Add to_numpy() and as_numpy() methods (#5568)

* added to_numpy() and as_numpy() methods * remove special-casing of cupy arrays in .values in favour of using .to_numpy() * lint * Fix mypy (I think?) * added Dataset.as_numpy() * improved docstrings * add what's new * add to API docs * linting * fix failures by only importing pint when needed * refactor pycompat into class * compute instead of load * added tests * fixed sparse test * tests and fixes for ds.as_numpy() * fix sparse tests * fix linting * tests for Variable * test IndexVariable too * use numpy.asarray to avoid a copy * also convert coords * Force tests again after #5600 * Apply suggestions from code review * Update xarray/core/variable.py * fix import * formatting * remove type check Co-authored-by: Stephan Hoyer <[email protected]> * remove attempt to call to_numpy Co-authored-by: Maximilian Roos <[email protected]> Co-authored-by: Deepak Cherian <[email protected]> Co-authored-by: Stephan Hoyer <[email protected]>
pydata · Jul 21, 2021 · c5ee050 · c5ee050
1 parent 86ca67e
commit c5ee050
Show file tree

Hide file tree

Showing 10 changed files with 363 additions and 36 deletions.
diff --git a/doc/api.rst b/doc/api.rst
@@ -686,6 +686,7 @@ Dataset methods
    open_zarr
    Dataset.to_netcdf
    Dataset.to_pandas
+   Dataset.as_numpy
    Dataset.to_zarr
    save_mfdataset
    Dataset.to_array
@@ -716,6 +717,8 @@ DataArray methods
    DataArray.to_pandas
    DataArray.to_series
    DataArray.to_dataframe
+   DataArray.to_numpy
+   DataArray.as_numpy
    DataArray.to_index
    DataArray.to_masked_array
    DataArray.to_cdms2

diff --git a/doc/whats-new.rst b/doc/whats-new.rst
@@ -56,6 +56,8 @@ New Features
 - Allow removal of the coordinate attribute ``coordinates`` on variables by setting ``.attrs['coordinates']= None``
   (:issue:`5510`).
   By `Elle Smith <https://github.com/ellesmith88>`_.
+- Added :py:meth:`DataArray.to_numpy`, :py:meth:`DataArray.as_numpy`, and :py:meth:`Dataset.as_numpy`. (:pull:`5568`).
+  By `Tom Nicholas <https://github.com/TomNicholas>`_.
 
 Breaking changes
 ~~~~~~~~~~~~~~~~

diff --git a/xarray/core/dataarray.py b/xarray/core/dataarray.py
@@ -426,12 +426,12 @@ def __init__(
         self._close = None
 
     def _replace(
-        self,
+        self: T_DataArray,
         variable: Variable = None,
         coords=None,
         name: Union[Hashable, None, Default] = _default,
         indexes=None,
-    ) -> "DataArray":
+    ) -> T_DataArray:
         if variable is None:
             variable = self.variable
         if coords is None:
@@ -623,7 +623,16 @@ def __len__(self) -> int:
 
     @property
     def data(self) -> Any:
-        """The array's data as a dask or numpy array"""
+        """
+        The DataArray's data as an array. The underlying array type
+        (e.g. dask, sparse, pint) is preserved.
+
+        See Also
+        --------
+        DataArray.to_numpy
+        DataArray.as_numpy
+        DataArray.values
+        """
         return self.variable.data
 
     @data.setter
@@ -632,13 +641,46 @@ def data(self, value: Any) -> None:
 
     @property
     def values(self) -> np.ndarray:
-        """The array's data as a numpy.ndarray"""
+        """
+        The array's data as a numpy.ndarray.
+
+        If the array's data is not a numpy.ndarray this will attempt to convert
+        it naively using np.array(), which will raise an error if the array
+        type does not support coercion like this (e.g. cupy).
+        """
         return self.variable.values
 
     @values.setter
     def values(self, value: Any) -> None:
         self.variable.values = value
 
+    def to_numpy(self) -> np.ndarray:
+        """
+        Coerces wrapped data to numpy and returns a numpy.ndarray.
