Points: Rasterize scatter plots with geom_point_rast()
Sometimes you need to publish a figure in a vector format:
library(ggplot2)
library(ggrastr)
points_num <- 10000
df <- data.frame(x=rnorm(points_num), y=rnorm(points_num), c=as.factor(1:points_num %% 2))
gg <- ggplot(df, aes(x=x, y=y, color=c)) + scale_color_discrete(guide="none")
gg_vec <- gg + geom_point(size=0.5)
print(gg_vec)
But in other cases, your figure contains thousands of points, e.g. try points_num <- 500000 in the example above, and you will notice the performance issues—it takes significantly longer to render the plot.
In this case, a reasonable solution would be to rasterize the plot. But the problem is that all text becomes rasterized as well.
Raster layers with ggrastr were developed to prevent such a situation, using `rasterized
gg_rasterized <- gg + rasterise(geom_point(), dpi = 300, scale = 1)
print(gg_rasterized)
The legacy function used in older versions of ggrastr was geom_point_rast():
gg_rast <- gg + geom_point_rast(size=0.5)
print(gg_rast)
The plots look the same, but the difference in size can be seen when they are exported to pdfs. Unfortunately, there is a longer rendering time to produce such plots:
PrintFileSize <- function(gg, name) {
invisible(ggsave('tmp.pdf', gg, width=4, height=4))
cat(name, ': ', file.info('tmp.pdf')$size / 1024, ' Kb.\n', sep = '')
unlink('tmp.pdf')
}
PrintFileSize(gg_rast, 'Raster')
#> Raster: 291.5576 Kb.
PrintFileSize(gg_vec, 'Vector')
#> Vector: 556.1484 Kb.
As expected, the difference becomes larger with growth of number of points:
points_num <- 1000000
df <- data.frame(x=rnorm(points_num), y=rnorm(points_num), c=as.factor(1:points_num %% 2))
gg <- ggplot(df, aes(x=x, y=y, color=c)) + scale_color_discrete(guide="none")
gg_vec <- gg + geom_point(size=0.5)
gg_rast <- gg + geom_point_rast(size=0.5)
PrintFileSize(gg_rast, 'Raster')
#> Raster: 358.6611 Kb.
PrintFileSize(gg_vec, 'Vector')
#> Vector: 54786.11 Kb.
Jitter: Rasterize jittered scatter plots with geom_jitter_rast()
Users may also opt to create rasterized scatter plots with jitter:
library(ggplot2)
library(ggrastr)
points_num <- 5000
df <- data.frame(x=rnorm(points_num), y=rnorm(points_num), c=as.factor(1:points_num %% 2))
gg <- ggplot(df, aes(x=x, y=y, color=c)) + scale_color_discrete(guide="none")
gg_jitter_rast <- gg + rasterise(geom_jitter(), dpi = 300, scale = 1)
print(gg_jitter_rast)
The legacy wrapper geom_jitter_rast() used the following syntax:
library(ggplot2)
library(ggrastr)
points_num <- 5000
df <- data.frame(x=rnorm(points_num), y=rnorm(points_num), c=as.factor(1:points_num %% 2))
gg <- ggplot(df, aes(x=x, y=y, color=c)) + scale_color_discrete(guide=FALSE)
gg_jitter_rast <- gg + geom_jitter_rast(raster.dpi=600)
print(gg_jitter_rast)
#> Warning: It is deprecated to specify `guide = FALSE` to remove a guide. Please
#> use `guide = "none"` instead.
Tiles: Rasterize heatmaps with geom_tile_rast()
Heatmaps also have similar issues with the default vectorized formats:
library(ggplot2)
library(ggrastr)
coords <- expand.grid(1:500, 1:500)
coords$Value <- 1 / apply(as.matrix(coords), 1, function(x) sum((x - c(50, 50))^2)^0.01)
gg_tile_vec <- ggplot(coords) + geom_tile(aes(x=Var1, y=Var2, fill=Value))
gg_tile_rast <- ggplot(coords) + rasterise(geom_tile(aes(x=Var1, y=Var2, fill=Value)), dpi = 300, scale = 1)
print(gg_tile_rast)
The legacy function geom_tile_rast() used the following syntax:
gg_tile_rast <- ggplot(coords) + geom_tile_rast(aes(x=Var1, y=Var2, fill=Value))
print(gg_tile_rast)
Note that we can see that the rasterized plots using ggrastr are lighter in size when rendered to pdf:
PrintFileSize(gg_tile_rast, 'Raster')
#> Raster: 46.77637 Kb.
