Implement shape inference for boolean advanced indexing

Author: ricardoV94

pytensor/tensor/extra_ops.py

@@ -21,14 +21,18 @@
 from pytensor.raise_op import Assert
 from pytensor.scalar import int32 as int_t
 from pytensor.scalar import upcast
+from pytensor.tensor import as_tensor_variable
 from pytensor.tensor import basic as at
 from pytensor.tensor import get_vector_length
 from pytensor.tensor.exceptions import NotScalarConstantError
-from pytensor.tensor.math import abs as at_abs
+from pytensor.tensor.math import abs as pt_abs
 from pytensor.tensor.math import all as pt_all
 from pytensor.tensor.math import eq as pt_eq
-from pytensor.tensor.math import ge, lt, maximum, minimum, prod
+from pytensor.tensor.math import ge, lt
+from pytensor.tensor.math import max as pt_max
+from pytensor.tensor.math import maximum, minimum, prod
 from pytensor.tensor.math import sum as at_sum
+from pytensor.tensor.math import switch
 from pytensor.tensor.subtensor import advanced_inc_subtensor1, set_subtensor
 from pytensor.tensor.type import TensorType, dvector, int_dtypes, integer_dtypes, vector
 from pytensor.tensor.var import TensorVariable
@@ -1063,7 +1067,7 @@ def grad(self, inp, cost_grad):
         # only valid for matrices
         wr_a = fill_diagonal_offset(grad, 0, offset)
 
-        offset_abs = at_abs(offset)
+        offset_abs = pt_abs(offset)
         pos_offset_flag = ge(offset, 0)
         neg_offset_flag = lt(offset, 0)
         min_wh = minimum(width, height)
@@ -1442,6 +1446,7 @@ def ravel_multi_index(multi_index, dims, mode="raise", order="C"):
     "axes that have a statically known length 1. Use `specify_broadcastable` to "
     "inform PyTensor of a known shape."
 )
+_runtime_broadcast_assert = Assert("Could not broadcast dimensions.")
 
 
 def broadcast_shape(*arrays, **kwargs) -> Tuple[aes.ScalarVariable, ...]:
@@ -1465,6 +1470,7 @@ def broadcast_shape(*arrays, **kwargs) -> Tuple[aes.ScalarVariable, ...]:
 def broadcast_shape_iter(
     arrays: Iterable[Union[TensorVariable, Tuple[TensorVariable, ...]]],
     arrays_are_shapes: bool = False,
+    allow_runtime_broadcast: bool = False,
 ) -> Tuple[aes.ScalarVariable, ...]:
     r"""Compute the shape resulting from broadcasting arrays.
 
@@ -1480,22 +1486,24 @@ def broadcast_shape_iter(
     arrays
         An iterable of tensors, or a tuple of shapes (as tuples),
         for which the broadcast shape is computed.
-    arrays_are_shapes
+    arrays_are_shapes: bool, default False
         Indicates whether or not the `arrays` contains shape tuples.
         If you use this approach, make sure that the broadcastable dimensions
         are (scalar) constants with the value ``1``--or simply the integer
-        ``1``.
+        ``1``. This is not revelant if `allow_runtime_broadcast` is True.
+    allow_runtime_broadcast: bool, default False
+        Whether to allow non-statically known broadcast on the shape computation.
 
     """
-    one_at = pytensor.scalar.ScalarConstant(pytensor.scalar.int64, 1)
+    one = pytensor.scalar.ScalarConstant(pytensor.scalar.int64, 1)
 
     if arrays_are_shapes:
         max_dims = max(len(a) for a in arrays)
 
         array_shapes = [
-            (one_at,) * (max_dims - len(a))
+            (one,) * (max_dims - len(a))
             + tuple(
-                one_at
+                one
                 if sh == 1 or isinstance(sh, Constant) and sh.value == 1
                 else (aes.as_scalar(sh) if not isinstance(sh, Variable) else sh)
                 for sh in a
@@ -1508,10 +1516,8 @@ def broadcast_shape_iter(
         _arrays = tuple(at.as_tensor_variable(a) for a in arrays)
 
         array_shapes = [
-            (one_at,) * (max_dims - a.ndim)
-            + tuple(
-                one_at if t_sh == 1 else sh for sh, t_sh in zip(a.shape, a.type.shape)
-            )
+            (one,) * (max_dims - a.ndim)
+            + tuple(one if t_sh == 1 else sh for sh, t_sh in zip(a.shape, a.type.shape))
             for a in _arrays
         ]
 
@@ -1520,11 +1526,11 @@ def broadcast_shape_iter(
     for dim_shapes in zip(*array_shapes):
         # Get the shapes in this dimension that are not broadcastable
         # (i.e. not symbolically known to be broadcastable)
-        non_bcast_shapes = [shape for shape in dim_shapes if shape != one_at]
+        non_bcast_shapes = [shape for shape in dim_shapes if shape != one]
 
         if len(non_bcast_shapes) == 0:
             # Every shape was broadcastable in this dimension
-            result_dims.append(one_at)
+            result_dims.append(one)
         elif len(non_bcast_shapes) == 1:
             # Only one shape might not be broadcastable in this dimension
             result_dims.extend(non_bcast_shapes)
@@ -1554,9 +1560,26 @@ def broadcast_shape_iter(
                 result_dims.append(first_length)
                 continue
 
