两个boost :: accumulators :: accumulator_set相互干扰

Question

I have created a class Histogram in my code which is intended as a wrapper for boost::accumulators::accumulator_set from Boost 1.54. The things that seem important to my problem are those lines from the Histogram.hpp file:

using namespace boost::accumulators;

class Histogram {
    public:
        Histogram(int bins, size_t cache);
        accumulator_set<double,
                        features<tag::min, tag::max, tag::mean, tag::density>> acc;
};

Then in Histogram.cpp I have the constructor:

Histogram::Histogram(int bins, size_t cache)
    : acc(accumulator_set<double,
          features<tag::min, tag::max, tag::mean, tag::density>>(
              tag::density::num_bins = bins,
              tag::density::cache_size = std::min(cache, MAX_CACHE_ENTRIES))) {
}

The code that uses this histogram ( do_iterations() in main-metropolis.cpp ) starts with the following:

Histogram position_histogram{settings.position_hist_bins, settings.time_sites * settings.iterations};
//Histogram action_histogram{settings.action_hist_bins, settings.iterations};

It works just like I expect when I run it with the second line deactivated. My simulation generates some data points, puts it into the Histogram::acc and lets me extract it afterwards:

-2.86958    0
-2.37393    0.0002
-1.87829    0.0071
-1.38265    0.06621
-0.887001   0.23902
-0.391356   0.33247
0.104288    0.2342
0.599932    0.08449
1.09558 0.02843
1.59122 0.00775
2.08687 0.00012
2.58251 1e-05
# Min -2.37393
# Max 2.58251
# Mean -0.0809983

Then I activate the line, and the position_histogram works in a really strange way. The bins are all zero, but the data is distributed into the overflow bins in the first and last bin:

0   0.57785
0   0
0   0
0   0
0   0
0   0
0   0
0   0
0   0
0   0
0   0
0   0.42215
# Min -2.37393
# Max 2.58251
# Mean -0.0809983

If I swap the lines, it is action_histogram that breaks. So the second one always breaks the first one. Why does an initialization of a second Histogram and therefore a second accumulator_set causes the first one to break?

---

Please use revision d3081a1ef7 when you browse through the code since I build my own histogram implementation by now to continue work.

Answer 1

You're going to have to debug this or supply more information.

I have used accumulators and always with more than one instance simultaneously in my research proof-of-concepts, and I have not encountered this. Then I realized I never did a density histogram in parallel, so I tested it out.

It pans out in my test based on your declarations, see it Live On Coliru :

#include <boost/accumulators/statistics.hpp>
#include <boost/accumulators/accumulators.hpp>
#include <boost/random.hpp>
#include <boost/bind.hpp>

using namespace boost::accumulators;

static const size_t MAX_CACHE_ENTRIES = 32;

class Histogram {
    public:
        Histogram(int bins, size_t cache)
            : acc(accumulator_set<double,
                    features<tag::min, tag::max, tag::mean, tag::density>>(
                        tag::density::num_bins = bins,
                        tag::density::cache_size = std::min(cache, MAX_CACHE_ENTRIES))) {
            }        

        accumulator_set<double,
                        features<tag::min, tag::max, tag::mean, tag::density>> acc;
};

int main()
{
    Histogram position_histogram { 10, 32 };
    Histogram action_histogram   { 10, 32 };

    auto random = boost::bind(boost::uniform_real<double>(-100,100), boost::mt19937(42));

    size_t samples = 1<<20;
    while (samples--)
    {
        auto v = random();
        position_histogram.acc(v);
        action_histogram.acc(v);
    }

    for (auto& acc : { position_histogram.acc, action_histogram.acc })
    {
        auto hist = density(acc);

        double total = 0.0;

        for( int i = 0; i < hist.size(); i++ ) 
        {
            std::cout << "Bin lower bound: " << hist[i].first << ", Value: " << hist[i].second << std::endl; 
            total += hist[i].second;
        }

        std::cout << "Total: " << total << std::endl; //should be 1 (and it is)
    }
}

Output, as expected:

Bin lower bound: -119.673, Value: 0.000766754
Bin lower bound: -99.8442, Value: 0.099205
Bin lower bound: -80.0156, Value: 0.0987797
Bin lower bound: -60.1869, Value: 0.0990477
Bin lower bound: -40.3583, Value: 0.0991993
Bin lower bound: -20.5296, Value: 0.0989904
Bin lower bound: -0.700967, Value: 0.0993652
Bin lower bound: 19.1277, Value: 0.0993567
Bin lower bound: 38.9563, Value: 0.0993252
Bin lower bound: 58.785, Value: 0.0993109
Bin lower bound: 78.6137, Value: 0.0989342
Bin lower bound: 98.4423, Value: 0.00771904
Total: 1
Bin lower bound: -119.673, Value: 0.000766754
Bin lower bound: -99.8442, Value: 0.099205
Bin lower bound: -80.0156, Value: 0.0987797
Bin lower bound: -60.1869, Value: 0.0990477
Bin lower bound: -40.3583, Value: 0.0991993
Bin lower bound: -20.5296, Value: 0.0989904
Bin lower bound: -0.700967, Value: 0.0993652
Bin lower bound: 19.1277, Value: 0.0993567
Bin lower bound: 38.9563, Value: 0.0993252
Bin lower bound: 58.785, Value: 0.0993109
Bin lower bound: 78.6137, Value: 0.0989342
Bin lower bound: 98.4423, Value: 0.00771904
Total: 1

Also, when feeding both accumulators different samples, I couldn't make it display any obvious malfunctions.

Hope this helps you realize what is different about your situation (eg do you actually feed both accumulators the right samples?)

I've test with boost 1.53-1.55