Default norm when using `levels` in contour(f) produce non-intuitive behavior.

[Ockenfuss]

Ultraplot ignores the vmin/vmax values, if the levels are explicitly set in contour plots.
This is documented behaviour.
However, I can't see the reason why this is the case?
Matplotlib is not ignoring vmin/vmax in contour plots, if they are explicitly set.
Only, if they are not given, the color boundaries are derived from the levels list.

from matplotlib.colors import Normalize
import matplotlib as mpl
import matplotlib.pyplot as plt
import numpy as np
import ultraplot as uplt

fig, ax=uplt.subplots()
x=np.linspace(-1,1,100)
y=np.linspace(-1,1,100)
X,Y=np.meshgrid(x,y)
Z=np.exp(-(X**2+Y**2))
im=ax.pcolormesh(X,Y,Z, cmap='turbo', alpha=0.6)
cs=ax.contour(X,Y,Z, levels=[0.3, 0.6, 0.9], cmap='turbo', vmin=0, vmax=1) # vmin/max are ignored here

[Ockenfuss]

Also when passing colors explicitly, the colors do not match the background (in matplotlib, this code works as expected):

from matplotlib.colors import Normalize
import matplotlib as mpl
import matplotlib.pyplot as plt
import numpy as np
import ultraplot as uplt

fig, ax=uplt.subplots()
x=np.linspace(-1,1,100)
y=np.linspace(-1,1,100)
X,Y=np.meshgrid(x,y)
Z=np.exp(-(X**2+Y**2))
im=ax.pcolormesh(X,Y,Z, cmap='turbo', alpha=0.7, vmin=0, vmax=1)
norm = Normalize(vmin=0, vmax=1)
turbo=mpl.colormaps.get_cmap('turbo')
colors= turbo(norm([0.3, 0.6, 0.9]))
cs=ax.contour(X,Y,Z, levels=[0.3, 0.6, 0.9], colors=colors, linewidths=1)

[cvanelteren]

Thanks for reporting! This is indeed a grave bug. The countourf function does parse the colors correctly. Will have a look what the difference is with contour

[cvanelteren]

The issue is in the normalizer

fig, ax = uplt.subplots()
x = np.linspace(-1, 1, 100)
y = np.linspace(-1, 1, 100)
X, Y = np.meshgrid(x, y)
Z = np.exp(-(X**2 + Y**2))
im = ax.pcolormesh(X, Y, Z, cmap="turbo", alpha=0.7, colorbar="right")
cs = ax.contour(
    X, Y, Z, levels=[0.3, 0.6, 0.9], cmap="turbo", alpha=1, lw=2, norm=im.norm
)

[Ockenfuss]

I was suspecting it has to do with discrete=False or True. I should have mentioned that I set cmap.discrete: False in my ultraplotrc, sorry for that. However, setting ax.contour(..., discrete=False) gives me a warning. Maybe, contour is always using discrete=True internally?

[cvanelteren]

The default is set to True I believe.

The source states

        # Determine the appropriate 'vmin', 'vmax', and/or 'levels'
        # NOTE: Unlike xarray, but like matplotlib, vmin and vmax only approximately
        # determine level range. Levels are selected with Locator.tick_values().
        levels = None  # unused

[cvanelteren]

Setting the normalizer like I did above will get you colors but not the correct levels. This comment I noticed before in the source, but find it a bit puzzling. Why approximate if the users specifies the levels.

[cvanelteren]

The issue is in the combination with the levels and the norm. Turning off the levels will give you the correct colors for the contour lines. Will look into this 👀

[cvanelteren]

I know what the issue is.

The issue is that when levels is given, a norm is created with vmin/vmax=min/max(levels). In the case above, this creates a new mapping of the colormap to parse the range [0,1] into 3 pieces. The fact that parsing the other norm in confirms this. So the colors are not aligned within the same mapping. I am not sure if this is a bug or a feature, and also not sure yet how to address this directly.

[cvanelteren]

For example

fig, ax = uplt.subplots()
x = np.linspace(-1, 1, 100)
y = np.linspace(-1, 1, 100)
X, Y = np.meshgrid(x, y)
Z = np.exp(-(X**2 + Y**2))

levels = [0.3, 0.6, 0.9]
im = ax.pcolormesh(
   X,
   Y,
   Z,
   cmap="turbo",
   alpha=0.8,
   colorbar="right",
   levels=levels,
)
cs = ax.contour(
   X,
   Y,
   Z,
   levels=levels,
   cmap="turbo",
   alpha=1,
   lw=2,
   norm_kw=dict(vmin=Z.min(), vmax=Z.max()),
   discrete=False,
   # norm=im.norm,
)

[cvanelteren]

I traced how vmin and vmax is handled. This call self._parse_level_vals` will override the vmin and vmax based on levels. I tried overriding it with vmin and vmax from the data but somehow that is not able to remove the labels to the correct range.

[cvanelteren]

Ok the issue is caused by pcolors.DiscreteNorm; it uses the levels only and ignores the data range -- again not sure if it is a bug or a feature. When commenting out the block at 3363 in plot.py we get a normal matplotlib object that works as expected.

[cvanelteren]

@Ockenfuss not sure how to proceed here. I think what you want is something like this

boundaries = np.concatenate(([Z.min()], levels, [Z.max()]))
norm = mcolors.BoundaryNorm(boundaries=boundaries, ncolors=256)
cs = ax.contour(
    X,
    Y,
    Z,
    levels=levels,
    cmap="turbo",
    alpha=1,
    lw=2,
    norm=norm,
)

UltraPlot currently divides the space completely based on the levels, showing a region of interest that when combined with a larger space will not match. We could either make the default behavior the boundary norm or make it a user input like above.