Fix boxplot stat unscaling

seaborn/categorical.py

@@ -626,7 +626,12 @@ def get_props(element, artist=mpl.lines.Line2D):
             capwidth = plot_kws.get("capwidths", 0.5 * data["width"])
 
             self._invert_scale(ax, data)
-            linear_scale = getattr(ax, f"get_{self.orient}scale")() == "linear"
+            _, inv = _get_transform_functions(ax, value_var)
+            for stat in ["mean", "med", "q1", "q3", "cilo", "cihi", "whislo", "whishi"]:
+                stats[stat] = inv(stats[stat])
+            stats["fliers"] = stats["fliers"].map(inv)
+
+            linear_orient_scale = getattr(ax, f"get_{self.orient}scale")() == "linear"
 
             maincolor = self._hue_map(sub_vars["hue"]) if "hue" in sub_vars else color
             if fill:
@@ -652,7 +657,7 @@ def get_props(element, artist=mpl.lines.Line2D):
                 bxpstats=stats.to_dict("records"),
                 positions=data[self.orient],
                 # Set width to 0 to avoid going out of domain
-                widths=data["width"] if linear_scale else 0,
+                widths=data["width"] if linear_orient_scale else 0,
                 patch_artist=fill,
                 vert=self.orient == "x",
                 manage_ticks=False,
@@ -674,7 +679,7 @@ def get_props(element, artist=mpl.lines.Line2D):
             # Reset artist widths after adding so everything stays positive
             ori_idx = ["x", "y"].index(self.orient)
 
-            if not linear_scale:
+            if not linear_orient_scale:
                 for i, box in enumerate(data.to_dict("records")):
                     p0 = box["edge"]
                     p1 = box["edge"] + box["width"]

tests/test_categorical.py

@@ -806,13 +806,13 @@ def check_box(self, bxp, data, orient, pos, width=0.8):
         p25, p50, p75 = np.percentile(data, [25, 50, 75])
 
         box = self.get_box_verts(bxp.box)
-        assert box[val_idx].min() == p25
-        assert box[val_idx].max() == p75
+        assert box[val_idx].min() == approx(p25, 1e-3)
+        assert box[val_idx].max() == approx(p75, 1e-3)
         assert box[pos_idx].min() == approx(pos - width / 2)
         assert box[pos_idx].max() == approx(pos + width / 2)
 
         med = bxp.median.get_xydata().T
-        assert tuple(med[val_idx]) == (p50, p50)
+        assert np.allclose(med[val_idx], (p50, p50), rtol=1e-3)
         assert np.allclose(med[pos_idx], (pos - width / 2, pos + width / 2))
 
     def check_whiskers(self, bxp, data, orient, pos, capsize=0.4, whis=1.5):
@@ -831,13 +831,13 @@ def check_whiskers(self, bxp, data, orient, pos, capsize=0.4, whis=1.5):
         adj_lo = data[data >= (p25 - iqr * whis)].min()
         adj_hi = data[data <= (p75 + iqr * whis)].max()
 
-        assert whis_lo[val_idx].max() == p25
+        assert whis_lo[val_idx].max() == approx(p25, 1e-3)
         assert whis_lo[val_idx].min() == approx(adj_lo)
         assert np.allclose(whis_lo[pos_idx], (pos, pos))
         assert np.allclose(caps_lo[val_idx], (adj_lo, adj_lo))
         assert np.allclose(caps_lo[pos_idx], (pos - capsize / 2, pos + capsize / 2))
 
-        assert whis_hi[val_idx].min() == p75
+        assert whis_hi[val_idx].min() == approx(p75, 1e-3)
         assert whis_hi[val_idx].max() == approx(adj_hi)
         assert np.allclose(whis_hi[pos_idx], (pos, pos))
         assert np.allclose(caps_hi[val_idx], (adj_hi, adj_hi))
@@ -944,6 +944,18 @@ def test_dodge_native_scale_log(self, long_df):
                 widths.append(np.ptp(coords))
         assert np.std(widths) == approx(0)
 
+    @pytest.mark.parametrize("orient", ["x", "y"])
+    def test_log_data_scale(self, long_df, orient):
+
+        var = {"x": "y", "y": "x"}[orient]
+        s = long_df["z"]
+        ax = mpl.figure.Figure().subplots()
+        getattr(ax, f"set_{var}scale")("log")
+        boxplot(**{var: s}, whis=np.inf, ax=ax)
+        bxp = ax.containers[0][0]
+        self.check_box(bxp, s, orient, 0)
+        self.check_whiskers(bxp, s, orient, 0, whis=np.inf)
+
     def test_color(self, long_df):
 
         color = "#123456"