Adding expand_dims for xtensor

Author: AllenDowney

pytensor/xtensor/rewriting/shape.py

@@ -1,6 +1,11 @@
+import numpy as np
+
 from pytensor.graph import node_rewriter
+from pytensor.raise_op import Assert
 from pytensor.tensor import (
     broadcast_to,
+    expand_dims,
+    gt,
     join,
     moveaxis,
     specify_shape,
@@ -10,6 +15,7 @@
 from pytensor.xtensor.rewriting.basic import register_lower_xtensor
 from pytensor.xtensor.shape import (
     Concat,
+    ExpandDims,
     Squeeze,
     Stack,
     Transpose,
@@ -132,3 +138,29 @@ def local_squeeze_reshape(fgraph, node):
 
     new_out = xtensor_from_tensor(x_tensor_squeezed, dims=node.outputs[0].type.dims)
     return [new_out]
+
+
+@register_lower_xtensor
+@node_rewriter([ExpandDims])
+def local_expand_dims_reshape(fgraph, node):
+    """Rewrite ExpandDims to tensor.expand_dims and optionally broadcast_to or specify_shape."""
+    [x] = node.inputs
+    x_tensor = tensor_from_xtensor(x)
+    x_tensor_expanded = expand_dims(x_tensor, axis=0)
+
+    target_shape = node.outputs[0].type.shape
+
+    size = getattr(node.op, "size", 1)
+    if isinstance(size, int | np.integer):
+        if size != 1 and None not in target_shape:
+            x_tensor_expanded = broadcast_to(x_tensor_expanded, target_shape)
+    else:
+        # Symbolic size: enforce shape so broadcast happens downstream correctly
+        # Also validate that size is positive
+        x_tensor_expanded = Assert(msg="size must be positive")(
+            x_tensor_expanded, gt(size, 0)
+        )
+        x_tensor_expanded = specify_shape(x_tensor_expanded, target_shape)
+
+    new_out = xtensor_from_tensor(x_tensor_expanded, dims=node.outputs[0].type.dims)
+    return [new_out]

pytensor/xtensor/shape.py

@@ -3,6 +3,8 @@
 from types import EllipsisType
 from typing import Literal
 
+import numpy as np
+
 from pytensor.graph import Apply
 from pytensor.scalar import discrete_dtypes, upcast
 from pytensor.tensor import as_tensor, get_scalar_constant_value
@@ -380,3 +382,75 @@ def squeeze(x, dim=None):
         return x  # no-op if nothing to squeeze
 
     return Squeeze(dims=dims)(x)
+
+
+class ExpandDims(XOp):
+    """Add a new dimension to an XTensorVariable."""
+
+    __props__ = ("dims", "size")
+
+    def __init__(self, dim, size=1):
+        self.dims = dim
+        self.size = size
+
+    def make_node(self, x):
+        x = as_xtensor(x)
+
+        if self.dims is None:
+            # No-op: return same variable
+            return Apply(self, [x], [x])
+
+        # Insert new dim at front
+        new_dims = (self.dims, *x.type.dims)
+
+        # Determine shape
+        if isinstance(self.size, int | np.integer):
+            new_shape = (self.size, *x.type.shape)
+        else:
+            new_shape = (None, *x.type.shape)  # symbolic size
+
+        out = xtensor(
+            dtype=x.type.dtype,
+            shape=new_shape,
+            dims=new_dims,
+        )
+        return Apply(self, [x], [out])
+
+
+def expand_dims(x, dim: str | None, size=1):
+    """Add a new dimension to an XTensorVariable.
+
+    Parameters
+    ----------
+    x : XTensorVariable
+        Input tensor
+    dim : str or None
+        Name of new dimension. If None, returns x unchanged.
+    size : int or symbolic, optional
+        Size of the new dimension (default 1)
+
+    Returns
+    -------
+    XTensorVariable
+        Tensor with the new dimension inserted
+    """
+    x = as_xtensor(x)
+
+    if dim is None:
+        return x  # No-op
+
+    if not isinstance(dim, str):
+        raise TypeError(f"`dim` must be a string or None, got: {type(dim)}")
+
+    if dim in x.type.dims:
+        raise ValueError(f"Dimension {dim} already exists in {x.type.dims}")
+
+    if isinstance(size, int | np.integer):
+        if size <= 0:
+            raise ValueError(f"size must be positive, got: {size}")
+    elif not (
+        hasattr(size, "ndim") and getattr(size, "ndim", None) == 0  # symbolic scalar
+    ):
+        raise TypeError(f"size must be an int or scalar variable, got: {type(size)}")
+
+    return ExpandDims(dim=dim, size=size)(x)

tests/xtensor/test_shape.py

@@ -8,12 +8,14 @@
 from itertools import chain, combinations
 
 import numpy as np
-import pytest
+import xarray as xr
 from xarray import DataArray
 from xarray import concat as xr_concat
 
+from pytensor.tensor import scalar
 from pytensor.xtensor.shape import (
     concat,
+    expand_dims,
     squeeze,
     stack,
     transpose,
@@ -301,6 +303,15 @@ def test_squeeze_explicit_dims():
     fn3d = xr_function([x3], y3d)
     xr_assert_allclose(fn3d(x3_test), x3_test)
 
