Alpha appears to be ignored if it's constant for all datapoints:

(why does github treat no text and enter as a valid issue… sorry about that).

An alpha aes is ignored when all instances have the same value. Observe:

p = plotnine.ggplot()
p += plotnine.geom_rect(
    plotnine.aes(xmin='xmin',xmax='xmax',ymin='ymin',ymax='ymax',fill='fill',
                alpha='alpha'),
          data = pd.DataFrame({
              'xmin': [0,0],
              'xmax': [1,0.5],
              'ymin': [0,0],
              'ymax': [1,0.5],
              'fill': ['red', 'blue'],
              'alpha': [0.5, 0.51]
          }),)
p += plotnine.scale_fill_identity()
p

Produces this, as expected.

But replacing that second alpha value with 0.5:

p = plotnine.ggplot()
p += plotnine.geom_rect(
    plotnine.aes(xmin='xmin',xmax='xmax',ymin='ymin',ymax='ymax',fill='fill',
                alpha='alpha'),
          data = pd.DataFrame({
              'xmin': [0,0],
              'xmax': [1,0.5],
              'ymin': [0,0],
              'ymax': [1,0.5],
              'fill': ['red', 'blue'],
              'alpha': [0.5, 0.50]
          }),)
p += plotnine.scale_fill_identity()
p

I suspect this is also the underlying issue with #50 and #43 - arguably setting an aes to a scalar is something a user would expect - it’s documented to work for x and y in the API. But please note that this is a different case - we have a valid aes mapping that just happens to contain the same value for all instances and it’s completely ignored 😦.

[TyberiusPrime]

Hassan, I have the deepest respect for your work on plotnine. It is a truly beautiful library. May I suggest that “If you have any specific values, put them outside the aes() call.” would be better phrased as ‘if you have specific values, use an identity mapping’. This is the more general, and to my mind, more useful advice.

Yes that is the issue. This specific is not about discrete or continuous distinctions.

Indeed, it is about sensible default mappings.

Having scale_alpha_continuous(range(0.1, 1)) as the default scale trips new users, as evidenced by multiple bug reports.

How about switching to scale_alpha_identity() by default. It does not change ‘the meaning of what it means to map to a scale’ and it will explode with a (semi-useful) error message if you pass in something that needs one of the ‘true’ mappings.

But separately from the issue of the default scale for alpha, why does training scale_alpha_continuous((0.1,1)) on a single value map to an alpha of 1, while a scale_x_continuous will put your value in the middle of the plot?

I’d have expected it to map to 0.5.

[has2k1]

I think rescale_pal is clear. The documentation that needs more examples is that for rescale, rescale_mid and rescale_max.