Use infer_shape of core_op to infer Blockwise core shapes

This can only be done when the output of infer_shape of the core_op depends only on the input shapes, and not their values.

Author: ricardoV94

pytensor/tensor/blockwise.py

@@ -7,7 +7,8 @@
 from pytensor import config
 from pytensor.compile.builders import OpFromGraph
 from pytensor.gradient import DisconnectedType
-from pytensor.graph.basic import Apply, Constant
+from pytensor.graph import FunctionGraph
+from pytensor.graph.basic import Apply, Constant, ancestors
 from pytensor.graph.null_type import NullType
 from pytensor.graph.op import Op
 from pytensor.graph.replace import (
@@ -179,15 +180,37 @@ def infer_shape(
 
         batch_shape = broadcast_shape(*batch_shapes, arrays_are_shapes=True)
 
+        # Try to extract the core shapes from the core_op
+        if hasattr(self.core_op, "infer_shape"):
+            dummy_core_node = self._create_dummy_core_node(node.inputs)
+            dummy_core_inputs = dummy_core_node.inputs
+            dummy_fgraph = FunctionGraph(outputs=dummy_core_node.outputs, clone=False)
+            core_input_shapes = [
+                input_shape[batch_ndims:] for input_shape in input_shapes
+            ]
+            core_output_shapes = self.core_op.infer_shape(
+                dummy_fgraph, dummy_core_node, core_input_shapes
+            )
+
         out_shapes = []
-        for output, sig in zip(node.outputs, self.outputs_sig):
+        for o, (output, sig) in enumerate(zip(node.outputs, self.outputs_sig)):
             core_out_shape = []
             for i, dim_name in enumerate(sig):
                 # The output dim is the same as another input dim
                 if dim_name in core_dims:
                     core_out_shape.append(core_dims[dim_name])
                 else:
-                    # TODO: We could try to make use of infer_shape of core_op
+                    if hasattr(self.core_op, "infer_shape"):
+                        # If the input values are needed to compute the dimension length, we can't use the infer_shape
+                        # of the core_node as the value is not constant across batch dims of the Blockwise
+                        core_out_dim = core_output_shapes[o][i]
+                        if not (
+                            set(dummy_core_inputs) & set(ancestors([core_out_dim]))
+                        ):
+                            core_out_shape.append(core_out_dim)
+                            continue
+
+                    # Fallback shape requires evaluating the Blockwise Op
                     core_out_shape.append(Shape_i(batch_ndims + i)(output))
             out_shapes.append((*batch_shape, *core_out_shape))
 

tests/tensor/test_blockwise.py

@@ -215,6 +215,58 @@ def test_blockwise_shape():
     assert tuple(shape_fn(inp1_test, inp2_test)[1]) == (7, 5, 4)
 
 
+def test_blockwise_infer_core_shape():
+    class TestOpWithInferShape(Op):
+        def make_node(self, a, b):
+            assert a.type.ndim == 1
+            assert b.type.ndim == 1
+            c = tensor(shape=(None,))
+            d = tensor(shape=(None,))
+            return Apply(self, [a, b], [c, d])
+
+        def perform(self, node, inputs, outputs):
+            a, b = inputs
+            c, d = outputs
+            c[0] = np.arange(a.size + b.size)
+            d[0] = np.arange(a.sum() + b.sum())
+
+        def infer_shape(self, fgraph, node, input_shapes):
+            # First output shape depends only on input_shapes
+            # Second output shape depends on input values
+            x, y = node.inputs
+            [(x_shape,), (y_shape,)] = input_shapes
+            return (x_shape + y_shape,), (x.sum() + y.sum(),)
+
+    blockwise_op = Blockwise(
+        core_op=TestOpWithInferShape(), signature="(a),(b)->(c),(d)"
+    )
+
+    a = tensor("a", shape=(5, 3))
+    b = tensor("b", shape=(1, 4))
+    c, d = blockwise_op(a, b)
+    assert c.type.shape == (5, None)
+    assert d.type.shape == (5, None)
+
+    c_shape_fn = pytensor.function([a, b], c.shape)
+    # c_shape can be computed from the input shapes alone
+    assert not any(
+        isinstance(getattr(n.op, "core_op", n.op), TestOpWithInferShape)
+        for n in c_shape_fn.maker.fgraph.apply_nodes
+    )
+
+    d_shape_fn = pytensor.function([a, b], d.shape)
+    # d_shape cannot be computed from the input shapes alone
+    assert any(
+        isinstance(getattr(n.op, "core_op", n.op), TestOpWithInferShape)
+        for n in d_shape_fn.maker.fgraph.apply_nodes
+    )
+
+    a_test = np.zeros(a.type.shape)
+    b_test = np.zeros(b.type.shape)
+    assert tuple(c_shape_fn(a_test, b_test)) == (5, 7)
+    assert tuple(d_shape_fn(a_test, b_test)) == (5, 0)
+
+
 class BlockwiseOpTester:
     """Base class to test Blockwise works for specific Ops"""