Accelerating ggplot2: use a canvas to speed up plots creation

Too wrong; don’t read

Basically, this post turned out to be just a wrong, premature, and unnecessary attempt of code optimization. If you still want to have look, make sure that later you read this post by Thomas Lin Pedersen. You are warned

This post is updated on 2017-07-15

The earlier version had a terminology mistake pointed out by Hadley Wickham. I wrongly called creation time of the plots as rendering time.

One of the nice features of the ggapproach to plotting is that one can save plots as R objects at any step and use later to render and/or modify. I used that feature extensively while creating maps with ggplot2 (see my previous posts: one, two, three, four, five). It is just convenient to first create a canvas with all the theme parameters appropriate for a map, and then overlay the map layer. At some point I decided to check if that workflow was computationally efficient or not. To my surprise, the usage of canvas reduces the creation time of a ggplot quite a lot. To my further surprise, this finding holds for simple plots as well as maps.

Let’s start with a simple check.

# load required packages
library(tidyverse)      # data manipulation and viz
library(ggthemes)       # themes for ggplot2
library(viridis)        # the best color palette
library(rgdal)          # deal with shapefiles
library(microbenchmark) # measure the speed of executing
library(extrafont)      # nice font
myfont <- "Roboto Condensed"
library(RColorBrewer)

# create a canvas 
canv_mt <- ggplot(mtcars, aes(hp, mpg, color = cyl))+
        coord_cartesian()

# test speed with mocrobenchmark
test <- microbenchmark(
        without_canvas = ggplot(mtcars, aes(hp, mpg, color = cyl))+
                coord_cartesian()+
                geom_point()
        
        ,
        
        with_canvas = canv_mt+
                geom_point()
       
        ,
        
        times = 100
)

test

autoplot(test)+
        aes(fill = expr)+
        scale_fill_viridis(discrete = T)+
        theme_bw(base_size = 15, base_family = myfont)+
        theme(legend.position = "none",
              axis.text = element_text(size = 15))+
        labs(title = "The speed of creating a simple ggplot")

Figure 1. Microbenchmark output for a simple plot

The median time of execution is 3.24 milliseconds for the plot without canvas and 2.29 milliseconds for the plot with canvas.

Next, let’s do the same check for a map. For that, I will use the data prepared for one of my earlier posts and recreate the simple map that shows the division of European Union 27 into three subregions.

Figure 2. The map we use to test the plot creation speed

# load the already prepared data
load(url("https://ikashnitsky.github.io/doc/misc/map-subplots/df-27-261-urb-rur.RData"))
load(url("https://ikashnitsky.github.io/doc/misc/map-subplots/spatial-27-261.RData"))

# fortify spatial objects
neib <- fortify(Sneighbors)
bord <- fortify(Sborders)
fort <- fortify(Sn2, region = "id")

# join spatial and statistical data
fort_map <- left_join(df, fort, "id")

# pal for the subregions
brbg3 <- brewer.pal(11,"BrBG")[c(8,2,11)]

# create a blank map
basemap <- ggplot()+
        geom_polygon(data = neib,
                     aes(x = long, y = lat, group = group),
                     fill = "grey90",color = "grey90")+
        coord_equal(ylim = c(1350000,5450000), 
                    xlim = c(2500000, 6600000), 
                    expand = c(0,0))+
        theme_map(base_family = myfont)+
        theme(panel.border = element_rect(color = "black",size = .5,fill = NA),
              legend.position = c(1, 1),
              legend.justification = c(1, 1),
              legend.background = element_rect(colour = NA, fill = NA),
              legend.title = element_text(size = 15),
              legend.text = element_text(size = 15))+
        labs(x = NULL, y = NULL)