+
+        See also
+        --------
+        DataArray.as_numpy : Same but returns the surrounding DataArray instead.
+        Dataset.as_numpy
+        DataArray.values
+        DataArray.data
+        """
+        return self.variable.to_numpy()
+
+    def as_numpy(self: T_DataArray) -> T_DataArray:
+        """
+        Coerces wrapped data and coordinates into numpy arrays, returning a DataArray.
+
+        See also
+        --------
+        DataArray.to_numpy : Same but returns only the data as a numpy.ndarray object.
+        Dataset.as_numpy : Converts all variables in a Dataset.
+        DataArray.values
+        DataArray.data
+        """
+        coords = {k: v.as_numpy() for k, v in self._coords.items()}
+        return self._replace(self.variable.as_numpy(), coords, indexes=self._indexes)
+
     @property
     def _in_memory(self) -> bool:
         return self.variable._in_memory
@@ -931,7 +973,7 @@ def persist(self, **kwargs) -> "DataArray":
         ds = self._to_temp_dataset().persist(**kwargs)
         return self._from_temp_dataset(ds)
 
-    def copy(self, deep: bool = True, data: Any = None) -> "DataArray":
+    def copy(self: T_DataArray, deep: bool = True, data: Any = None) -> T_DataArray:
         """Returns a copy of this array.
 
         If `deep=True`, a deep copy is made of the data array.

diff --git a/xarray/core/dataset.py b/xarray/core/dataset.py
@@ -1323,6 +1323,18 @@ def copy(self, deep: bool = False, data: Mapping = None) -> "Dataset":
 
         return self._replace(variables, attrs=attrs)
 
+    def as_numpy(self: "Dataset") -> "Dataset":
+        """
+        Coerces wrapped data and coordinates into numpy arrays, returning a Dataset.
+
+        See also
+        --------
+        DataArray.as_numpy
+        DataArray.to_numpy : Returns only the data as a numpy.ndarray object.
+        """
+        numpy_variables = {k: v.as_numpy() for k, v in self.variables.items()}
+        return self._replace(variables=numpy_variables)
+
     @property
     def _level_coords(self) -> Dict[str, Hashable]:
         """Return a mapping of all MultiIndex levels and their corresponding

diff --git a/xarray/core/pycompat.py b/xarray/core/pycompat.py
@@ -1,47 +1,63 @@
 from distutils.version import LooseVersion
+from importlib import import_module
 
 import numpy as np
 
 from .utils import is_duck_array
 
 integer_types = (int, np.integer)
 
-try:
-    import dask
-    import dask.array
-    from dask.base import is_dask_collection
 
-    dask_version = LooseVersion(dask.__version__)
+class DuckArrayModule:
+    """
+    Solely for internal isinstance and version checks.
 
-    # solely for isinstance checks
-    dask_array_type = (dask.array.Array,)
+    Motivated by having to only import pint when required (as pint currently imports xarray)
+    https://github.com/pydata/xarray/pull/5561#discussion_r664815718
+    """
 
-    def is_duck_dask_array(x):
-        return is_duck_array(x) and is_dask_collection(x)
+    def __init__(self, mod):
+        try:
+            duck_array_module = import_module(mod)
+            duck_array_version = LooseVersion(duck_array_module.__version__)
+
+            if mod == "dask":
+                duck_array_type = (import_module("dask.array").Array,)
+            elif mod == "pint":
+                duck_array_type = (duck_array_module.Quantity,)
+            elif mod == "cupy":
+                duck_array_type = (duck_array_module.ndarray,)
+            elif mod == "sparse":
+                duck_array_type = (duck_array_module.SparseArray,)
+            else:
