awkward behavior from scale_fill_viridis_c(trans = "log")

[twest820]

Three interrelated issues:

1. Log transformation of zero values for a continuous fill produces a warning about infinite values in a discrete y axis. This warning could be more clearly worded since the fill is neither discrete nor a y axis. There also appears to be differentiation of NaN from -Inf, leading to a cryptic NaN warning from negative fill values.
2. There doesn't appear to be any way to tell ggplot these fill values are expected in order to suppress the warnings. The closest I can see is something like scale_fill_viridis_c(na.value = "transparent", trans = "log"), which plots fine since apparently -Inf, NaN, and NA are treated interchangeably but is orthogonal to warning generation.
3. The resulting labels on the fill colorbar include 8+ decimal places, which is a curiously verbose default, though format_format(digits = n) provides a simple mitigation.

library(ggplot2)
library(magrittr)
library(scales)
library(tidyverse)

data = crossing(x = seq(0, 1, length.out = 100), y = seq(0, 1, length.out = 100))
data %<>% mutate(value = x - y)
ggplot(data) + geom_raster(aes(x = x, y = y, fill = value)) + scale_fill_viridis_c(trans = "log")

Warning messages:
1: In self$trans$transform(x) : NaNs produced
2: Transformation introduced infinite values in discrete y-axis 

# change for more concise labeling
ggplot(data) + geom_raster(aes(x = x, y = y, fill = value)) + scale_fill_viridis_c(labels = format_format(digits = 2), trans = "log")

Observed on R 4.0.2 with tidyverse 1.3.0 (ggplot 3.3.2). I suspect this might also occur with other scale_fill functions.

[EmilHvitfeldt]

Some of the issues you are getting are related to what happens when you take logarithms of non-positive numbers.

seq(-1, 1, by = 0.5)
#> [1] -1.0 -0.5  0.0  0.5  1.0

log(seq(-1, 1, by = 0.5))
#> Warning in log(seq(-1, 1, by = 0.5)): NaNs produced
#> [1]        NaN        NaN       -Inf -0.6931472  0.0000000

The NaN's appear because logarithms for negative numbers are undefined. NaN is generally treated differently than NA.
We are getting infinite values after transformation since log(0) = -Inf.

[teunbrand]

I also ran into this and I think these lines are responsible:

ggplot2/R/scale-.r

Lines 597 to 598 in aaaec2f

	axis <- if ("x" %in% self$aesthetics) "x" else "y"
	check_transformation(x, new_x, self$scale_name, axis)

Which is fine for position scales, but generates these awkward messages for other types of scales.

I'm not sure how safe this is, but it might be possible to replace the first line in the chunk above by

axis <- self$aesthetics[1]

At least, that is how ggplot2:::guide_train.axis() retrieves the aesthetic from scales if not explicitly given as argument.

In a similar vein, ggplot2:::check_transformation() only considers the names of position scales for generating the message itself:

ggplot2/R/scale-.r

Lines 1185 to 1191 in aaaec2f

	type <- if (name == "position_b") {
	"binned"
	} else if (name == "position_c") {
	"continuous"
	} else {
	"discrete"
	}

It think an alternative might be to consider to use the inherits(scale, "ScaleContinuous") pattern above (which should be TRUE for both position and non-position continuous scales, but not for binned or discrete scales), but the scale itself is not passed as an argument.

[clauswilke]

In general, you have to be prepared for self$aesthetics to be a vector of multiple aesthetics that you all have to consider. So any code of the form self$aesthetics[1] makes me uncomfortable. There may be places where it's fine, but it can also go spectacularly wrong.