element_text size using vector

[mpgerstl]

Hi,

When using a vector for axis text size in a flipped bar chart, the bars seems to be moved far outside the plot region, and also the text is shifted. I am using ggplot2 version 3.2.1.
I was able to produce the wished plot with ggplot2 version 3.1.1

library(tidyverse)
library(tibble)
ds <- mtcars %>%
  rownames_to_column() %>%
  select(car = rowname, mpg)

ds_size = rep(8, nrow(ds))
ds_size[10] <- 12

ds %>%
  ggplot(aes(x = car, y = mpg)) +
  coord_flip() +
  theme(axis.text.y = element_text(size = ds_size))

Thank you

[batpigandme]

Would you mind turning this into a reprex (short for minimal reproducible example)? It's especially helpful for ggplot2, since it automatically generates and uploads the plots (or lack thereof), which makes it much easier to go through at a glance.

If you've never heard of a reprex before, you might want to start by reading the tidyverse.org help page.

Right now the best way to install reprex is:

# install.packages("devtools")
devtools::install_github("tidyverse/reprex")

If you run into problems with access to your clipboard, you can specify an outfile for the reprex, and then copy and paste (or drag and drop) the contents here.

reprex::reprex(input = "fruits_stringdist.R", outfile = "fruits_stringdist.md")

Thanks

[mpgerstl]

Thank you for the hint with the clipboard

library(tidyverse)
library(tibble)
ds <- mtcars %>%
  rownames_to_column() %>%
  select(car = rowname, mpg)

ds_size = rep(8, nrow(ds))
ds_size[10] <- 12

ds %>%
  ggplot(aes(x = car, y = mpg)) +
  coord_flip() +
  theme(axis.text.y = element_text(size = ds_size))

^{Created on 2019-08-19 by the reprex package (v0.3.0)}

[batpigandme]

Hmm, I get the same results you've shown above with the dev version.

Here it is with ggplot2 3.0.0

library(tidyverse)

ds <- mtcars %>%
  rownames_to_column() %>%
  select(car = rowname, mpg)

ds_size = rep(8, nrow(ds))
ds_size[10] <- 12

ds %>%
  ggplot(aes(x = car, y = mpg)) +
  coord_flip() +
  theme(axis.text.y = element_text(size = ds_size))

---

And below is with 3.1.1

library(tidyverse)

ds <- mtcars %>%
  rownames_to_column() %>%
  select(car = rowname, mpg)

ds_size = rep(8, nrow(ds))
ds_size[10] <- 12

ds %>%
  ggplot(aes(x = car, y = mpg)) +
  coord_flip() +
  theme(axis.text.y = element_text(size = ds_size))

^{Created on 2019-08-19 by the reprex package (v0.3.0)}

[yutannihilation]

It seems the element takes the sum of the values as margin. BTW, is element_*() supposed to accept vectors? I've seen this kind of tricks somewhere several times, but I guess none of the official docs refer to this.

library(ggplot2)

d <- data.frame(x = letters[1:10], y = 1:10, stringsAsFactors = FALSE)

f <- function(n) {
  ggplot(d) +
    geom_point(aes(x, y)) +
    theme(axis.text.x = element_text(size = rep(8, n)))
}

f(1)

f(10)

f(100)

^{Created on 2019-08-19 by the reprex package (v0.3.0)}

[mpgerstl]

@batpigandme Sorry, I forgot the geometry in my reprex. Here is a new one

library(ggplot2)

ds <- data.frame(
  car = rownames(mtcars),
  mpg = mtcars$mpg
)

ds_size = rep(8, nrow(ds))
ds_size[10] <- 12

ggplot(ds, ggplot2::aes(x = car, y = mpg)) +
  geom_col() +
  coord_flip() +
  theme(axis.text.y = element_text(size = ds_size))

^{Created on 2019-08-19 by the reprex package (v0.3.0)}

@yutannihilation I cannot remember where I got the hint, but using a vector for face still works

library(ggplot2)

ds <- data.frame(
  car = rownames(mtcars),
  mpg = mtcars$mpg
)

ds_face = rep("plain", nrow(ds))
ds_face[10] <- "bold"

ggplot(ds, ggplot2::aes(x = car, y = mpg)) +
  geom_col() +
  coord_flip() +
  theme(axis.text.y = element_text(face = ds_face))

^{Created on 2019-08-19 by the reprex package (v0.3.0)}

Thank you

[yutannihilation]

Yeah, it works and probably should, but I feet the individual specifications should belong to the scale, not to the theme.

@clauswilke
What do you think? I expect this kind of tricks will be unnecessary once gridtext gets on CRAN.

[clauswilke]

I'm not a big fan of setting text parameters as vectors in this way, mostly because it's unpredictable when it works and when it doesn't. It works for axes but it doesn't work for legends, for internal technical reasons that are not visible to the enduser.

Having said that, I don't understand why this specific example fails. We should track down the cause, because it suggests something weird is going on with the size argument.

[yutannihilation]

Thanks, agreed. Let's see what's happening here.

[clauswilke]

Here is what I think is happening. (I don't quite understand yet why it only happens for font size, and not for font face, so maybe I'm wrong.)

