Support PyTensor deterministic operations as observations

pymc/data.py

@@ -33,6 +33,7 @@
 from pytensor.scalar import Cast
 from pytensor.tensor.elemwise import Elemwise
 from pytensor.tensor.random.basic import IntegersRV
+from pytensor.tensor.random.var import RandomGeneratorSharedVariable
 from pytensor.tensor.type import TensorType
 from pytensor.tensor.variable import TensorConstant, TensorVariable
 
@@ -148,6 +149,24 @@ def __str__(self):
         return "Minibatch"
 
 
+def first_inputs(r):
+    if not r.owner:
+        return
+
+    inputs = r.owner.inputs
+
+    if not inputs:
+        return
+
+    first_input = inputs[0]
+    yield first_input
+    yield from first_inputs(first_input)
+
+
+def has_random_ancestor(r):
+    return any(isinstance(i, RandomGeneratorSharedVariable) for i in first_inputs(r))
+
+
 def is_valid_observed(v) -> bool:
     if not isinstance(v, Variable):
         # Non-symbolic constant
@@ -165,6 +184,7 @@ def is_valid_observed(v) -> bool:
             and isinstance(v.owner.op.scalar_op, Cast)
             and is_valid_observed(v.owner.inputs[0])
         )
+        or not has_random_ancestor(v)
         # Or Minibatch
         or (
             isinstance(v.owner.op, MinibatchOp)

tests/test_data.py

@@ -623,3 +623,14 @@ def test_multiple_vars(self):
         [draw_mA, draw_mB] = pm.draw([mA, mB])
         assert draw_mA.shape == (10,)
         np.testing.assert_allclose(draw_mA, -draw_mB)
+
+
+def test_scaling_data_works_in_likelihood() -> None:
+    data = np.array([10, 11, 12, 13, 14, 15])
+
+    with pm.Model() as model:
+        target = pm.Data("target", data)
+        scale = 12
+        scaled_target = target / scale
+        mu = pm.Normal("mu", mu=0, sigma=1)
+        pm.Normal("x", mu=mu, sigma=1, observed=scaled_target)