Implement vectorize_node for CheckAndRaise Op

Author: ricardoV94

pytensor/raise_op.py

@@ -6,6 +6,7 @@
 
 from pytensor.gradient import DisconnectedType
 from pytensor.graph.basic import Apply, Variable
+from pytensor.graph.replace import _vectorize_node
 from pytensor.link.c.op import COp
 from pytensor.link.c.params_type import ParamsType
 from pytensor.link.c.type import Generic
@@ -198,3 +199,19 @@ def __str__(self):
 
 
 assert_op = Assert()
+
+
+@_vectorize_node.register(CheckAndRaise)
+def vectorize_check_and_raise(op, node, batch_x, batch_cond):
+    from pytensor.tensor.extra_ops import broadcast_arrays
+    from pytensor.tensor.shape import shape_padright
+
+    batch_cond_dims = batch_cond.type.ndim
+
+    if batch_cond_dims:
+        # Condition may broadcast batch dims of x
+        x_core_ndim = node.inputs[0].type.ndim
+        batch_x, _ = broadcast_arrays(batch_x, shape_padright(batch_cond, x_core_ndim))
+        batch_cond = batch_cond.all()
+
+    return op.make_node(batch_x, batch_cond)

tests/test_raise_op.py

@@ -5,6 +5,7 @@
 import pytensor
 import pytensor.tensor as pt
 from pytensor.compile.mode import OPT_FAST_RUN, Mode
+from pytensor.graph import vectorize_graph
 from pytensor.graph.basic import Constant, equal_computations
 from pytensor.raise_op import Assert, CheckAndRaise, assert_op
 from pytensor.scalar.basic import ScalarType, float64
@@ -184,3 +185,64 @@ def test_CheckAndRaise_sparse_variable():
     a2 = check_and_raise(aspe1, aspe2.sum() > 2)
     with pytest.raises(ValueError, match="sparse_check"):
         a2.sum().eval()
+
+
+@pytensor.config.change_flags(cxx="")  # For speed-up
+def test_vectorize():
+    floatX = pytensor.config.floatX
+    x = pt.vector("x")
+    y = pt.vector("y")
+    cond = pt.all(y >= 0)
+    out = assert_op(x, cond)
+
+    batch_x = pt.matrix("batch_x", shape=(2, None))
+    batch_y = pt.matrix("batch_y", shape=(2, None))
+
+    test_x = np.arange(3).astype(floatX)
+    test_y = np.arange(4).astype(floatX)
+    test_batch_x = np.arange(6).reshape(2, 3).astype(floatX)
+    test_batch_y = np.arange(8).reshape(2, 4).astype(floatX)
+
+    # Only x is batched
+    vect_out = vectorize_graph(out, {x: batch_x, y: y})
+    assert vect_out.type.shape == (2, None)
+    assert isinstance(vect_out.owner.op, CheckAndRaise)
+    np.testing.assert_array_equal(
+        vect_out.eval({batch_x: test_batch_x, y: test_y}),
+        test_batch_x,
+    )
+    with pytest.raises(AssertionError):
+        vect_out.eval({batch_x: test_batch_x, y: -test_y})
+
+    # Only y is batched
+    vect_out = vectorize_graph(out, {x: x, y: batch_y})
+    assert vect_out.type.shape == (2, None)
+    assert isinstance(vect_out.owner.op, CheckAndRaise)
+    np.testing.assert_array_equal(
+        vect_out.eval({x: test_x, batch_y: test_batch_y}),
+        np.broadcast_to(test_x, (2, *test_x.shape)),
+    )
+    with pytest.raises(AssertionError):
+        vect_out.eval({x: test_x, batch_y: -test_batch_y})
+
+    # Both x, and y are batched
+    vect_out = vectorize_graph(out, {x: batch_x, y: batch_y})
+    assert vect_out.type.shape == (2, None)
+    assert isinstance(vect_out.owner.op, CheckAndRaise)
+    np.testing.assert_array_equal(
+        vect_out.eval({batch_x: test_batch_x, batch_y: test_batch_y}),
+        test_batch_x,
+    )
+    with pytest.raises(AssertionError):
+        vect_out.eval({batch_x: test_batch_x, batch_y: -test_batch_y})
+
+    # Both x, and y are batched and broadcast each other
+    vect_out = vectorize_graph(out, {x: batch_x[:, None, :], y: batch_y[None, :, :]})
+    assert vect_out.type.shape == (2, 2, None)
+    assert isinstance(vect_out.owner.op, CheckAndRaise)
+    np.testing.assert_array_equal(
+        vect_out.eval({batch_x: test_batch_x, batch_y: test_batch_y}),
+        np.broadcast_to(test_batch_x[:, None, :], (2, *test_batch_x.shape)),
+    )
+    with pytest.raises(AssertionError):
+        vect_out.eval({batch_x: test_batch_x, batch_y: -test_batch_y})