This line returns now a vector of descent values rather than just one value:

ggplot2/R/margins.R

Line 69 in 1ae034e

descent <- font_descent(gp$fontfamily, gp$fontface, gp$fontsize, gp$cex)

ggplot2:::font_descent("", "plain", rep(10, 5), 1)
#> [1] 0.0292290581597222inches 0.0292290581597222inches
#> [3] 0.0292290581597222inches 0.0292290581597222inches
#> [5] 0.0292290581597222inches

^{Created on 2019-08-19 by the reprex package (v0.3.0)}

This vector results in vectors text_width and text_height, which are given to add_margins() here:

ggplot2/R/margins.R

Lines 197 to 205 in 1ae034e

	add_margins(
	grob = grob_details$text_grob,
	height = grob_details$text_height,
	width = grob_details$text_width,
	gp = gp,
	margin = margin,
	margin_x = margin_x,
	margin_y = margin_y
	)

And the add_margins() function then build layouts assuming width and height are single values rather than vectors:

ggplot2/R/margins.R

Lines 115 to 121 in 1ae034e

	widths <- unit.c(margin[4], width, margin[2])
	heights <- unit.c(margin[1], height, margin[3])
	vp <- viewport(
	layout = grid.layout(3, 3, heights = heights, widths = widths),
	gp = gp
	)

Again, this doesn't yet fully explain what exactly is happening, but I think it's on the right track. Importantly, it's based on a line that was changed with the latest release:
f5a88a7#diff-004d6436e77ba3e9705ab23890dace32

The previous function descentDetails() produced the wrong results, but it didn't create vectors.

[yutannihilation]

Thanks. For facet strips, it seems we do the right thing to pass a single height and width to add_margins() even when height and/or width is a vector.

ggplot2/R/labeller.r

Lines 608 to 631 in b842024

	if (horizontal) {
	height <- max_height(lapply(grobs, function(x) x$text_height))
	width <- unit(1, "null")
	} else {
	height <- unit(1, "null")
	width <- max_width(lapply(grobs, function(x) x$text_width))
	}
	# Add margins around text grob
	grobs <- apply(
	grobs,
	c(1, 2),
	function(x) {
	add_margins(
	grob = x[[1]]$text_grob,
	height = height,
	width = width,
	gp = gp,
	margin = element$margin,
	margin_x = TRUE,
	margin_y = TRUE
	)
	}
	)

edit:

But, facet strips doesn't accept variable sizes...

library(ggplot2)

d <- data.frame(x = letters[1:10], y = 1:10, stringsAsFactors = FALSE)

ggplot(d) +
  geom_point(aes(x, y)) +
  facet_wrap(vars(x)) +
  theme(strip.text = element_text(size = 1:10 * 5))

^{Created on 2019-08-20 by the reprex package (v0.3.0)}

[clauswilke]

I made a PR that seems to fix the problem, though I still don't fully understand how exactly things go wrong. #3493

[yutannihilation]

Confirmed #3493 fixes the issue. But, I too don't understand how the things are supposed to work here...

Is font_descent() really intended to work with multiple sizes? I see it returns the same values for different sizes in the vector (Does this happen only on Windows?).

ggplot2:::font_descent("", "plain", 10, 1)
#> [1] 0.03125inches
ggplot2:::font_descent("", "plain", c(10, 10), 1)
#> [1] 0.03125inches 0.03125inches

ggplot2:::font_descent("", "plain", c(10, 20), 1)
#> [1] 0.03125inches 0.03125inches
ggplot2:::font_descent("", "plain", 1:10 * 10, 1)
#>  [1] 0.03125inches 0.03125inches 0.03125inches 0.03125inches 0.03125inches
#>  [6] 0.03125inches 0.03125inches 0.03125inches 0.03125inches 0.03125inches

# probably the cache works wrong here...
ggplot2:::font_descent("", "plain", 1:10 * 20, 1)
#>  [1] 0.03125inches            0.03125inches           
#>  [3] 0.03125inches            0.03125inches           
#>  [5] 0.03125inches            0.0520833333333333inches
#>  [7] 0.0520833333333333inches 0.0520833333333333inches
#>  [9] 0.0520833333333333inches 0.0520833333333333inches

^{Created on 2019-08-20 by the reprex package (v0.3.0)}

[clauswilke]

Yes, the font cache is not working correctly for vectorized input. It loops over the keys but then uses the same parameters for descent lookup each time:

ggplot2/R/margins.R

Lines 339 to 355 in b842024

	descents <- lapply(key, function(k) {
	descent <- descent_cache[[k]]
	if (is.null(descent)) {
	descent <- convertHeight(grobDescent(textGrob(
	label = "gjpqyQ",
	gp = gpar(
	fontsize = size,
	cex = cex,
	fontfamily = family,
	fontface = face
	)
	)), 'inches')
	descent_cache[[k]] <- descent
	}
	descent
	})

@thomasp85 Do you have any strong opinions on how to fix this? Since font_descent() is only used by title_spec() and that is not vectorized, maybe it's best to let font_descent() return only the first result? In general, the fact that title_spec() is not properly vectorized causes all sorts of problems, but fixing that is a pretty substantial rewrite, and probably better left to something like the gridtext package.