Intercept UserWarning on JAX random function tests

Author: ricardoV94

tests/link/jax/test_random.py

@@ -22,6 +22,17 @@
 from pytensor.link.jax.dispatch.random import numpyro_available  # noqa: E402
 
 
+# turns all warnings into errors for this module
+pytestmark = pytest.mark.filterwarnings("error")
+
+
+def random_function(*args, **kwargs):
+    with pytest.warns(
+        UserWarning, match=r"The RandomType SharedVariables \[.+\] will not be used"
+    ):
+        return function(*args, **kwargs)
+
+
 def test_random_RandomStream():
     """Two successive calls of a compiled graph using `RandomStream` should
     return different values.
@@ -30,11 +41,7 @@ def test_random_RandomStream():
     srng = RandomStream(seed=123)
     out = srng.normal() - srng.normal()
 
-    with pytest.warns(
-        UserWarning,
-        match=r"The RandomType SharedVariables \[.+\] will not be used",
-    ):
-        fn = function([], out, mode=jax_mode)
+    fn = random_function([], out, mode=jax_mode)
     jax_res_1 = fn()
     jax_res_2 = fn()
 
@@ -47,13 +54,7 @@ def test_random_updates(rng_ctor):
     rng = shared(original_value, name="original_rng", borrow=False)
     next_rng, x = at.random.normal(name="x", rng=rng).owner.outputs
 
-    with pytest.warns(
-        UserWarning,
-        match=re.escape(
-            "The RandomType SharedVariables [original_rng] will not be used"
-        ),
-    ):
-        f = pytensor.function([], [x], updates={rng: next_rng}, mode=jax_mode)
+    f = random_function([], [x], updates={rng: next_rng}, mode=jax_mode)
     assert f() != f()
 
     # Check that original rng variable content was not overwritten when calling jax_typify
@@ -81,17 +82,14 @@ def test_random_updates_input_storage_order():
 
     # This function replaces inp by input_shared in the update expression
     # This is what caused the RNG to appear later than inp_shared in the input_storage
-    with pytest.warns(
-        UserWarning,
-        match=r"The RandomType SharedVariables \[.+\] will not be used",
-    ):
-        fn = pytensor.function(
-            inputs=[],
-            outputs=[],
-            updates={inp_shared: inp_update},
-            givens={inp: inp_shared},
-            mode="JAX",
-        )
+
+    fn = random_function(
+        inputs=[],
+        outputs=[],
+        updates={inp_shared: inp_update},
+        givens={inp: inp_shared},
+        mode="JAX",
+    )
     fn()
     np.testing.assert_allclose(inp_shared.get_value(), 5, rtol=1e-3)
     fn()
@@ -455,7 +453,7 @@ def test_random_RandomVariable(rv_op, dist_params, base_size, cdf_name, params_c
     else:
         rng = shared(np.random.RandomState(29402))
     g = rv_op(*dist_params, size=(10_000,) + base_size, rng=rng)
-    g_fn = function(dist_params, g, mode=jax_mode)
+    g_fn = random_function(dist_params, g, mode=jax_mode)
     samples = g_fn(
         *[
             i.tag.test_value
@@ -479,7 +477,7 @@ def test_random_RandomVariable(rv_op, dist_params, base_size, cdf_name, params_c
 def test_random_bernoulli(size):
     rng = shared(np.random.RandomState(123))
     g = at.random.bernoulli(0.5, size=(1000,) + size, rng=rng)
-    g_fn = function([], g, mode=jax_mode)
+    g_fn = random_function([], g, mode=jax_mode)
     samples = g_fn()
     np.testing.assert_allclose(samples.mean(axis=0), 0.5, 1)
 
@@ -490,7 +488,7 @@ def test_random_mvnormal():
     mu = np.ones(4)
     cov = np.eye(4)
     g = at.random.multivariate_normal(mu, cov, size=(10000,), rng=rng)
-    g_fn = function([], g, mode=jax_mode)
+    g_fn = random_function([], g, mode=jax_mode)
     samples = g_fn()
     np.testing.assert_allclose(samples.mean(axis=0), mu, atol=0.1)
 
@@ -505,7 +503,7 @@ def test_random_mvnormal():
 def test_random_dirichlet(parameter, size):
     rng = shared(np.random.RandomState(123))
     g = at.random.dirichlet(parameter, size=(1000,) + size, rng=rng)
-    g_fn = function([], g, mode=jax_mode)
+    g_fn = random_function([], g, mode=jax_mode)
     samples = g_fn()
     np.testing.assert_allclose(samples.mean(axis=0), 0.5, 1)
 