# test speed with mocrobenchmark
test_map <- microbenchmark(
        without_canvas = 
                ggplot()+
                geom_polygon(data = neib,
                             aes(x = long, y = lat, group = group),
                             fill = "grey90",color = "grey90")+
                coord_equal(ylim = c(1350000,5450000), 
                            xlim = c(2500000, 6600000), 
                            expand = c(0,0))+
                theme_map(base_family = myfont)+
                theme(panel.border = element_rect(color = "black",
                                                  size = .5,fill = NA),
                      legend.position = c(1, 1),
                      legend.justification = c(1, 1),
                      legend.background = element_rect(colour = NA, fill = NA),
                      legend.title = element_text(size = 15),
                      legend.text = element_text(size = 15))+
                labs(x = NULL, y = NULL) +
                geom_polygon(data = fort_map, 
                             aes(x = long, y = lat, group = group, 
                                 fill = subregion), color = NA)+
                scale_fill_manual(values = rev(brbg3)) +
                theme(legend.position = "none")
        
        ,
        
        with_canvas = 
                basemap +
                geom_polygon(data = fort_map, 
                             aes(x = long, y = lat, group = group, 
                                 fill = subregion), color = NA)+
                scale_fill_manual(values = rev(brbg3)) +
                theme(legend.position = "none")
        
        ,
        
        times = 100
)
      

autoplot(test_map)+
        aes(fill = expr)+
        scale_fill_viridis(discrete = T)+
        theme_bw(base_size = 15, base_family = myfont)+
        theme(legend.position = "none",
              axis.text = element_text(size = 15))+
        labs(title = "The speed of creating a map with ggplot2")

Figure 3. Microbenchmark output for a map

The median time of execution is 18.8 milliseconds for the map without canvas and 6.3 milliseconds for the map with canvas.