+    # Reversibility with expand_dims
+    x6 = xtensor("x6", dims=("a", "b", "c"), shape=(2, 1, 3))
+    y6 = squeeze(x6, "b")
+    # First expand_dims adds at front, then transpose puts it in the right place
+    z6 = transpose(expand_dims(y6, "b"), "a", "b", "c")
+    fn6 = xr_function([x6], z6)
+    x6_test = xr_arange_like(x6)
+    xr_assert_allclose(fn6(x6_test), x6_test)
+
 
 def test_squeeze_implicit_dims():
     """Test squeeze with implicit dim=None (all size-1 dimensions)."""
@@ -369,3 +380,199 @@ def test_squeeze_errors():
     fn2 = xr_function([x2], y2)
     with pytest.raises(Exception):
         fn2(x2_test)
+
+
+def test_expand_dims_explicit():
+    """Test expand_dims with explicitly named dimensions and sizes."""
+
+    # 1D case
+    x = xtensor("x", dims=("city",), shape=(3,))
+    y = expand_dims(x, "country")
+    fn = xr_function([x], y)
+    x_xr = xr_arange_like(x)
+    xr_assert_allclose(fn(x_xr), x_xr.expand_dims("country"))
+
+    # 2D case
+    x = xtensor("x", dims=("city", "year"), shape=(2, 2))
+    y = expand_dims(x, "country")
+    fn = xr_function([x], y)
+    xr_assert_allclose(fn(xr_arange_like(x)), xr_arange_like(x).expand_dims("country"))
+
+    # 3D case
+    x = xtensor("x", dims=("city", "year", "month"), shape=(2, 2, 2))
+    y = expand_dims(x, "country")
+    fn = xr_function([x], y)
+    xr_assert_allclose(fn(xr_arange_like(x)), xr_arange_like(x).expand_dims("country"))
+
+    # Prepending various dims
+    x = xtensor("x", dims=("a", "b"), shape=(2, 3))
+    for new_dim in ("x", "y", "z"):
+        y = expand_dims(x, new_dim)
+        assert y.type.dims == (new_dim, "a", "b")
+        assert y.type.shape == (1, 2, 3)
+
+    # Explicit size=1 behaves like default
+    y1 = expand_dims(x, "batch", size=1)
+    y2 = expand_dims(x, "batch")
+    fn1 = xr_function([x], y1)
+    fn2 = xr_function([x], y2)
+    x_test = xr_arange_like(x)
+    xr_assert_allclose(fn1(x_test), fn2(x_test))
+
+    # Scalar expansion
+    x = xtensor("x", dims=(), shape=())
+    y = expand_dims(x, "batch")
+    assert y.type.dims == ("batch",)
+    assert y.type.shape == (1,)
+    fn = xr_function([x], y)
+    xr_assert_allclose(fn(xr_arange_like(x)), xr_arange_like(x).expand_dims("batch"))
+
+    # Static size > 1: broadcast
+    x = xtensor("x", dims=("a", "b"), shape=(2, 3))
+    y = expand_dims(x, "batch", size=4)
+    fn = xr_function([x], y)
+    expected = xr.DataArray(
+        np.broadcast_to(xr_arange_like(x).data, (4, 2, 3)),
+        dims=("batch", "a", "b"),
+        coords={"a": xr_arange_like(x).coords["a"], "b": xr_arange_like(x).coords["b"]},
+    )
+    xr_assert_allclose(fn(xr_arange_like(x)), expected)
+
+    # Insert new dim between existing dims
+    x = xtensor("x", dims=("a", "b"), shape=(2, 3))
+    y = expand_dims(x, "new")
+    # Insert new dim between a and b: ("a", "new", "b")
+    y = transpose(y, "a", "new", "b")
+    fn = xr_function([x], y)
+    x_test = xr_arange_like(x)
+    expected = x_test.expand_dims("new").transpose("a", "new", "b")
+    xr_assert_allclose(fn(x_test), expected)
+
+    # Expand with multiple dims
+    x = xtensor("x", dims=(), shape=())
+    y = expand_dims(expand_dims(x, "a"), "b")
+    fn = xr_function([x], y)
+    expected = xr_arange_like(x).expand_dims("a").expand_dims("b")
+    xr_assert_allclose(fn(xr_arange_like(x)), expected)
+
+
+def test_expand_dims_implicit():
+    """Test expand_dims with default or symbolic sizes and dim=None."""
+
+    # Symbolic size=1: same as default
+    size_sym_1 = scalar("size_sym_1", dtype="int64")
+    x = xtensor("x", dims=("a", "b"), shape=(2, 3))
+    y = expand_dims(x, "batch", size=size_sym_1)
+    fn = xr_function([x, size_sym_1], y, on_unused_input="ignore")
+    x_test = xr_arange_like(x)
+    xr_assert_allclose(fn(x_test, 1), x_test.expand_dims("batch"))
+
+    # Symbolic size > 1 (but expand only adds dim=1)
+    size_sym_4 = scalar("size_sym_4", dtype="int64")
+    y = expand_dims(x, "batch", size=size_sym_4)
+    fn = xr_function([x, size_sym_4], y, on_unused_input="ignore")