@@ -515,29 +513,29 @@ def test_random_choice():
     num_samples = 10000
     rng = shared(np.random.RandomState(123))
     g = at.random.choice(np.arange(4), size=num_samples, rng=rng)
-    g_fn = function([], g, mode=jax_mode)
+    g_fn = random_function([], g, mode=jax_mode)
     samples = g_fn()
     np.testing.assert_allclose(np.sum(samples == 3) / num_samples, 0.25, 2)
 
     # `replace=False` produces unique results
     rng = shared(np.random.RandomState(123))
     g = at.random.choice(np.arange(100), replace=False, size=99, rng=rng)
-    g_fn = function([], g, mode=jax_mode)
+    g_fn = random_function([], g, mode=jax_mode)
     samples = g_fn()
     assert len(np.unique(samples)) == 99
 
     # We can pass an array with probabilities
     rng = shared(np.random.RandomState(123))
     g = at.random.choice(np.arange(3), p=np.array([1.0, 0.0, 0.0]), size=10, rng=rng)
-    g_fn = function([], g, mode=jax_mode)
+    g_fn = random_function([], g, mode=jax_mode)
     samples = g_fn()
     np.testing.assert_allclose(samples, np.zeros(10))
 
 
 def test_random_categorical():
     rng = shared(np.random.RandomState(123))
     g = at.random.categorical(0.25 * np.ones(4), size=(10000, 4), rng=rng)
-    g_fn = function([], g, mode=jax_mode)
+    g_fn = random_function([], g, mode=jax_mode)
     samples = g_fn()
     np.testing.assert_allclose(samples.mean(axis=0), 6 / 4, 1)
 
@@ -546,7 +544,7 @@ def test_random_permutation():
     array = np.arange(4)
     rng = shared(np.random.RandomState(123))
     g = at.random.permutation(array, rng=rng)
-    g_fn = function([], g, mode=jax_mode)
+    g_fn = random_function([], g, mode=jax_mode)
     permuted = g_fn()
     with pytest.raises(AssertionError):
         np.testing.assert_allclose(array, permuted)
@@ -556,7 +554,7 @@ def test_random_geometric():
     rng = shared(np.random.RandomState(123))
     p = np.array([0.3, 0.7])
     g = at.random.geometric(p, size=(10_000, 2), rng=rng)
-    g_fn = function([], g, mode=jax_mode)
+    g_fn = random_function([], g, mode=jax_mode)
     samples = g_fn()
     np.testing.assert_allclose(samples.mean(axis=0), 1 / p, rtol=0.1)
     np.testing.assert_allclose(samples.std(axis=0), np.sqrt((1 - p) / p**2), rtol=0.1)
@@ -567,7 +565,7 @@ def test_negative_binomial():
     n = np.array([10, 40])
     p = np.array([0.3, 0.7])
     g = at.random.negative_binomial(n, p, size=(10_000, 2), rng=rng)
-    g_fn = function([], g, mode=jax_mode)
+    g_fn = random_function([], g, mode=jax_mode)
     samples = g_fn()
     np.testing.assert_allclose(samples.mean(axis=0), n * (1 - p) / p, rtol=0.1)
     np.testing.assert_allclose(
@@ -581,7 +579,7 @@ def test_binomial():
     n = np.array([10, 40])
     p = np.array([0.3, 0.7])
     g = at.random.binomial(n, p, size=(10_000, 2), rng=rng)
-    g_fn = function([], g, mode=jax_mode)
+    g_fn = random_function([], g, mode=jax_mode)
     samples = g_fn()
     np.testing.assert_allclose(samples.mean(axis=0), n * p, rtol=0.1)
     np.testing.assert_allclose(samples.std(axis=0), np.sqrt(n * p * (1 - p)), rtol=0.1)
@@ -596,7 +594,7 @@ def test_beta_binomial():
     a = np.array([1.5, 13])
     b = np.array([0.5, 9])
     g = at.random.betabinom(n, a, b, size=(10_000, 2), rng=rng)
-    g_fn = function([], g, mode=jax_mode)
+    g_fn = random_function([], g, mode=jax_mode)
     samples = g_fn()
     np.testing.assert_allclose(samples.mean(axis=0), n * a / (a + b), rtol=0.1)
     np.testing.assert_allclose(
@@ -614,7 +612,7 @@ def test_multinomial():
     n = np.array([10, 40])
     p = np.array([[0.3, 0.7, 0.0], [0.1, 0.4, 0.5]])
     g = at.random.multinomial(n, p, size=(10_000, 2), rng=rng)
-    g_fn = function([], g, mode=jax_mode)
+    g_fn = random_function([], g, mode=jax_mode)
     samples = g_fn()
     np.testing.assert_allclose(samples.mean(axis=0), n[..., None] * p, rtol=0.1)
     np.testing.assert_allclose(
@@ -630,7 +628,7 @@ def test_vonmises_mu_outside_circle():
     mu = np.array([-30, 40])
     kappa = np.array([100, 10])
     g = at.random.vonmises(mu, kappa, size=(10_000, 2), rng=rng)
-    g_fn = function([], g, mode=jax_mode)
+    g_fn = random_function([], g, mode=jax_mode)
     samples = g_fn()
     np.testing.assert_allclose(
         samples.mean(axis=0), (mu + np.pi) % (2.0 * np.pi) - np.pi, rtol=0.1
@@ -676,7 +674,10 @@ def rng_fn(cls, rng, size):
     fgraph = FunctionGraph([out.owner.inputs[0]], [out], clone=False)
 
