Home / Questions / 295767
Accepted
Patricio Moracho
Asked: 2020-09-24 19:49:41 +0800 CST 2020-09-24 19:49:41 +0800 CST

How to improve performance of a geom_histogram()?

  • 772

I am graphing a histogram from a very important set of data using geom_histogram(), I have realized that as the "definition" of it increases by increasing the number of binsbars, the result is slower and slower. The ratio, with a histogram of R base is at least 10 to 1 in time. Example:

library("ggplot2")
library("microbenchmark")
set.seed(2019)
x <- rnorm(100000)
df <- data.frame(x=x)

ggplot_hist <- function(data, bins=100000){
  print(ggplot(data, aes(x=x)) + geom_histogram(bins=bins))
}

base_hist <- function(x, breaks=100000){
  print(hist(x, breaks=length(x)))
}

microbenchmark(
  base_hist(x), 
  ggplot_hist(df), 
  times=3L 
)

Unit: seconds
            expr       min        lq      mean    median        uq       max neval
    base_hist(x)  4.503556  4.632358  4.680143  4.761159  4.768436 4.775713     3
 ggplot_hist(df) 56.330033 57.249490 60.182923 58.168946 62.109369 66.049791    3

Is there a way to optimize a histogram in ggplot?

r

1 Answers

  1. Best Answer

    According to the hypothesis of this interesting answer , the bottleneck would be in the calculation of the binsor bars. We can try to test it:

    microbenchmark(
  ggplot_hist(df, bins=1), 
  ggplot_hist(df, bins=100), 
  ggplot_hist(df, bins=1000), 
  ggplot_hist(df, bins=10000), 
  ggplot_hist(df, bins=100000), 
  times=3
) -> tiempos
summary(tiempos)[, c(1,4), drop=FALSE]

                           expr       mean
1     ggplot_hist(df, bins = 1)   502.0619
2   ggplot_hist(df, bins = 100)   705.1292
3  ggplot_hist(df, bins = 1000)  1209.0673
4 ggplot_hist(df, bins = 10000)  6143.7755
5 ggplot_hist(df, bins = 1e+05) 55068.0312

    We can see that as we incorporate more level of detail, increasing the binstime grows rapidly. On the other hand, if we study the "base" histogram like this:

    microbenchmark(
  base_hist(df$x, breaks=1), 
  base_hist(df$x, breaks=100), 
  base_hist(df$x, breaks=1000), 
  base_hist(df$x, breaks=10000), 
  base_hist(df$x, breaks=100000), 
  times=3
) -> tiempos_base

summary(tiempos_base)[, c(1,4), drop=FALSE]
                             expr     mean
1     base_hist(df$x, breaks = 1) 10.49051
2   base_hist(df$x, breaks = 100) 10.46801
3  base_hist(df$x, breaks = 1000) 13.60592
4 base_hist(df$x, breaks = 10000) 16.48910
5 base_hist(df$x, breaks = 1e+05) 13.49401

    With the basic histogram, we see that the growth over time as the increases binsis minimal. So the idea, which is proposed in the mentioned answer, is to replace the calculation of the binswith the base function hist()and then draw the bars using a geom_rect(). Let's see:

    quick_hist = function(x, bins=100000) {

  res = hist(x, plot=FALSE, breaks=bins)
  dat = data.frame(xmin=head(res$breaks, -1L),
                   xmax=tail(res$breaks, -1L),
                   ymin=0.0,
                   ymax=res$counts)

  print(ggplot(dat, aes(xmin=xmin, xmax=xmax, ymin=ymin, ymax=ymax)) +
    geom_rect(size=0.5))
}

print(microbenchmark(
  base_hist(x), 
  ggplot_hist(df), 
  quick_hist(df$x), 
  times=3L 
), signif=3)

Unit: seconds
             expr   min    lq  mean median    uq  max neval
     base_hist(x)  4.75  4.76  7.70   4.76  9.17 13.6     3
  ggplot_hist(df) 57.30 57.40 57.60  57.40 57.80 58.1     3
 quick_hist(df$x)  6.45  6.64  9.57   6.83 11.10 15.4     3

    And we see that with the ad-hoc function quick_hist()we have managed ggplotto improve the performance of the histogram in a radical way and with a very similar visual result.

    • 2