Avoid missing_rvs warning when using RandomWalk

pymc/distributions/timeseries.py

@@ -21,7 +21,7 @@
 import pytensor
 import pytensor.tensor as at
 
-from pytensor.graph.basic import Node
+from pytensor.graph.basic import Node, ancestors
 from pytensor.graph.replace import clone_replace
 from pytensor.tensor import TensorVariable
 from pytensor.tensor.random.op import RandomVariable
@@ -33,7 +33,7 @@
     _moment,
     moment,
 )
-from pymc.distributions.logprob import ignore_logprob, logp
+from pymc.distributions.logprob import ignore_logprob, logp, reconsider_logprob
 from pymc.distributions.multivariate import MvNormal, MvStudentT
 from pymc.distributions.shape_utils import (
     _change_dist_size,
@@ -106,6 +106,15 @@ def dist(cls, init_dist, innovation_dist, steps=None, **kwargs) -> at.TensorVari
                 "init_dist and innovation_dist must have the same support dimensionality"
             )
 
+        # We need to check this, because we clone the variables when we ignore their logprob next
+        if init_dist in ancestors([innovation_dist]) or innovation_dist in ancestors([init_dist]):
+            raise ValueError("init_dist and innovation_dist must be completely independent")
+
+        # PyMC should not be concerned that these variables don't have values, as they will be
+        # accounted for in the logp of RandomWalk
+        init_dist = ignore_logprob(init_dist)
+        innovation_dist = ignore_logprob(innovation_dist)
+
         steps = cls.get_steps(
             innovation_dist=innovation_dist,
             steps=steps,
@@ -225,12 +234,14 @@ def random_walk_moment(op, rv, init_dist, innovation_dist, steps):
 
 
 @_logprob.register(RandomWalkRV)
-def random_walk_logp(op, values, *inputs, **kwargs):
+def random_walk_logp(op, values, init_dist, innovation_dist, steps, **kwargs):
     # Although we can derive the logprob of random walks, it does not collapse
     # what we consider the core dimension of steps. We do it manually here.
     (value,) = values
     # Recreate RV and obtain inner graph
-    rv_node = op.make_node(*inputs)
+    rv_node = op.make_node(
+        reconsider_logprob(init_dist), reconsider_logprob(innovation_dist), steps
+    )
     rv = clone_replace(
         op.inner_outputs, replace={u: v for u, v in zip(op.inner_inputs, rv_node.inputs)}
     )[op.default_output]

pymc/tests/distributions/test_timeseries.py

@@ -47,6 +47,10 @@
 from pymc.tests.distributions.util import assert_moment_is_expected
 from pymc.tests.helpers import select_by_precision
 
+# Turn all warnings into errors for this module
+# Ignoring NumPy deprecation warning tracked in https://github.com/pymc-devs/pytensor/issues/146
+pytestmark = pytest.mark.filterwarnings("error", "ignore: NumPy will stop allowing conversion")
+
 
 class TestRandomWalk:
     def test_dists_types(self):
@@ -92,6 +96,14 @@ def test_dists_not_registered_check(self):
             ):
                 RandomWalk("rw", init_dist=init_dist, innovation_dist=innovation, steps=5)
 
+    def test_dists_independent_check(self):
+        init_dist = Normal.dist()
+        innovation_dist = Normal.dist(init_dist)
+        with pytest.raises(
+            ValueError, match="init_dist and innovation_dist must be completely independent"
+        ):
+            RandomWalk.dist(init_dist=init_dist, innovation_dist=innovation_dist)
+
     @pytest.mark.parametrize(
         "init_dist, innovation_dist, steps, size, shape, "
         "init_dist_size, innovation_dist_size, rw_shape",
@@ -423,15 +435,18 @@ def test_mvgaussian_with_chol_cov_rv(self, param):
                 "chol_cov", n=3, eta=2, sd_dist=sd_dist, compute_corr=True
             )
             # pylint: enable=unpacking-non-sequence
-            if param == "chol":
-                mv = MvGaussianRandomWalk("mv", mu, chol=chol, shape=(10, 7, 3))
-            elif param == "cov":
-                mv = MvGaussianRandomWalk("mv", mu, cov=pm.math.dot(chol, chol.T), shape=(10, 7, 3))
-            else:
-                raise ValueError
+            with pytest.warns(UserWarning, match="Initial distribution not specified"):
+                if param == "chol":
+                    mv = MvGaussianRandomWalk("mv", mu, chol=chol, shape=(10, 7, 3))
+                elif param == "cov":
+                    mv = MvGaussianRandomWalk(
+                        "mv", mu, cov=pm.math.dot(chol, chol.T), shape=(10, 7, 3)
+                    )
+                else:
+                    raise ValueError
         assert draw(mv, draws=5).shape == (5, 10, 7, 3)
 
-    @pytest.mark.parametrize("param", ["cov", "chol", "tau"])
+    @pytest.mark.parametrize("param", ["scale", "chol", "tau"])
     def test_mvstudentt(self, param):
         x = MvStudentTRandomWalk.dist(
             nu=4,
@@ -853,7 +868,13 @@ def sde_fn(x, k, d, s):
         with Model() as t0:
             init_dist = pm.Normal.dist(0, 10, shape=(batch_size,))
             y = EulerMaruyama(
-                "y", dt=0.02, sde_fn=sde_fn, sde_pars=sde_pars, init_dist=init_dist, **kwargs
+                "y",
+                dt=0.02,
+                sde_fn=sde_fn,
+                sde_pars=sde_pars,
+                init_dist=init_dist,
+                initval="prior",
+                **kwargs,
             )
 
         y_eval = draw(y, draws=2, random_seed=numpy_rng)
@@ -873,6 +894,7 @@ def sde_fn(x, k, d, s):
                     sde_fn=sde_fn,
                     sde_pars=sde_pars_slice,
                     init_dist=init_dist,
+                    initval="prior",
                     **kwargs,
                 )