-            # Add assert that all remaining shapes are equal
-            condition = pt_all([pt_eq(first_length, other) for other in other_lengths])
-            result_dims.append(_broadcast_assert(first_length, condition))
+            if not allow_runtime_broadcast:
+                # Add assert that all remaining shapes are equal
+                condition = pt_all(
+                    [pt_eq(first_length, other) for other in other_lengths]
+                )
+                result_dims.append(_broadcast_assert(first_length, condition))
+            else:
+                lengths = as_tensor_variable((first_length, *other_lengths))
+                runtime_broadcastable = pt_eq(lengths, one)
+                result_dim = pt_abs(
+                    pt_max(switch(runtime_broadcastable, -one, lengths))
+                )
+                condition = pt_all(
+                    switch(
+                        ~runtime_broadcastable,
+                        pt_eq(lengths, result_dim),
+                        np.array(True),
+                    )
+                )
+                result_dims.append(_runtime_broadcast_assert(result_dim, condition))
 
     return tuple(result_dims)
 

pytensor/tensor/subtensor.py

@@ -20,15 +20,11 @@
 from pytensor.printing import Printer, pprint, set_precedence
 from pytensor.scalar.basic import ScalarConstant
 from pytensor.tensor import _get_vector_length, as_tensor_variable, get_vector_length
-from pytensor.tensor.basic import alloc, get_underlying_scalar_constant_value
+from pytensor.tensor.basic import alloc, get_underlying_scalar_constant_value, nonzero
 from pytensor.tensor.elemwise import DimShuffle
-from pytensor.tensor.exceptions import (
-    AdvancedIndexingError,
-    NotScalarConstantError,
-    ShapeError,
-)
+from pytensor.tensor.exceptions import AdvancedIndexingError, NotScalarConstantError
 from pytensor.tensor.math import clip
-from pytensor.tensor.shape import Reshape, specify_broadcastable
+from pytensor.tensor.shape import Reshape, shape_i, specify_broadcastable
 from pytensor.tensor.type import (
     TensorType,
     bscalar,
@@ -510,7 +506,11 @@ def indexed_result_shape(array_shape, indices, indices_are_shapes=False):
             from pytensor.tensor.extra_ops import broadcast_shape
 
             res_shape += broadcast_shape(
-                *grp_indices, arrays_are_shapes=indices_are_shapes
+                *grp_indices,
+                arrays_are_shapes=indices_are_shapes,
+                # The AdvancedIndexing Op relies on the Numpy implementation which allows runtime broadcasting.
+                # As long as that is true, the shape inference has to respect that this is not an error.
+                allow_runtime_broadcast=True,
             )
 
     res_shape += tuple(array_shape[dim] for dim in remaining_dims)
@@ -2584,26 +2584,47 @@ def R_op(self, inputs, eval_points):
         return self.make_node(eval_points[0], *inputs[1:]).outputs
 
     def infer_shape(self, fgraph, node, ishapes):
-        indices = node.inputs[1:]
-        index_shapes = list(ishapes[1:])
-        for i, idx in enumerate(indices):
-            if (
+        def is_bool_index(idx):
+            return (
                 isinstance(idx, (np.bool_, bool))
                 or getattr(idx, "dtype", None) == "bool"
-            ):
-                raise ShapeError(
-                    "Shape inference for boolean indices is not implemented"
+            )
+
+        indices = node.inputs[1:]
+        index_shapes = []
+        for idx, ishape in zip(indices, ishapes[1:]):
+            # Mixed bool indexes are converted to nonzero entries
+            if is_bool_index(idx):
+                index_shapes.extend(
+                    (shape_i(nz_dim, 0, fgraph=fgraph),) for nz_dim in nonzero(idx)
                 )
             # The `ishapes` entries for `SliceType`s will be None, and
             # we need to give `indexed_result_shape` the actual slices.
-            if isinstance(getattr(idx, "type", None), SliceType):
-                index_shapes[i] = idx
+            elif isinstance(getattr(idx, "type", None), SliceType):
+                index_shapes.append(idx)
+            else:
+                index_shapes.append(ishape)
 
