R:大数据的区别？计算两个矩阵之间的最小距离

Question

我有两个矩阵,一个是200K行,另一个是20K.对于第一个矩阵中的每一行(也就是一个点),我试图找到第二个矩阵中哪一行(也是一个点)最接近第一个矩阵中的点.这是我在样本数据集上尝试的第一种方法:

#Test dataset
pixels.latlon=cbind(runif(200000,min=-180, max=-120), runif(200000, min=50, max=85))
grwl.latlon=cbind(runif(20000,min=-180, max=-120), runif(20000, min=50, max=85))
#calculate the distance matrix
library(geosphere)
dist.matrix=distm(pixels.latlon, grwl.latlon, fun=distHaversine)
#Pick out the indices of the minimum distance
rnum=apply(dist.matrix, 1, which.min)

但是,Error: cannot allocate vector of size 30.1 Gb当我使用该distm功能时出现错误.

关于这个主题有几个帖子:

这个bigmemory用于计算SAME数据帧中各点之间的距离,但我不确定如何调整它来计算两个不同矩阵中点之间的距离... https://stevemosher.wordpress.com/2012/04/12 /缺口司托克斯-距离代码现在与-大存储器/

这个也适用于计算SAME矩阵中各点之间的距离矩阵... 用于重复距离矩阵计算的高效(记忆方式)函数和超大距离矩阵的分块

这个与我想做的几乎完全相同,但他们实际上没有提出适用于大数据的解决方案:R:使用大内存的distm我试过这个方法,它使用bigmemory但是得到Error in CreateFileBackedBigMatrix(as.character(backingfile), as.character(backingpath), : Problem creating filebacked matrix.错误,我认为因为数据帧太大了.

有没有人想出这个问题的好方法？我对其他包装的想法持开放态度!

更新了解决问题的代码

pixels.latlon=cbind(runif(200000,min=-180, max=-120), runif(200000, min=50, max=85))
grwl.tibble = tibble(long=runif(20000,min=-180, max=-120), lat=runif(20000, min=50, max=85), id=runif(20000, min=0, max=20000))
rnum <- apply(pixels.latlon, 1, function(x) {
    xlon=x[1]
    xlat=x[2]
    grwl.filt = grwl.tibble %>% 
      filter(long < (xlon+0.3) & long >(xlon-0.3) & lat < (xlat+0.3)&lat >(xlat-.3))
    grwl.latlon.filt = cbind(grwl.filt$long, grwl.filt$lat)
    dm <- distm(x, grwl.latlon.filt, fun=distHaversine)
    rnum=apply(dm, 1, which.min)
    id = grwl.filt$id[rnum]
    return(id)
                     })

Answer 1

你可以使用这个R(cpp)函数:

#include <Rcpp.h>
using namespace Rcpp;

double compute_a(double lat1, double long1, double lat2, double long2) {

  double sin_dLat = ::sin((lat2 - lat1) / 2);
  double sin_dLon = ::sin((long2 - long1) / 2);

  return sin_dLat * sin_dLat + ::cos(lat1) * ::cos(lat2) * sin_dLon * sin_dLon;
}

int find_min(double lat1, double long1,
             const NumericVector& lat2,
             const NumericVector& long2,
             int current0) {

  int m = lat2.size();
  double lat_k, lat_min, lat_max, a, a0;
  int k, current = current0;

  a0 = compute_a(lat1, long1, lat2[current], long2[current]);
  // Search before current0
  lat_min = lat1 - 2 * ::asin(::sqrt(a0));
  for (k = current0 - 1; k >= 0; k--) {
    lat_k = lat2[k];
    if (lat_k > lat_min) {
      a = compute_a(lat1, long1, lat_k, long2[k]);
      if (a < a0) {
        a0 = a;
        current = k;
        lat_min = lat1 - 2 * ::asin(::sqrt(a0));
      }
    } else {
      // No need to search further
      break;
    }
  }
  // Search after current0
  lat_max = lat1 + 2 * ::asin(::sqrt(a0));
  for (k = current0 + 1; k < m; k++) {
    lat_k = lat2[k];
    if (lat_k < lat_max) {
      a = compute_a(lat1, long1, lat_k, long2[k]);
      if (a < a0) {
        a0 = a;
        current = k;
        lat_max = lat1 + 2 * ::asin(::sqrt(a0));
      }
    } else {
      // No need to search further
      break;
    }
  }

  return current;
} 

// [[Rcpp::export]]
IntegerVector find_closest_point(const NumericVector& lat1,
                                 const NumericVector& long1,
                                 const NumericVector& lat2,
                                 const NumericVector& long2) {

  int n = lat1.size();
  IntegerVector res(n);

  int current = 0;
  for (int i = 0; i < n; i++) {
    res[i] = current = find_min(lat1[i], long1[i], lat2, long2, current);
  }

  return res; // need +1
}


/*** R
N <- 2000  # 2e6
M <- 500   # 2e4

pixels.latlon=cbind(runif(N,min=-180, max=-120), runif(N, min=50, max=85))
grwl.latlon=cbind(runif(M,min=-180, max=-120), runif(M, min=50, max=85))
# grwl.latlon <- grwl.latlon[order(grwl.latlon[, 2]), ]

library(geosphere)
system.time({
  #calculate the distance matrix
  dist.matrix = distm(pixels.latlon, grwl.latlon, fun=distHaversine)
  #Pick out the indices of the minimum distance
  rnum=apply(dist.matrix, 1, which.min)
})


find_closest <- function(lat1, long1, lat2, long2) {

  toRad <- pi / 180
  lat1  <- lat1  * toRad
  long1 <- long1 * toRad
  lat2  <- lat2  * toRad
  long2 <- long2 * toRad

  ord1  <- order(lat1)
  rank1 <- match(seq_along(lat1), ord1)
  ord2  <- order(lat2)

  ind <- find_closest_point(lat1[ord1], long1[ord1], lat2[ord2], long2[ord2])

  ord2[ind + 1][rank1]
}

system.time(
  test <- find_closest(pixels.latlon[, 2], pixels.latlon[, 1], 
                       grwl.latlon[, 2], grwl.latlon[, 1])
)
all.equal(test, rnum)

N <- 2e4
M <- 2e4
pixels.latlon=cbind(runif(N,min=-180, max=-120), runif(N, min=50, max=85))
grwl.latlon=cbind(long = runif(M,min=-180, max=-120), lat = runif(M, min=50, max=85))
system.time(
  test <- find_closest(pixels.latlon[, 2], pixels.latlon[, 1], 
                       grwl.latlon[, 2], grwl.latlon[, 1])
)
*/

它需要0.5秒,N = 2e4而4.2秒N = 2e5.我无法让你的代码进行比较.