Drop object variables/coords before saving to_netcdf or to_zarr

Author: michaelosthege

CHANGELOG.md

@@ -16,6 +16,7 @@
 * Correctly (re)order dimensions for `bfmi` and `plot_energy` ([2126](https://github.com/arviz-devs/arviz/pull/2126))
 * Fix bug with the dimension order dependency ([2103](https://github.com/arviz-devs/arviz/pull/2103))
 * Add testing module for labeller classes ([2095](https://github.com/arviz-devs/arviz/pull/2095))
+* Variables or coordinates with an `object` dtype are dropped automatically when doing `InferenceData.to_netcdf()` or `InferenceData.to_zarr()`. Log messages are emitted about these droppings. ([2134](https://github.com/arviz-devs/arviz/pull/2134))
 
 ### Deprecation
 * Removed `fill_last`, `contour` and `plot_kwargs` arguments from `plot_pair` function ([2085](https://github.com/arviz-devs/arviz/pull/2085))

arviz/data/inference_data.py

@@ -1,5 +1,6 @@
 # pylint: disable=too-many-lines,too-many-public-methods
 """Data structure for using netcdf groups with xarray."""
+import logging
 import sys
 import uuid
 import warnings
@@ -29,6 +30,7 @@
 import xarray as xr
 from packaging import version
 
+
 from ..rcparams import rcParams
 from ..utils import HtmlTemplate, _subset_list, either_dict_or_kwargs
 from .base import _extend_xr_method, _make_json_serializable, dict_to_dataset
@@ -50,6 +52,8 @@
     # https://github.com/python/mypy/issues/1153
     import json  # type: ignore
 
+_log = logging.getLogger("arviz")
+
 
 SUPPORTED_GROUPS = [
     "posterior",
@@ -411,15 +415,21 @@ def to_netcdf(
         str
             Location of netcdf file
         """
+        # We don't support saving of object-typed variables or coords.
+        # The below helper function removes them while logging it.
+        idata = drop_objects_from_inferencedata(self)
+
         mode = "w"  # overwrite first, then append
-        if self._groups_all:  # check's whether a group is present or not.
+        # check's whether a group is present or not.
+        groups_all = idata._groups_all  # pylint: disable=W0212
+        if groups_all:
             if groups is None:
-                groups = self._groups_all
+                groups = groups_all
             else:
-                groups = [group for group in self._groups_all if group in groups]
+                groups = [group for group in groups_all if group in groups]
 
             for group in groups:
-                data = getattr(self, group)
+                data = getattr(idata, group)
                 kwargs = {}
                 if compress:
                     kwargs["encoding"] = {var_name: {"zlib": True} for var_name in data.variables}
@@ -664,6 +674,10 @@ def to_zarr(self, store=None):
         except (ImportError, AssertionError) as err:
             raise ImportError("'to_zarr' method needs Zarr (2.5.0+) installed.") from err
 
+        # We don't support saving of object-typed variables or coords.
+        # The below helper function removes them while logging it.
+        idata = drop_objects_from_inferencedata(self)
+
         # Check store type and create store if necessary
         if store is None:
             store = zarr.storage.TempStore(suffix="arviz")
@@ -672,14 +686,14 @@ def to_zarr(self, store=None):
         elif not isinstance(store, MutableMapping):
             raise TypeError(f"No valid store found: {store}")
 
-        groups = self.groups()
+        groups = idata.groups()
 
         if not groups:
             raise TypeError("No valid groups found!")
 
         for group in groups:
             # Create zarr group in store with same group name
-            getattr(self, group).to_zarr(store=store, group=group, mode="w")
+            getattr(idata, group).to_zarr(store=store, group=group, mode="w")
 
         return zarr.open(store)  # Open store to get overarching group
 
@@ -2210,3 +2224,50 @@ def concat(*args, dim=None, copy=True, inplace=False, reset_dim=True):
         inference_data_dict["attrs"] = combined_attr
 
     return None if inplace else InferenceData(**inference_data_dict)
+
+
+def drop_objects_from_dataset(data: xr.Dataset) -> Tuple[xr.Dataset, List[str], List[str]]:
+    """Returns a new ``Dataset`` without object variables and coords."""
+    vals = {}
+    vals_dropped = []
+    for vname, val in data.data_vars.items():
+        if val.dtype.hasobject:
+            vals_dropped.append(vname)
+            continue
+        vals[vname] = val
+
+    coords = {}
+    coords_dropped = []
+    for cname, cval in data.coords.items():
+        if cval.dtype.hasobject:
+            coords_dropped.append(cname)
+            continue
+        coords[cname] = cval
+
+    ndata = xr.Dataset(
+        data_vars=vals,
+        coords=coords,
+        attrs=data.attrs,
+    )
+    return ndata, vals_dropped, coords_dropped
+
+
+def drop_objects_from_inferencedata(idata: InferenceData) -> InferenceData:
+    """Returns a new InferenceData without object variables and coords.
+
+    All droppings are logged at WARNING level.
+    """
+    nidata = InferenceData(attrs=idata.attrs)
+    for gname, group in idata.items():
+        ndata, vars_dropped, coords_dropped = drop_objects_from_dataset(group)
+        if "warning" in vars_dropped and "sample_stats" in gname:
+            vars_dropped.remove("warning")
+            _log.debug(
+                "Dropped 'warning' variable from '%s' group because it's dtyped object.", gname
+            )
+        if vars_dropped:
+            _log.warning("Dropped object variables from '%s' group: %s", gname, vars_dropped)
+        if coords_dropped:
+            _log.warning("Dropped object coords from '%s' group: %s", gname, coords_dropped)
+        nidata.add_groups({gname: ndata}, coords=ndata.coords, dims=ndata.dims)
+    return nidata

