Fix numpy array conversion for LindiH5pyDataset

lindi/LindiH5pyFile/LindiH5pyDataset.py

@@ -168,6 +168,21 @@ def compression_opts(self):
     def chunks(self):
         return self._zarr_array.chunks
 
+    @property
+    def _is_empty(self):
+        # LINDI datasets backed by zarr are never HDF5 NULL-space datasets
+        return False
+
+    def __array__(self, dtype=None, copy=None):
+        if copy is False:
+            raise ValueError(
+                "LindiH5pyDataset.__array__ received copy=False "
+                "but memory allocation cannot be avoided on read"
+            )
+        if self._is_scalar:
+            return np.array(self[()], dtype=dtype)
+        return np.array(self[:], dtype=dtype)
+
     def __repr__(self):  # type: ignore
         return f"<{self.__class__.__name__}: {self.name}>"
 

tests/test_numpy_array_conversion.py

@@ -0,0 +1,76 @@
+import tempfile
+import numpy as np
+import h5py
+import lindi
+
+
+def test_numpy_array_conversion():
+    """Test that LindiH5pyDataset supports np.asarray() and np.atleast_2d().
+
+    Regression test for https://github.com/NeurodataWithoutBorders/lindi/issues/120
+    """
+    with tempfile.TemporaryDirectory() as tmpdir:
+        h5_fname = f'{tmpdir}/test.h5'
+        lindi_json_fname = f'{tmpdir}/test.lindi.json'
+
+        # Create a simple HDF5 file with a 1D dataset
+        with h5py.File(h5_fname, 'w') as f:
+            f.create_dataset('data', data=np.arange(100, dtype=np.float64))
+            f.create_dataset('scalar', data=42.0)
+
+        # Convert to lindi
+        with lindi.LindiH5pyFile.from_hdf5_file(h5_fname, url=h5_fname) as f:
+            f.write_lindi_file(lindi_json_fname)
+
+        # Open the lindi file and test numpy conversions
+        with lindi.LindiH5pyFile.from_lindi_file(lindi_json_fname) as f:
+            ds = f['data']
+
+            # Test _is_empty
+            assert ds._is_empty is False
+
+            # Test np.asarray - this triggers __array__
+            arr = np.asarray(ds)
+            assert arr.shape == (100,)
+            np.testing.assert_array_equal(arr, np.arange(100, dtype=np.float64))
+
+            # Test np.atleast_2d - this is what failed in the issue
+            arr2d = np.atleast_2d(ds)
+            assert arr2d.shape == (1, 100)
+
+            # Test np.array with dtype conversion
+            arr_int = np.array(ds, dtype=np.int32)
+            assert arr_int.dtype == np.int32
+
+            # Test scalar dataset
+            sc = f['scalar']
+            arr_sc = np.asarray(sc)
+            assert arr_sc.shape == ()
+            assert float(arr_sc) == 42.0
+
+
+def test_numpy_array_conversion_compound():
+    """Test that __array__ works for compound dtype datasets."""
+    with tempfile.TemporaryDirectory() as tmpdir:
+        h5_fname = f'{tmpdir}/test.h5'
+        lindi_json_fname = f'{tmpdir}/test.lindi.json'
+
+        compound_dtype = np.dtype([('x', np.int32), ('y', np.float64)])
+        data = np.array([(1, 2.5), (3, 4.5), (5, 6.5)], dtype=compound_dtype)
+
+        with h5py.File(h5_fname, 'w') as f:
+            f.create_dataset('compound', data=data)
+
+        with lindi.LindiH5pyFile.from_hdf5_file(h5_fname, url=h5_fname) as f:
+            f.write_lindi_file(lindi_json_fname)
+
+        with lindi.LindiH5pyFile.from_lindi_file(lindi_json_fname) as f:
+            ds = f['compound']
+            assert ds._is_empty is False
+
+            # Test field access followed by np.asarray
+            x_vals = np.asarray(ds['x'][:])
+            np.testing.assert_array_equal(x_vals, np.array([1, 3, 5], dtype=np.int32))
+
+            y_vals = np.asarray(ds['y'][:])
+            np.testing.assert_array_equal(y_vals, np.array([2.5, 4.5, 6.5], dtype=np.float64))