+                raise NotImplementedError
+
+        except ImportError:  # pragma: no cover
+            duck_array_module = None
+            duck_array_version = LooseVersion("0.0.0")
+            duck_array_type = ()
 
+        self.module = duck_array_module
+        self.version = duck_array_version
+        self.type = duck_array_type
+        self.available = duck_array_module is not None
 
-except ImportError:  # pragma: no cover
-    dask_version = LooseVersion("0.0.0")
-    dask_array_type = ()
-    is_duck_dask_array = lambda _: False
-    is_dask_collection = lambda _: False
 
-try:
-    # solely for isinstance checks
-    import sparse
+def is_duck_dask_array(x):
+    if DuckArrayModule("dask").available:
+        from dask.base import is_dask_collection
+
+        return is_duck_array(x) and is_dask_collection(x)
+    else:
+        return False
+
 
-    sparse_version = LooseVersion(sparse.__version__)
-    sparse_array_type = (sparse.SparseArray,)
-except ImportError:  # pragma: no cover
-    sparse_version = LooseVersion("0.0.0")
-    sparse_array_type = ()
+dsk = DuckArrayModule("dask")
+dask_version = dsk.version
+dask_array_type = dsk.type
 
-try:
-    # solely for isinstance checks
-    import cupy
+sp = DuckArrayModule("sparse")
+sparse_array_type = sp.type
+sparse_version = sp.version
 
-    cupy_version = LooseVersion(cupy.__version__)
-    cupy_array_type = (cupy.ndarray,)
-except ImportError:  # pragma: no cover
-    cupy_version = LooseVersion("0.0.0")
-    cupy_array_type = ()
+cupy_array_type = DuckArrayModule("cupy").type
diff --git a/xarray/core/variable.py b/xarray/core/variable.py
@@ -29,10 +29,12 @@
 from .indexing import BasicIndexer, OuterIndexer, VectorizedIndexer, as_indexable
 from .options import _get_keep_attrs
 from .pycompat import (
+    DuckArrayModule,
     cupy_array_type,
     dask_array_type,
     integer_types,
     is_duck_dask_array,
+    sparse_array_type,
 )
 from .utils import (
     NdimSizeLenMixin,
@@ -259,7 +261,7 @@ def _as_array_or_item(data):
 
     TODO: remove this (replace with np.asarray) once these issues are fixed
     """
-    data = data.get() if isinstance(data, cupy_array_type) else np.asarray(data)
+    data = np.asarray(data)
     if data.ndim == 0:
         if data.dtype.kind == "M":
             data = np.datetime64(data, "ns")
@@ -1069,6 +1071,29 @@ def chunk(self, chunks={}, name=None, lock=False):
 
         return self._replace(data=data)
 
+    def to_numpy(self) -> np.ndarray:
+        """Coerces wrapped data to numpy and returns a numpy.ndarray"""
+        # TODO an entrypoint so array libraries can choose coercion method?
+        data = self.data
+        # TODO first attempt to call .to_numpy() once some libraries implement it
+        if isinstance(data, dask_array_type):
+            data = data.compute()
+        if isinstance(data, cupy_array_type):
+            data = data.get()
+        # pint has to be imported dynamically as pint imports xarray
+        pint_array_type = DuckArrayModule("pint").type
+        if isinstance(data, pint_array_type):
+            data = data.magnitude
+        if isinstance(data, sparse_array_type):
+            data = data.todense()
+        data = np.asarray(data)
+
+        return data
+
+    def as_numpy(self: VariableType) -> VariableType:
+        """Coerces wrapped data into a numpy array, returning a Variable."""
+        return self._replace(data=self.to_numpy())
+
     def _as_sparse(self, sparse_format=_default, fill_value=dtypes.NA):
         """
         use sparse-array as backend.

