Implement index update for XTensorVariables

Author: ricardoV94

pytensor/xtensor/indexing.py

@@ -4,6 +4,7 @@
 # https://numpy.org/neps/nep-0021-advanced-indexing.html
 # https://docs.xarray.dev/en/latest/user-guide/indexing.html
 # https://tutorial.xarray.dev/intermediate/indexing/advanced-indexing.html
+from typing import Literal
 
 from pytensor.graph.basic import Apply, Constant, Variable
 from pytensor.scalar.basic import discrete_dtypes
@@ -184,3 +185,35 @@ def combine_dim_info(idx_dim, idx_dim_shape):
 
 
 index = Index()
+
+
+class IndexUpdate(XOp):
+    __props__ = ("mode",)
+
+    def __init__(self, mode: Literal["set", "inc"]):
+        if mode not in ("set", "inc"):
+            raise ValueError("mode must be 'set' or 'inc'")
+        self.mode = mode
+
+    def make_node(self, x, y, *idxs):
+        # Call Index on (x, *idxs) to process inputs and infer output type
+        x_view_node = index.make_node(x, *idxs)
+        x, *idxs = x_view_node.inputs
+        [x_view] = x_view_node.outputs
+
+        try:
+            y = as_xtensor(y)
+        except TypeError:
+            y = as_xtensor(as_tensor(y), dims=x_view.type.dims)
+
+        if not set(y.type.dims).issubset(x_view.type.dims):
+            raise ValueError(
+                f"Value dimensions {y.type.dims} must be a subset of the indexed dimensions {x_view.type.dims}"
+            )
+
+        out = x.type()
+        return Apply(self, [x, y, *idxs], [out])
+
+
+index_assignment = IndexUpdate("set")
+index_increment = IndexUpdate("inc")

pytensor/xtensor/rewriting/indexing.py

@@ -3,10 +3,10 @@
 from pytensor import as_symbolic
 from pytensor.graph import Constant, node_rewriter
 from pytensor.tensor import TensorType, arange, specify_shape
-from pytensor.tensor.subtensor import _non_consecutive_adv_indexing
+from pytensor.tensor.subtensor import _non_consecutive_adv_indexing, inc_subtensor
 from pytensor.tensor.type_other import NoneTypeT, SliceType
 from pytensor.xtensor.basic import tensor_from_xtensor, xtensor_from_tensor
-from pytensor.xtensor.indexing import Index
+from pytensor.xtensor.indexing import Index, IndexUpdate, index
 from pytensor.xtensor.rewriting.utils import register_lower_xtensor
 from pytensor.xtensor.type import XTensorType
 
@@ -35,9 +35,7 @@ def to_basic_idx(idx):
     raise TypeError("Cannot convert idx to basic idx")
 
 
-@register_lower_xtensor
-@node_rewriter(tracks=[Index])
-def lower_index(fgraph, node):
+def _lower_index(node):
     """Lower XTensorVariable indexing to regular TensorVariable indexing.
 
     xarray-like indexing has two modes:
@@ -59,12 +57,18 @@ def lower_index(fgraph, node):
     We do this by creating an `arange` tensor that matches the shape of the dimension being indexed,
     and then indexing it with the original slice. This index is then handled as a regular advanced index.
 
-    Note: The IndexOp has only 2 types of indices: Slices and XTensorVariables. Regular array indices
-    are converted to the appropriate XTensorVariable by `Index.make_node`
+    Finally, the location of views resulting from advanced indices follows two distinct behaviors in numpy.
+    When all advanced indices are consecutive, the respective view is located in the "original" location.
+    However, if advanced indices are separated by basic indices (slices in our case), the output views
+    always show up at the front of the array. This information is returned as the second output of this function,
+    which labels the final position of the indexed dimensions under this rule.
     """
 
+    assert isinstance(node.op, Index)
+
     x, *idxs = node.inputs
     [out] = node.outputs
+    x_tensor_indexed_dims = out.type.dims
     x_tensor = tensor_from_xtensor(x)
 
     if all(
@@ -141,10 +145,68 @@ def lower_index(fgraph, node):
             x_tensor_indexed_dims = [
                 dim for dim in out_dims if dim not in x_tensor_indexed_basic_dims
             ] + x_tensor_indexed_basic_dims
-            transpose_order = [x_tensor_indexed_dims.index(dim) for dim in out_dims]
-            x_tensor_indexed = x_tensor_indexed.transpose(transpose_order)
+
+    return x_tensor_indexed, x_tensor_indexed_dims
+
+
+@register_lower_xtensor
+@node_rewriter(tracks=[Index])
+def lower_index(fgraph, node):
+    """Lower XTensorVariable indexing to regular TensorVariable indexing.
+
+    The bulk of the work is done by `_lower_index`, except for special logic to control the
+    location of non-consecutive advanced indices, and to preserve static shape information.
+    """
+
+    [out] = node.outputs
+    out_dims = out.type.dims
+
+    x_tensor_indexed, x_tensor_indexed_dims = _lower_index(node)
+    if x_tensor_indexed_dims != out_dims:
+        # Numpy moves advanced indexing dimensions to the front when they are not consecutive
+        # We need to transpose them back to the expected output order
+        transpose_order = [x_tensor_indexed_dims.index(dim) for dim in out_dims]
+        x_tensor_indexed = x_tensor_indexed.transpose(transpose_order)
 
