bincounts + bincounts2, 'BinEdges', BE);

While I did concatenate x and y together to identify the minimum and maximum of the combined data to generate a set of bins that would work for both data sets, you may be able to generate a common set of bin edges without needing x and y to be in memory simultaneously.

If you're planning on working with data too large to fit in memory regularly, I would like to point out that the histcounts and histogram functions support tall arrays with some limitations given in the Extended Capabilities section on their documentation pages.

Answer 2

[Image Analyst]

I use bar(), along with other functions like grid, xlabel, ylabel, title, and whatever else I want to use to fancy up the chart.

[Brando Miranda]

do u have a working example? I keep having that the output by histcounts is one dimension less than edges. Then bar freaks out.

[Walter Roberson]

N = histcounts(X, edges);
centers = (edges(1:end-1) + edges(2:end))/2;
bar(centers, N)

Provided, that is, that none of the edges are infinite.

[Bhavin Khatri]

This is of course correct, but begs the question why this change was made, so as to output edges rather than bin centres!!

[Image Analyst]

I don't think that was a change, at least it's been that for as long as I remember. Perhaps you're thinking of imhist() which uses bin centers, which makes sense since it's (usually) dealing with integer values.

[Bruno Luong]

"I don't think that was a change".

Well, not sure if it's related but HISTC returns size of count (N here) equals to number of edges, where as HISTCOUNTS has one less.

Though in both cases the center of the edges must be associated with the returned count.

To me no big deal to compute the center, much ado about nothing.

[Andrew]

good morning, it's been a couple years, but as I recall there is a matlab function called 'hist' that would output histogram counts and bin centers. Those outputs could be passed directly to the bar function, or the plot function, for plotting. The latest versions of matlab documentation, however do not recommend using hist to compute histograms, instead they suggest the 'histogram' function, which outputs N+1 bin edges for a N-bin histogram. The function also automatically makes a plot, so if you want to look at a single histogram, you don't need to do anything more. But if you want to generate many histograms and plot multiple distributions, or perhaps make line plots instead of bar plots from the histogram outputs, the matlab-recommended way of computing histograms then requires the user to compute bin centers from the bin edges to make the plotting functions work.

As a couple users have pointed out, it simply requires a single short line of code, nothing complicated about it, and for the past two years that's what I've been doing. Really the thrust of my intial comment on the matter was the new way matlab is recommending seems to be ever so slightly less efficient at plotting multiple histograms, in that it requires the user to do a little more than they had do to when using the 'old' way. And again, it's not like I'm throwing up my hands and quitting Matlab over this, it's just a little nit to pick.

And since I'm nit-picking, there's another thing about histogram computations I've noticed. It seems like histcounts and histogram do not operate along array dimensions like other matlab functions do. For example, if I input a NxM matrix X into histogram, it sorts all the data X into a single histogram object. So if I have M populations with N elements each, and I want to get a distribution for each population, I have to call histogram M times for each population. There are many other matlab functions that will operate along a given array dimension, allowing a single function call to work on multiple 'instances' of data. Histogram doesn't seem to have that functionality. Am I missing how to do that with histogram, or is there a reason why the histogram function shouldn't have that functionality? Does that make the output too complicated to use? And again, I understand that it's not difficult to call a function multiple times, I'm just poking on the fact that matlab has other functions that seem to be tailored to operate in a slightly different way.

Regards,

[Bruno Luong]

"I've noticed. It seems like histcounts and histogram do not operate along array dimensions like other matlab functions do. "

Here is a work around.

[Image Analyst]

Personally I like the fact that histogram does the whole array and doesn't force me to do (:) after the array. I never would want the histogram of just the columns or just the rows. If I pass an image, why would I want , say, four thousand histograms instead of just one for the whole image? I wouldn't. And I'd guess that others also would rarely want dozens or hundreds of histograms. So while I agree it's inconsistent because it doesn't give a result for every row or every column, I believe it's an inconsistency for the better.

[Bruno Luong]

It's not question of inconsistency, it's a question of a feature that has been removed (working along a dimension).

If you don't use it good for you, IA.

But some people like me do. I do not work with not an image but array of data that I need to bin along some dimension that I select.

But if you prefer image processing, I can give you a concrete example where such thing is needed IA: if one one to do sliding histogram on a small rectangle area of an gray image, one could conver the image to im2col then use HISTC along the first dimension. This is very convenient.

Try to do the same with HISTCOUNTS, you"ll end up with few slow for-loops instead of fast and compact code.