Improve support for datetime (and categorical) data in relational plo…

…ts (#2138)

* Improve support for datetime (and categorical) data in relational plots

Fixes #2130

* Disable pandas unit conversion during testing

* Force matplotlib date converters in test fixture

doc/releases/v0.11.0.txt

@@ -60,6 +60,8 @@ Other
 
 - Improved :meth:`FacetGrid.set_titles` with `margin titles=True`, such that texts representing the original row titles are removed before adding new ones. GH2083
 
+- Improved support for datetime variables in :func:`scatterplot` and :func:`lineplot`. GH2138
+
 - Fixed a bug where :func:`lineplot` did not pass the ``linestyle`` parameter down to matplotlib. GH2095
 
 - Improved the error messages produced when categorical plots process the orientation parameter.

seaborn/_core.py

@@ -920,7 +920,11 @@ def _semantic_subsets(
     def comp_data(self):
         """Dataframe with numeric x and y, after unit conversion and log scaling."""
         if not hasattr(self, "ax"):
-            raise AttributeError("No Axes attached to plotter")
+            # Probably a good idea, but will need a bunch of tests updated
+            # Most of these tests should just use the external interface
+            # Then this can be reeneabled.
+            # raise AttributeError("No Axes attached to plotter")
+            return self.plot_data
 
         if not hasattr(self, "_comp_data"):
 

seaborn/conftest.py

@@ -1,10 +1,24 @@
 import numpy as np
 import pandas as pd
+import datetime
+import matplotlib as mpl
 import matplotlib.pyplot as plt
 
 import pytest
 
 
+@pytest.fixture(scope="session", autouse=True)
+def remove_pandas_unit_conversion():
+    # Prior to pandas 1.0, it registered its own datetime converters,
+    # but they are less powerful than what matplotlib added in 2.2,
+    # and we rely on that functionality in seaborn.
+    # https://github.com/matplotlib/matplotlib/pull/9779
+    # https://github.com/pandas-dev/pandas/issues/27036
+    mpl.units.registry[np.datetime64] = mpl.dates.DateConverter()
+    mpl.units.registry[datetime.date] = mpl.dates.DateConverter()
+    mpl.units.registry[datetime.datetime] = mpl.dates.DateConverter()
+
+
 @pytest.fixture(autouse=True)
 def close_figs():
     yield
@@ -138,7 +152,7 @@ def long_df(rng):
         a=rng.choice(list("abc"), n),
         b=rng.choice(list("mnop"), n),
         c=rng.choice([0, 1], n, [.3, .7]),
-        t=rng.choice(np.arange("2005-02-25", "2005-02-28", dtype="datetime64[D]"), n),
+        t=rng.choice(np.arange("2004-07-30", "2007-07-30", dtype="datetime64[Y]"), n),
         s=rng.choice([2, 4, 8], n),
         f=rng.choice([0.2, 0.3], n),
     ))

seaborn/relational.py

@@ -288,7 +288,9 @@ def plot(self, ax, kws):
 
         # Loop over the semantic subsets and add to the plot
         grouping_semantics = "hue", "size", "style"
-        for sub_vars, sub_data in self._semantic_subsets(grouping_semantics):
+        for sub_vars, sub_data in self._semantic_subsets(
+            grouping_semantics, from_comp_data=True
+        ):
 
             if self.sort:
                 sub_data = sub_data.sort_values(["units", "x", "y"])
@@ -641,6 +643,8 @@ def lineplot(
     if ax is None:
         ax = plt.gca()
 
+    p._attach(ax)
+
     p.plot(ax, kwargs)
     return ax
 
@@ -920,6 +924,8 @@ def scatterplot(
     if ax is None:
         ax = plt.gca()
 
+    p._attach(ax)
+
     p.plot(ax, kwargs)
 
     return ax

seaborn/tests/test_core.py

@@ -897,8 +897,10 @@ def test_comp_data(self, long_df):
 
         p = VectorPlotter(data=long_df, variables={"x": "x", "y": "t"})
 
-        with pytest.raises(AttributeError):
-            p.comp_data
+        # We have disabled this check for now, while it remains part of
+        # the internal API, because it will require updating a number of tests
+        # with pytest.raises(AttributeError):
+        #     p.comp_data
 
         _, ax = plt.subplots()
         p._attach(ax)

seaborn/tests/test_relational.py

@@ -1326,6 +1326,9 @@ def test_lineplot_smoke(
         lineplot(x="x", y=long_df.y.values, data=long_df)
         ax.clear()
 
+        lineplot(x="x", y="t", data=long_df)
+        ax.clear()
+
         lineplot(x="x", y="y", hue="a", data=long_df)
         ax.clear()
 
@@ -1692,6 +1695,13 @@ def test_linewidths(self, long_df):
         scatterplot(data=long_df, x="x", y="y", linewidth=lw)
         assert ax.collections[0].get_linewidths().item() == lw
 
+    def test_datetime_scale(self, long_df):
+
+        ax = scatterplot(data=long_df, x="t", y="y")
+        # Check that we avoid weird matplotlib default auto scaling
+        # https://github.com/matplotlib/matplotlib/issues/17586
+        ax.get_xlim()[0] > ax.xaxis.convert_units(np.datetime64("2002-01-01"))
+
     def test_scatterplot_vs_relplot(self, long_df, long_semantics):
 
         ax = scatterplot(data=long_df, **long_semantics)