     # Add lost shape information
     x_tensor_indexed = specify_shape(x_tensor_indexed, out.type.shape)
-    new_out = xtensor_from_tensor(x_tensor_indexed, dims=out.type.dims)
+
+    new_out = xtensor_from_tensor(x_tensor_indexed, dims=out.dims)
+    return [new_out]
+
+
+@register_lower_xtensor
+@node_rewriter(tracks=[IndexUpdate])
+def lower_index_update(fgraph, node):
+    """Lower XTensorVariable index update to regular TensorVariable indexing update.
+
+    This rewrite requires converting the index view to a tensor-based equivalent expression,
+    just like `lower_index`. It then requires aligning the dimensions of y with the
+    dimensions of the index view, with special care for non-consecutive dimensions being
+    pulled to the front axis according to numpy rules.
+    """
+    x, y, *idxs = node.inputs
+
+    # Lower the indexing part first
+    indexed_node = index.make_node(x, *idxs)
+    x_tensor_indexed, x_tensor_indexed_dims = _lower_index(indexed_node)
+    y_tensor = tensor_from_xtensor(y)
+
+    # Align dimensions of y with those of the indexed tensor x
+    y_dims = y.type.dims
+    y_dims_set = set(y_dims)
+    y_order = tuple(
+        y_dims.index(x_dim) if x_dim in y_dims_set else "x"
+        for x_dim in x_tensor_indexed_dims
+    )
+    # Remove useless left expand_dims
+    while len(y_order) > 0 and y_order[0] == "x":
+        y_order = y_order[1:]
+    if y_order != tuple(range(y_tensor.type.ndim)):
+        y_tensor = y_tensor.dimshuffle(y_order)
+
+    x_tensor_updated = inc_subtensor(
+        x_tensor_indexed, y_tensor, set_instead_of_inc=node.op.mode == "set"
+    )
+    new_out = xtensor_from_tensor(x_tensor_updated, dims=x.type.dims)
     return [new_out]

pytensor/xtensor/type.py

@@ -411,8 +411,10 @@ def item(self):
 
     # Indexing
     # https://docs.xarray.dev/en/latest/api.html#id2
-    def __setitem__(self, key, value):
-        raise TypeError("XTensorVariable does not support item assignment.")
+    def __setitem__(self, idx, value):
+        raise TypeError(
+            "XTensorVariable does not support item assignment. Use the output of `x[idx].set` or `x[idx].inc` instead."
+        )
 
     @property
     def loc(self):
@@ -476,6 +478,28 @@ def isel(
 
         return px.indexing.index(self, *indices)
 
+    def set(self, value):
+        if not (
+            self.owner is not None and isinstance(self.owner.op, px.indexing.Index)
+        ):
+            raise ValueError(
+                f"set can only be called on the output of an index (or isel) operation. Self is the result of {self.owner}"
+            )
+
+        x, *idxs = self.owner.inputs
+        return px.indexing.index_assignment(x, value, *idxs)
+
+    def inc(self, value):
+        if not (
+            self.owner is not None and isinstance(self.owner.op, px.indexing.Index)
+        ):
+            raise ValueError(
+                f"inc can only be called on the output of an index (or isel) operation. Self is the result of {self.owner}"
+            )
+
+        x, *idxs = self.owner.inputs
+        return px.indexing.index_increment(x, value, *idxs)
+
     def _head_tail_or_thin(
         self,
         indexers: dict[str, Any] | int | None,

tests/xtensor/test_indexing.py

@@ -11,7 +11,12 @@
 
 from pytensor.tensor import tensor
 from pytensor.xtensor import xtensor
-from tests.xtensor.util import xr_arange_like, xr_assert_allclose, xr_function
+from tests.xtensor.util import (
+    xr_arange_like,
+    xr_assert_allclose,
+    xr_function,
+    xr_random_like,
+)
 