diff --git a/xarray/tests/__init__.py b/xarray/tests/__init__.py
@@ -83,6 +83,7 @@ def LooseVersion(vstring):
 has_numbagg, requires_numbagg = _importorskip("numbagg")
 has_seaborn, requires_seaborn = _importorskip("seaborn")
 has_sparse, requires_sparse = _importorskip("sparse")
+has_cupy, requires_cupy = _importorskip("cupy")
 has_cartopy, requires_cartopy = _importorskip("cartopy")
 # Need Pint 0.15 for __dask_tokenize__ tests for Quantity wrapped Dask Arrays
 has_pint_0_15, requires_pint_0_15 = _importorskip("pint", minversion="0.15")

diff --git a/xarray/tests/test_dataarray.py b/xarray/tests/test_dataarray.py
@@ -36,10 +36,12 @@
     has_dask,
     raise_if_dask_computes,
     requires_bottleneck,
+    requires_cupy,
     requires_dask,
     requires_iris,
     requires_numbagg,
     requires_numexpr,
+    requires_pint_0_15,
     requires_scipy,
     requires_sparse,
     source_ndarray,
@@ -7375,3 +7377,87 @@ def test_drop_duplicates(keep):
     expected = xr.DataArray(data, dims="time", coords={"time": time}, name="test")
     result = ds.drop_duplicates("time", keep=keep)
     assert_equal(expected, result)
+
+
+class TestNumpyCoercion:
+    # TODO once flexible indexes refactor complete also test coercion of dimension coords
+    def test_from_numpy(self):
+        da = xr.DataArray([1, 2, 3], dims="x", coords={"lat": ("x", [4, 5, 6])})
+
+        assert_identical(da.as_numpy(), da)
+        np.testing.assert_equal(da.to_numpy(), np.array([1, 2, 3]))
+        np.testing.assert_equal(da["lat"].to_numpy(), np.array([4, 5, 6]))
+
+    @requires_dask
+    def test_from_dask(self):
+        da = xr.DataArray([1, 2, 3], dims="x", coords={"lat": ("x", [4, 5, 6])})
+        da_chunked = da.chunk(1)
+
+        assert_identical(da_chunked.as_numpy(), da.compute())
+        np.testing.assert_equal(da.to_numpy(), np.array([1, 2, 3]))
+        np.testing.assert_equal(da["lat"].to_numpy(), np.array([4, 5, 6]))
+
+    @requires_pint_0_15
+    def test_from_pint(self):
+        from pint import Quantity
+
+        arr = np.array([1, 2, 3])
+        da = xr.DataArray(
+            Quantity(arr, units="Pa"),
+            dims="x",
+            coords={"lat": ("x", Quantity(arr + 3, units="m"))},
+        )
+
+        expected = xr.DataArray(arr, dims="x", coords={"lat": ("x", arr + 3)})
+        assert_identical(da.as_numpy(), expected)
+        np.testing.assert_equal(da.to_numpy(), arr)
+        np.testing.assert_equal(da["lat"].to_numpy(), arr + 3)
+
+    @requires_sparse
+    def test_from_sparse(self):
+        import sparse
+
+        arr = np.diagflat([1, 2, 3])
+        sparr = sparse.COO.from_numpy(arr)
+        da = xr.DataArray(
+            sparr, dims=["x", "y"], coords={"elev": (("x", "y"), sparr + 3)}
+        )
+
+        expected = xr.DataArray(
+            arr, dims=["x", "y"], coords={"elev": (("x", "y"), arr + 3)}
+        )
+        assert_identical(da.as_numpy(), expected)
+        np.testing.assert_equal(da.to_numpy(), arr)
+
+    @requires_cupy
+    def test_from_cupy(self):
+        import cupy as cp
+
+        arr = np.array([1, 2, 3])
+        da = xr.DataArray(
+            cp.array(arr), dims="x", coords={"lat": ("x", cp.array(arr + 3))}
+        )
+
+        expected = xr.DataArray(arr, dims="x", coords={"lat": ("x", arr + 3)})
+        assert_identical(da.as_numpy(), expected)
+        np.testing.assert_equal(da.to_numpy(), arr)
+
+    @requires_dask
+    @requires_pint_0_15
+    def test_from_pint_wrapping_dask(self):
+        import dask
+        from pint import Quantity
+
+        arr = np.array([1, 2, 3])
+        d = dask.array.from_array(arr)
+        da = xr.DataArray(
+            Quantity(d, units="Pa"),
+            dims="x",
+            coords={"lat": ("x", Quantity(d, units="m") * 2)},
+        )
+
+        result = da.as_numpy()
+        result.name = None  # remove dask-assigned name
+        expected = xr.DataArray(arr, dims="x", coords={"lat": ("x", arr * 2)})
+        assert_identical(result, expected)
+        np.testing.assert_equal(da.to_numpy(), arr)