+    xr_assert_allclose(fn(x_test, 4), x_test.expand_dims("batch"))
+
+    # Symbolic size > 1 with broadcasting
+    size_sym_4 = scalar("size_sym_4", dtype="int64")
+    x = xtensor("x", dims=("a", "b"), shape=(2, 3))
+    y = expand_dims(x, "batch", size=size_sym_4)
+    z = y + y  # This should broadcast along the batch dimension
+    fn = xr_function([x, size_sym_4], z, on_unused_input="ignore")
+    x_test = xr_arange_like(x)
+    out = fn(x_test, 4)
+    expected = x_test.expand_dims("batch") + x_test.expand_dims("batch")
+    xr_assert_allclose(out, expected)
+
+    # Symbolic size with shape validation
+    size_sym = scalar("size_sym", dtype="int64")
+    x = xtensor("x", dims=("a", "b"), shape=(2, 3))
+    y = expand_dims(x, "batch", size=size_sym)
+    z = y + y  # This should validate the shape
+    fn = xr_function([x, size_sym], z, on_unused_input="ignore")
+    x_test = xr_arange_like(x)
+    out = fn(x_test, 4)
+    expected = x_test.expand_dims("batch") + x_test.expand_dims("batch")
+    xr_assert_allclose(out, expected)
+
+    # Symbolic size with subsequent operations
+    size_sym = scalar("size_sym", dtype="int64")
+    x = xtensor("x", dims=("a", "b"), shape=(2, 3))
+    y = expand_dims(x, "batch", size=size_sym)
+    z = y.sum("batch")  # This should work with symbolic size
+    fn = xr_function([x, size_sym], z, on_unused_input="ignore")
+    x_test = xr_arange_like(x)
+    out = fn(x_test, 4)
+    expected = x_test.expand_dims("batch").sum("batch")
+    xr_assert_allclose(out, expected)
+
+    # Symbolic size with transpose and broadcasting
+    size_sym = scalar("size_sym", dtype="int64")
+    x = xtensor("x", dims=("a", "b"), shape=(2, 3))
+    y = expand_dims(x, "batch", size=size_sym)
+    z = transpose(y, "batch", "a", "b")  # This should work with symbolic size
+    fn = xr_function([x, size_sym], z, on_unused_input="ignore")
+    x_test = xr_arange_like(x)
+    out = fn(x_test, 4)
+    expected = x_test.expand_dims("batch").transpose("batch", "a", "b")
+    xr_assert_allclose(out, expected)
+
+    # Reversibility: expand then squeeze
+    x = xtensor("x", dims=("a", "b"), shape=(2, 3))
+    y = expand_dims(x, "batch")
+    z = squeeze(y, "batch")
+    fn = xr_function([x], z)
+    x_test = xr_arange_like(x)
+    xr_assert_allclose(fn(x_test), x_test)
+
+    # expand_dims with dim=None = no-op
+    x = xtensor("x", dims=("a",), shape=(3,))
+    y = expand_dims(x, None)
+    fn = xr_function([x], y)
+    x_test = xr_arange_like(x)
+    xr_assert_allclose(fn(x_test), x_test)
+
+    # broadcast after symbolic size
+    size_sym = scalar("size_sym", dtype="int64")
+    x = xtensor("x", dims=("a", "b"), shape=(2, 3))
+    y = expand_dims(x, "batch", size=size_sym)
+    z = y + y  # triggers shape alignment
+    fn = xr_function([x, size_sym], z, on_unused_input="ignore")
+    x_test = xr_arange_like(x)
+    out = fn(x_test, 1)
+    expected = x_test.expand_dims("batch") + x_test.expand_dims("batch")
+    xr_assert_allclose(out, expected)
+
+
+def test_expand_dims_errors():
+    """Test error handling in expand_dims."""
+
+    # Expanding existing dim
+    x = xtensor("x", dims=("city",), shape=(3,))
+    y = expand_dims(x, "country")
+    with pytest.raises(ValueError, match="already exists"):
+        expand_dims(y, "city")
+
+    # Size = 0 is invalid
+    with pytest.raises(ValueError, match="size must be.*positive"):
+        expand_dims(x, "batch", size=0)
+
+    # Invalid dim type
+    with pytest.raises(TypeError):
+        expand_dims(x, 123)
+
+    # Invalid size type
+    with pytest.raises(TypeError):
+        expand_dims(x, "new", size=[1])
+
+    # Duplicate dimension creation
+    y = expand_dims(x, "new")
+    with pytest.raises(ValueError):
+        expand_dims(y, "new")
+
+    # Symbolic size with invalid runtime value
+    size_sym = scalar("size_sym", dtype="int64")
+    y = expand_dims(x, "batch", size=size_sym)
+    fn = xr_function([x, size_sym], y, on_unused_input="ignore")
+    with pytest.raises(Exception):
+        fn(xr_arange_like(x), 0)