 
 @pytest.mark.parametrize(
@@ -351,3 +356,138 @@ def test_boolean_indexing():
     expected_res2 = x_test[bool_idx_test, int_idx_test.rename(a="b")]
     xr_assert_allclose(res1, expected_res1)
     xr_assert_allclose(res2, expected_res2)
+
+
+@pytest.mark.parametrize("mode", ("set", "inc"))
+def test_basic_index_update(mode):
+    x = xtensor("x", shape=(11, 7), dims=("a", "b"))
+    y = xtensor("y", shape=(7, 5), dims=("a", "b"))
+    x_indexed = x[2:-2, 2:]
+    update_method = getattr(x_indexed, mode)
+
+    x_updated = [
+        update_method(y),
+        update_method(y.T),
+        update_method(y.isel(a=-1)),
+        update_method(y.isel(b=-1)),
+        update_method(y.isel(a=-2, b=-2)),
+    ]
+
+    fn = xr_function([x, y], x_updated)
+    x_test = xr_random_like(x)
+    y_test = xr_random_like(y)
+    results = fn(x_test, y_test)
+
+    def update_fn(y):
+        x = x_test.copy()
+        if mode == "set":
+            x[2:-2, 2:] = y
+        elif mode == "inc":
+            x[2:-2, 2:] += y
+        return x
+
+    expected_results = [
+        update_fn(y_test),
+        update_fn(y_test.T),
+        update_fn(y_test.isel(a=-1)),
+        update_fn(y_test.isel(b=-1)),
+        update_fn(y_test.isel(a=-2, b=-2)),
+    ]
+    for result, expected_result in zip(results, expected_results):
+        xr_assert_allclose(result, expected_result)
+
+
+@pytest.mark.parametrize("mode", ("set", "inc"))
+@pytest.mark.parametrize("idx_dtype", (int, bool))
+def test_adv_index_update(mode, idx_dtype):
+    x = xtensor("x", shape=(5, 5), dims=("a", "b"))
+    y = xtensor("y", shape=(3,), dims=("b",))
+    idx = xtensor("idx", dtype=idx_dtype, shape=(None,), dims=("a",))
+
+    orthogonal_update1 = getattr(x[idx, -3:], mode)(y)
+    orthogonal_update2 = getattr(x[idx, -3:], mode)(y.rename(b="a"))
+    if idx_dtype is not bool:
+        # Vectorized booling indexing/update is not allowed
+        vectorized_update = getattr(x[idx.rename(a="b"), :3], mode)(y)
+    else:
+        with pytest.raises(
+            IndexError,
+            match="Boolean indexer should be unlabeled or on the same dimension to the indexed array.",
+        ):
+            getattr(x[idx.rename(a="b"), :3], mode)(y)
+        vectorized_update = x
+
+    outs = [orthogonal_update1, orthogonal_update2, vectorized_update]
+
+    fn = xr_function([x, idx, y], outs)
+    x_test = xr_random_like(x)
+    y_test = xr_random_like(y)
+    if idx_dtype is int:
+        idx_test = DataArray([0, 1, 2], dims=("a",))
+    else:
+        idx_test = DataArray([True, False, True, True, False], dims=("a",))
+    results = fn(x_test, idx_test, y_test)
+
+    def update_fn(x, idx, y):
+        x = x.copy()
+        if mode == "set":
+            x[idx] = y
+        else:
+            x[idx] += y
+        return x
+
+    expected_results = [
+        update_fn(x_test, (idx_test, slice(-3, None)), y_test),
+        update_fn(
+            x_test,
+            (idx_test, slice(-3, None)),
+            y_test.rename(b="a"),
+        ),
+        update_fn(x_test, (idx_test.rename(a="b"), slice(None, 3)), y_test)
+        if idx_dtype is not bool
+        else x_test,
+    ]
+    for result, expected_result in zip(results, expected_results):
+        xr_assert_allclose(result, expected_result)
+
+
+@pytest.mark.parametrize("mode", ("set", "inc"))
+def test_non_consecutive_idx_update(mode):
+    x = xtensor("x", shape=(2, 3, 5, 7), dims=("a", "b", "c", "d"))
+    y = xtensor("y", shape=(5, 4), dims=("c", "b"))
+    x_indexed = x[:, [0, 1, 2, 2], :, ("b", [0, 1, 1, 2])]
+    out = getattr(x_indexed, mode)(y)
+
+    fn = xr_function([x, y], out)
+    x_test = xr_random_like(x)
+    y_test = xr_random_like(y)
+
+    result = fn(x_test, y_test)
+    expected_result = x_test.copy()
+    # xarray fails inplace operation with the "tuple trick"
+    # https://github.com/pydata/xarray/issues/10387
+    d_indexer = DataArray([0, 1, 1, 2], dims=("b",))
+    if mode == "set":
+        expected_result[:, [0, 1, 2, 2], :, d_indexer] = y_test
+    else:
+        expected_result[:, [0, 1, 2, 2], :, d_indexer] += y_test
+    xr_assert_allclose(result, expected_result)
+
+
+def test_indexing_renames_into_update_variable():
+    x = xtensor("x", shape=(5, 5), dims=("a", "b"))
+    y = xtensor("y", shape=(3,), dims=("d",))
+    idx = xtensor("idx", dtype=int, shape=(None,), dims=("d",))
+
+    # define "d" dimension by slicing the "a" dimension so we can set y into x
+    orthogonal_update1 = x[idx].set(y)
+    fn = xr_function([x, idx, y], orthogonal_update1)
+
+    x_test = np.abs(xr_random_like(x))
+    y_test = -np.abs(xr_random_like(y))
+    idx_test = DataArray([0, 2, 3], dims=("d",))
+
+    result = fn(x_test, idx_test, y_test)
+    expected_result = x_test.copy()
+    expected_result[idx_test] = y_test
+    xr_assert_allclose(result, expected_result)