     with pytest.raises(NotImplementedError):
-        compare_jax_and_py(fgraph, [])
+        with pytest.warns(
+            UserWarning, match=r"The RandomType SharedVariables \[.+\] will not be used"
+        ):
+            compare_jax_and_py(fgraph, [])
 
 
 def test_random_custom_implementation():
@@ -707,7 +708,10 @@ def sample_fn(rng, size, dtype, *parameters):
     rng = shared(np.random.RandomState(123))
     out = nonexistentrv(rng=rng)
     fgraph = FunctionGraph([out.owner.inputs[0]], [out], clone=False)
-    compare_jax_and_py(fgraph, [])
+    with pytest.warns(
+        UserWarning, match=r"The RandomType SharedVariables \[.+\] will not be used"
+    ):
+        compare_jax_and_py(fgraph, [])
 
 
 def test_random_concrete_shape():
@@ -724,19 +728,15 @@ def test_random_concrete_shape():
     rng = shared(np.random.RandomState(123))
     x_at = at.dmatrix()
     out = at.random.normal(0, 1, size=x_at.shape, rng=rng)
-    jax_fn = function([x_at], out, mode=jax_mode)
+    jax_fn = random_function([x_at], out, mode=jax_mode)
     assert jax_fn(np.ones((2, 3))).shape == (2, 3)
 
 
 def test_random_concrete_shape_from_param():
     rng = shared(np.random.RandomState(123))
     x_at = at.dmatrix()
     out = at.random.normal(x_at, 1, rng=rng)
-    with pytest.warns(
-        UserWarning,
-        match="The RandomType SharedVariables \[.+\] will not be used"
-    ):
-        jax_fn = function([x_at], out, mode=jax_mode)
+    jax_fn = random_function([x_at], out, mode=jax_mode)
     assert jax_fn(np.ones((2, 3))).shape == (2, 3)
 
 
@@ -755,7 +755,7 @@ def test_random_concrete_shape_subtensor():
     rng = shared(np.random.RandomState(123))
     x_at = at.dmatrix()
     out = at.random.normal(0, 1, size=x_at.shape[1], rng=rng)
-    jax_fn = function([x_at], out, mode=jax_mode)
+    jax_fn = random_function([x_at], out, mode=jax_mode)
     assert jax_fn(np.ones((2, 3))).shape == (3,)
 
 
@@ -771,7 +771,7 @@ def test_random_concrete_shape_subtensor_tuple():
     rng = shared(np.random.RandomState(123))
     x_at = at.dmatrix()
     out = at.random.normal(0, 1, size=(x_at.shape[0],), rng=rng)
-    jax_fn = function([x_at], out, mode=jax_mode)
+    jax_fn = random_function([x_at], out, mode=jax_mode)
     assert jax_fn(np.ones((2, 3))).shape == (2,)
 
 
@@ -782,5 +782,5 @@ def test_random_concrete_shape_graph_input():
     rng = shared(np.random.RandomState(123))
     size_at = at.scalar()
     out = at.random.normal(0, 1, size=size_at, rng=rng)
-    jax_fn = function([size_at], out, mode=jax_mode)
+    jax_fn = random_function([size_at], out, mode=jax_mode)
     assert jax_fn(10).shape == (10,)