PrintFileSize(gg_tile_vec, 'Vector')
#> Vector: 817.8398 Kb.
Violin plots: Rasterize violin plots with geom_violin_rast()
One can see a similar effect with violin plots:
library(ggplot2)
library(ggrastr)
gg_violin_vec <- ggplot(mtcars, aes(factor(cyl), mpg)) + geom_violin()
gg_violin_rast <- ggplot(mtcars) + rasterise(geom_violin(aes(factor(cyl), mpg)))
print(gg_violin_rast)
The legacy function geom_violin_rast() had the following syntax:
gg_violin_rast <- ggplot(mtcars) + geom_violin_rast(aes(factor(cyl), mpg))
print(gg_violin_rast)
## difference in size shown
PrintFileSize(gg_tile_rast, 'Raster')
#> Raster: 46.77637 Kb.
PrintFileSize(gg_tile_vec, 'Vector')
#> Vector: 817.8398 Kb.
Box plots: Jitter outliers and rasterize box plots with geom_boxplot_jitter
Another type of plot with a potentially large number of small objects is geom_boxplot:
library(ggplot2)
library(ggrastr)
points_num <- 5000
df <- data.frame(x=as.factor(1:points_num %% 2), y=log(abs(rcauchy(points_num))))
gg <- ggplot(df, aes(x=x, y=y)) + scale_color_discrete(guide="none")
boxplot <- gg + geom_boxplot()
print(boxplot)
With a large number of objects, outlier points become noninformative. For example, here is the rendered plot with points_num <- 1000000.
For such a large number of points, it would be better to jitter them using geom_boxplot_jitter():
library(ggplot2)
library(ggrastr)
points_num <- 500000
df <- data.frame(x=as.factor(1:points_num %% 2), y=log(abs(rcauchy(points_num))))
gg <- ggplot(df, aes(x=x, y=y)) + scale_color_discrete(guide="none")
gg_box_vec <- gg + geom_boxplot_jitter(outlier.size=0.1, outlier.jitter.width=0.3, outlier.alpha=0.5)
print(gg_box_vec)
And this geom can be rasterized as well:
gg_box_rast <- gg + geom_boxplot_jitter(outlier.size=0.1, outlier.jitter.width=0.3, outlier.alpha=0.5, raster.dpi=200)
print(gg_box_rast)
PrintFileSize(gg_box_rast, 'Raster')
#> Raster: 122.5781 Kb.
PrintFileSize(gg_box_vec, 'Vector')
#> Vector: 233.0508 Kb.
Beeswarm-style plots: geom_beeswarm_rast and geom_quasirandom
ggrastr also allows users to create rasterized beeswarm plots. As described in the README for ggbeeswarm,
Beeswarm plots (aka column scatter plots or violin scatter plots) are a way of plotting points that would ordinarily overlap so that they fall next to each other instead. In addition to reducing overplotting, it helps visualize the density of the data at each point (similar to a violin plot), while still showing each data point individually. The ggrastr geom geom_beeswarm_rast is similar to ggbeeswarm::geom_beeswarm(), but it provides a rasterized layer:
library(ggplot2)
library(ggrastr)
ggplot(mtcars) + geom_beeswarm_rast(aes(x = factor(cyl), y=mpg), raster.dpi=600, cex=1.5)
Again, we strongly encourage users to simply use rasterise():
library(ggplot2)
library(ggrastr)
library(ggbeeswarm,)
ggplot(mtcars) + rasterise(geom_beeswarm(aes(x = factor(cyl), y=mpg)))
Analogously, the legacy wrapper geom_quasirandom_rast() is much like ggbeeswarm::geom_quasirandom(), but with a rasterized layer:
library(ggplot2)
library(ggrastr)
ggplot(mtcars) + geom_quasirandom_rast(aes(x = factor(cyl), y=mpg), raster.dpi=600)
We encourage users to visit both CRAN and the GitHub repo for ggbeeswam for more details.