Conclusion: Use canvas with `ggplot2`

For the full script to reproduce the results check out this gist

--- title: "Accelerating ggplot2: use a canvas to speed up plots creation" description-meta: "{{< meta website.description >}}" date: "2017-07-04" image: teaser.png categories: [r, ggplot2] --- *** ::: {.callout-important} # Too wrong; don't read Basically, this post turned out to be just a wrong, premature, and unnecessary attempt of code optimization. If you still want to have look, make sure that later you read [this post][tlp] by Thomas Lin Pedersen. You are warned ::: *** ::: {.callout-note} # This post is updated on 2017-07-15 The earlier version had a terminology mistake [pointed out by Hadley Wickham][hadley]. I wrongly called creation time of the plots as rendering time. ::: One of the nice features of the `gg`approach to plotting is that one can save plots as R objects at any step and use later to render and/or modify. I used that feature extensively while creating maps with `ggplot2` (see my previous posts: [one][one], [two][two], [three][three], [four][four], [five][five]). It is just convenient to first create a canvas with all the theme parameters appropriate for a map, and then overlay the map layer. At some point I decided to check if that workflow was computationally efficient or not. To my surprise, the usage of canvas reduces the creation time of a ggplot quite a lot. To my further surprise, this finding holds for simple plots as well as maps. Let's start with a simple check. ```{r, eval=FALSE} # load required packages library(tidyverse) # data manipulation and viz library(ggthemes) # themes for ggplot2 library(viridis) # the best color palette library(rgdal) # deal with shapefiles library(microbenchmark) # measure the speed of executing library(extrafont) # nice font myfont <- "Roboto Condensed" library(RColorBrewer) # create a canvas canv_mt <- ggplot(mtcars, aes(hp, mpg, color = cyl))+ coord_cartesian() # test speed with mocrobenchmark test <- microbenchmark( without_canvas = ggplot(mtcars, aes(hp, mpg, color = cyl))+ coord_cartesian()+ geom_point() , with_canvas = canv_mt+ geom_point() , times = 100 ) test autoplot(test)+ aes(fill = expr)+ scale_fill_viridis(discrete = T)+ theme_bw(base_size = 15, base_family = myfont)+ theme(legend.position = "none", axis.text = element_text(size = 15))+ labs(title = "The speed of creating a simple ggplot") ``` ![](fig-01-test-simple.png) **Figure 1. Microbenchmark output for a simple plot** The median time of execution is 3.24 milliseconds for the plot **without** canvas and 2.29 milliseconds for the plot **with** canvas. Next, let's do the same check for a map. For that, I will use the data prepared for [one of my earlier posts][four] and recreate the simple map that shows the division of European Union 27 into three subregions. ![](fig-02-sub.png) **Figure 2. The map we use to test the plot creation speed** ```{r, eval=FALSE} # load the already prepared data load(url("https://ikashnitsky.github.io/doc/misc/map-subplots/df-27-261-urb-rur.RData")) load(url("https://ikashnitsky.github.io/doc/misc/map-subplots/spatial-27-261.RData")) # fortify spatial objects neib <- fortify(Sneighbors) bord <- fortify(Sborders) fort <- fortify(Sn2, region = "id") # join spatial and statistical data fort_map <- left_join(df, fort, "id") # pal for the subregions brbg3 <- brewer.pal(11,"BrBG")[c(8,2,11)] # create a blank map basemap <- ggplot()+ geom_polygon(data = neib, aes(x = long, y = lat, group = group), fill = "grey90",color = "grey90")+ coord_equal(ylim = c(1350000,5450000), xlim = c(2500000, 6600000), expand = c(0,0))+ theme_map(base_family = myfont)+ theme(panel.border = element_rect(color = "black",size = .5,fill = NA), legend.position = c(1, 1), legend.justification = c(1, 1), legend.background = element_rect(colour = NA, fill = NA), legend.title = element_text(size = 15), legend.text = element_text(size = 15))+ labs(x = NULL, y = NULL) # test speed with mocrobenchmark test_map <- microbenchmark( without_canvas = ggplot()+ geom_polygon(data = neib, aes(x = long, y = lat, group = group), fill = "grey90",color = "grey90")+ coord_equal(ylim = c(1350000,5450000), xlim = c(2500000, 6600000), expand = c(0,0))+ theme_map(base_family = myfont)+ theme(panel.border = element_rect(color = "black", size = .5,fill = NA), legend.position = c(1, 1), legend.justification = c(1, 1), legend.background = element_rect(colour = NA, fill = NA), legend.title = element_text(size = 15), legend.text = element_text(size = 15))+ labs(x = NULL, y = NULL) + geom_polygon(data = fort_map, aes(x = long, y = lat, group = group, fill = subregion), color = NA)+ scale_fill_manual(values = rev(brbg3)) + theme(legend.position = "none") , with_canvas = basemap + geom_polygon(data = fort_map, aes(x = long, y = lat, group = group, fill = subregion), color = NA)+ scale_fill_manual(values = rev(brbg3)) + theme(legend.position = "none") , times = 100 ) autoplot(test_map)+ aes(fill = expr)+ scale_fill_viridis(discrete = T)+ theme_bw(base_size = 15, base_family = myfont)+ theme(legend.position = "none", axis.text = element_text(size = 15))+ labs(title = "The speed of creating a map with ggplot2") ``` ![](fig-03-test-map.png) **Figure 3. Microbenchmark output for a map** The median time of execution is 18.8 milliseconds for the map **without** canvas and 6.3 milliseconds for the map **with** canvas. **Conclusion: Use canvas with `ggplot2`** *** ::: {.callout-tip} # For the full script to reproduce the results check out this [gist][gist] ::: [tlp]: https://www.data-imaginist.com/2017/beneath-the-canvas/ [one]: https://ikashnitsky.github.io/2017/denmark-nuts-reconstruction/ [two]: https://ikashnitsky.github.io/2017/map-hacking/ [three]: https://ikashnitsky.github.io/2017/align-six-maps/ [four]: https://ikashnitsky.github.io/2017/subplots-in-maps/ [five]: https://ikashnitsky.github.io/2017/colorcoded-map/ [gist]: https://gist.github.com/ikashnitsky/b9c5d0b838daa2338066dbaa3e035dcc [hadley]: https://twitter.com/hadleywickham/status/882217871769837569

Conclusion: Use canvas with ggplot2

Conclusion: Use canvas with `ggplot2`