-        res_shape = indexed_result_shape(
-            ishapes[0], index_shapes, indices_are_shapes=True
+        res_shape = list(
+            indexed_result_shape(ishapes[0], index_shapes, indices_are_shapes=True)
         )
+
+        adv_indices = [idx for idx in indices if not is_basic_idx(idx)]
+        bool_indices = [idx for idx in adv_indices if is_bool_index(idx)]
+
+        # Special logic when the only advanced index group is of bool type.
+        # We can replace the nonzeros by a sum of the whole bool variable.
+        if len(bool_indices) == 1 and len(adv_indices) == 1:
+            [bool_index] = bool_indices
+            # Find the output dim associated with the bool index group
+            # Because there are no more advanced index groups, there is exactly
+            # one output dim per index variable up to the bool group.
+            # Note: Scalar integer indexing counts as advanced indexing.
+            start_dim = indices.index(bool_index)
+            res_shape[start_dim] = bool_index.sum()
+
         assert node.outputs[0].ndim == len(res_shape)
-        return [list(res_shape)]
+        return [res_shape]
 
     def perform(self, node, inputs, out_):
         (out,) = out_

tests/tensor/test_extra_ops.py

@@ -1087,9 +1087,17 @@ def shape_tuple(x, use_bcast=True):
     assert any(
         isinstance(node.op, Assert) for node in applys_between([x_at, y_at], b_at)
     )
-    # This should fail because it would need dynamic broadcasting
     with pytest.raises(AssertionError):
         assert np.array_equal([z.eval() for z in b_at], b.shape)
+    # But fine if we allow_runtime_broadcast
+    b_at = broadcast_shape(
+        shape_tuple(x_at, use_bcast=False),
+        shape_tuple(y_at, use_bcast=False),
+        arrays_are_shapes=True,
+        allow_runtime_broadcast=True,
+    )
+    assert np.array_equal([z.eval() for z in b_at], b.shape)
+    # Or if static bcast is known
     b_at = broadcast_shape(shape_tuple(x_at), shape_tuple(y_at), arrays_are_shapes=True)
     assert np.array_equal([z.eval() for z in b_at], b.shape)
 

tests/tensor/test_subtensor.py

@@ -63,6 +63,7 @@
     tensor,
     tensor3,
     tensor4,
+    tensor5,
     vector,
 )
 from pytensor.tensor.type_other import NoneConst, SliceConstant, make_slice, slicetype
@@ -2150,6 +2151,12 @@ def fun(x, y):
 
 
 class TestInferShape(utt.InferShapeTester):
+    @staticmethod
+    def random_bool_mask(shape, rng=None):
+        if rng is None:
+            rng = np.random.default_rng()
+        return rng.binomial(n=1, p=0.5, size=shape).astype(bool)
+
     def test_IncSubtensor(self):
         admat = dmatrix()
         bdmat = dmatrix()
@@ -2439,25 +2446,85 @@ def test_AdvancedSubtensor_bool(self):
         n = dmatrix()
         n_val = np.arange(6).reshape((2, 3))
 
-        # infer_shape is not implemented, but it should not crash
+        # Shape inference requires runtime broadcasting between the nonzero() shapes
         self._compile_and_check(
             [n],
             [n[n[:, 0] > 2, n[0, :] > 2]],
             [n_val],
             AdvancedSubtensor,
-            check_topo=False,
         )
         self._compile_and_check(
             [n],
             [n[n[:, 0] > 2]],
             [n_val],
             AdvancedSubtensor,
-            check_topo=False,
+        )
+        self._compile_and_check(
+            [n],
+            [n[:, np.array([True, False, True])]],
+            [n_val],
+            AdvancedSubtensor,
+        )
+        self._compile_and_check(
+            [n],
+            [n[np.array([False, False]), 1:]],
+            [n_val],
+            AdvancedSubtensor,
+        )
+        self._compile_and_check(
+            [n],
+            [n[np.array([True, True]), 0]],
+            [n_val],
+            AdvancedSubtensor,
+        )
+        self._compile_and_check(
+            [n],
+            [n[self.random_bool_mask(n_val.shape)]],
+            [n_val],
+            AdvancedSubtensor,
+        )
+        self._compile_and_check(
+            [n],
+            [n[None, self.random_bool_mask(n_val.shape), None]],
+            [n_val],
+            AdvancedSubtensor,
+        )
+        self._compile_and_check(
+            [n],
+            [n[slice(5, None), self.random_bool_mask(n_val.shape[1])]],
+            [n_val],
+            AdvancedSubtensor,
         )
 
         abs_res = n[~isinf(n)]
         assert abs_res.type.shape == (None,)
 
+    def test_AdvancedSubtensor_bool_mixed(self):
+        n = tensor5("x", dtype="float64")
+        shape = (18, 3, 4, 5, 6)
+        n_val = np.arange(np.prod(shape)).reshape(shape)
+        self._compile_and_check(
+            [n],
+            # Consecutive advanced index
+            [n[1:, self.random_bool_mask((3, 4)), 0, 1:]],
+            [n_val],
+            AdvancedSubtensor,
+        )
+        self._compile_and_check(
+            [n],
+            # Non-consecutive advanced index
+            [n[1:, self.random_bool_mask((3, 4)), 1:, 0]],
+            [n_val],
+            AdvancedSubtensor,
+        )
+        self._compile_and_check(
+            [n],
+            # Non-consecutive advanced index
+            [n[1:, self.random_bool_mask((3,)), 1:, None, np.zeros((6, 1), dtype=int)]],
+            [n_val],
+            AdvancedSubtensor,
+        )
+
 
 @config.change_flags(compute_test_value="raise")
 def test_basic_shape():