Split violin plot with ggplot2

I'd like to create a split violin density plot using ggplot, like the fourth example on this page of the seaborn documentation.

Here is some data:

set.seed(20160229)

my_data = data.frame(
    y=c(rnorm(1000), rnorm(1000, 0.5), rnorm(1000, 1), rnorm(1000, 1.5)),
    x=c(rep('a', 2000), rep('b', 2000)),
    m=c(rep('i', 1000), rep('j', 2000), rep('i', 1000))
)

I can plot dodged violins like this:

library('ggplot2')

ggplot(my_data, aes(x, y, fill=m)) +
  geom_violin()

But it's hard to visually compare the widths at different points in the side-by-side distributions. I haven't been able to find any examples of split violins in ggplot - is it possible?

I found a base R graphics solution but the function is quite long and I want to highlight distribution modes, which are easy to add as additional layers in ggplot but will be harder to do if I need to figure out how to edit that function.

Or, to avoid fiddling with the densities, you could extend ggplot2's GeomViolin like this:

GeomSplitViolin <- ggproto("GeomSplitViolin", GeomViolin, 
						   draw_group = function(self, data, ..., draw_quantiles = NULL) {
  data <- transform(data, xminv = x - violinwidth * (x - xmin), xmaxv = x + violinwidth * (xmax - x))
  grp <- data[1, "group"]
  newdata <- plyr::arrange(transform(data, x = if (grp %% 2 == 1) xminv else xmaxv), if (grp %% 2 == 1) y else -y)
  newdata <- rbind(newdata[1, ], newdata, newdata[nrow(newdata), ], newdata[1, ])
  newdata[c(1, nrow(newdata) - 1, nrow(newdata)), "x"] <- round(newdata[1, "x"])

  if (length(draw_quantiles) > 0 & !scales::zero_range(range(data$y))) {
	stopifnot(all(draw_quantiles >= 0), all(draw_quantiles <=
	  1))
	quantiles <- ggplot2:::create_quantile_segment_frame(data, draw_quantiles)
	aesthetics <- data[rep(1, nrow(quantiles)), setdiff(names(data), c("x", "y")), drop = FALSE]
	aesthetics$alpha <- rep(1, nrow(quantiles))
	both <- cbind(quantiles, aesthetics)
	quantile_grob <- GeomPath$draw_panel(both, ...)
	ggplot2:::ggname("geom_split_violin", grid::grobTree(GeomPolygon$draw_panel(newdata, ...), quantile_grob))
  }
  else {
	ggplot2:::ggname("geom_split_violin", GeomPolygon$draw_panel(newdata, ...))
  }
})

geom_split_violin <- function(mapping = NULL, data = NULL, stat = "ydensity", position = "identity", ..., 
							  draw_quantiles = NULL, trim = TRUE, scale = "area", na.rm = FALSE, 
							  show.legend = NA, inherit.aes = TRUE) {
  layer(data = data, mapping = mapping, stat = stat, geom = GeomSplitViolin, 
		position = position, show.legend = show.legend, inherit.aes = inherit.aes, 
		params = list(trim = trim, scale = scale, draw_quantiles = draw_quantiles, na.rm = na.rm, ...))
}

And use the new geom_split_violin like this:

ggplot(my_data, aes(x, y, fill = m)) + geom_split_violin()

Note: I think the answer by jan-glx is much better, and most people should use that instead. But sometimes, the manual approach is still helpful to do weird things.

---

You can achieve this by calculating the densities yourself beforehand, and then plotting polygons. See below for a rough idea.

Get densities

library(dplyr)
pdat <- my_data %>%
  group_by(x, m) %>%
  do(data.frame(loc = density(.$y)$x,
                dens = density(.$y)$y))

Flip and offset densities for the groups

pdat$dens <- ifelse(pdat$m == 'i', pdat$dens * -1, pdat$dens)
pdat$dens <- ifelse(pdat$x == 'b', pdat$dens + 1, pdat$dens)

Plot

ggplot(pdat, aes(dens, loc, fill = m, group = interaction(m, x))) + 
  geom_polygon() +
  scale_x_continuous(breaks = 0:1, labels = c('a', 'b')) +
  ylab('density') +
  theme_minimal() +
  theme(axis.title.x = element_blank())

Result