arviz/tests/base_tests/test_data.py

@@ -5,6 +5,7 @@
 from collections import namedtuple
 from copy import deepcopy
 from html import escape
+import logging
 from typing import Dict
 from urllib.parse import urlunsplit
 
@@ -14,6 +15,8 @@
 from xarray.core.options import OPTIONS
 from xarray.testing import assert_identical
 
+from ...data.inference_data import drop_objects_from_dataset, drop_objects_from_inferencedata
+
 from ... import (
     InferenceData,
     clear_data_home,
@@ -1466,3 +1469,85 @@ def test_subset_samples(self):
         post = extract(idata, num_samples=10)
         assert post.dims["sample"] == 10
         assert post.attrs == idata.posterior.attrs
+
+
+class TestDropObjectVariables:
+    """Tests for helper functions that drop object-typed variables/coords."""
+
+    def test_drop_variable_from_dataset(self):
+        a = np.ones((3,))
+        b = np.array([1, None, "test"])
+        ds = xr.Dataset(
+            data_vars={
+                "a": xr.DataArray(a, coords=dict(t=[0, 1, 2]), dims=("t",)),
+                "b": xr.DataArray(b, coords=dict(t=[0, 1, 2]), dims=("t",)),
+            }
+        )
+        new, dropped_vars, dropped_coords = drop_objects_from_dataset(ds)
+        assert isinstance(new, xr.Dataset)
+        assert new is not ds
+        assert "a" in new
+        assert "b" not in new
+        assert dropped_vars == ["b"]
+        assert not dropped_coords
+
+    def test_drop_coord_from_dataset(self):
+        a = np.ones((3,))
+        ds = xr.Dataset(
+            data_vars={
+                "a": xr.DataArray(a, coords=dict(adim=["A", "B", "C"]), dims=("adim",)),
+                "b": xr.DataArray(a, coords=dict(bdim=[0, None, 2]), dims=("bdim",)),
+            }
+        )
+        new, dropped_vars, dropped_coords = drop_objects_from_dataset(ds)
+        assert isinstance(new, xr.Dataset)
+        assert new is not ds
+        assert "a" in new
+        assert "b" in new
+        assert not dropped_vars
+        assert dropped_coords == ["bdim"]
+
+    @pytest.mark.parametrize("vname", ["warning", "other"])
+    def test_drop_objects_from_inferencedata(self, vname, caplog):
+        idata = from_dict(
+            sample_stats={
+                "a": np.ones((2, 5, 4)),
+                vname: np.ones((2, 5, 3), dtype=object),
+            },
+            attrs=dict(version="0.1.2"),
+            coords={
+                "adim": [0, 1, None, 3],
+                "vdim": list("ABC"),
+            },
+            dims={"a": ["adim"], vname: ["vdim"]},
+        )
+
+        # Capture logging messages about the droppings
+        with caplog.at_level(logging.DEBUG, logger="arviz"):
+            new = drop_objects_from_inferencedata(idata)
+
+        assert new is not idata
+        assert new.attrs.get("version") == "0.1.2"
+
+        ss = new.get("sample_stats", None)
+        assert isinstance(ss, xr.Dataset)
+        assert "a" in ss
+        assert vname not in ss
+        assert caplog.records
+
+        # Check the logging about the dropped variable
+        rec = caplog.records[0]
+        if vname == "warning":
+            # DEBUG level for the 'warning' stat
+            # which in PyMC is an object an very commonly present.
+            assert rec.levelno == logging.DEBUG
+            assert "Dropped 'warning' variable from 'sample_stats'" in rec.message
+        else:
+            # WARNING level otherwise
+            assert rec.levelno == logging.WARNING
+            assert f"from 'sample_stats' group: ['{vname}']" in rec.message
+
+        # And the logging about dropped coord
+        rec = caplog.records[1]
+        assert rec.levelno == logging.WARNING
+        assert "object coords from 'sample_stats' group: ['adim']" in rec.message