adapted uniform.icdf tests to run on new test suite, extended testing.py check_icdf with skip_paradomain_outside_edge_test param

Author: michaelraczycki

pymc/distributions/continuous.py

@@ -739,27 +739,6 @@ def logp(value, mu, sigma, lower, upper):
 
         return logp
 
-    def icdf(value, mu, sigma, lower, upper):
-        value = np.asarray(value)
-        eps = np.finfo(float).eps
-
-        # Standardize the lower and upper bounds
-        lower_std = (lower - mu) / sigma
-        upper_std = (upper - mu) / sigma
-
-        # Compute the cdf of the standard normal distribution at the bounds
-        Phi_a = 0.5 * (1 + at.erf(lower_std / at.sqrt(2)))
-        Phi_b = 0.5 * (1 + at.erf(upper_std / at.sqrt(2)))
-
-        # Compute the cdf of the truncated normal distribution at the quantiles
-        Phi_q = Phi_a + value * (Phi_b - Phi_a)
-
-        # Invert the cdf of the standard normal distribution at the quantiles
-        z = at.erfinv(2 * Phi_q - 1) * np.sqrt(2)
-
-        # Transform the samples back to the truncated normal distribution
-        return mu + sigma * at.clip(z, lower_std + eps, upper_std - eps)
-
 
 @_default_transform.register(TruncatedNormal)
 def truncated_normal_default_transform(op, rv):

pymc/testing.py

@@ -526,6 +526,7 @@ def check_icdf(
     pymc_dist: Distribution,
     paramdomains: Dict[str, Domain],
     scipy_icdf: Callable,
+    skip_paramdomain_outside_edge_test=False,
     decimal: Optional[int] = None,
     n_samples: int = 100,
 ) -> None:
@@ -548,7 +549,7 @@ def check_icdf(
     paramdomains : Dictionary of Parameter : Domain pairs
         Supported domains of distribution parameters
     scipy_icdf : Scipy icdf method
-        Scipy icdf (ppp) method of equivalent pymc_dist distribution
+        Scipy icdf (ppf) method of equivalent pymc_dist distribution
     decimal : int, optional
         Level of precision with which pymc_dist and scipy_icdf are compared.
         Defaults to 6 for float64 and 3 for float32
@@ -557,6 +558,9 @@ def check_icdf(
         are compared between pymc and scipy methods. If n_samples is below the
         total number of combinations, a random subset is evaluated. Setting
         n_samples = -1, will return all possible combinations. Defaults to 100
+    skip_paradomain_outside_edge_test : Bool
+        Whether to run test 2., which checks that pymc distribution icdf
+        returns nan for invalid parameter values outside the supported domain edge
 
     """
     if decimal is None:
@@ -586,19 +590,20 @@ def check_icdf(
     valid_params = {param: paramdomain.vals[0] for param, paramdomain in paramdomains.items()}
     valid_params["q"] = valid_value
 
-    # Test pymc distribution raises ParameterValueError for parameters outside the
-    # supported domain edges (excluding edges)
-    invalid_params = find_invalid_scalar_params(paramdomains)
-    for invalid_param, invalid_edges in invalid_params.items():
-        for invalid_edge in invalid_edges:
-            if invalid_edge is None:
-                continue
+    if not skip_paramdomain_outside_edge_test:
+        # Test pymc distribution raises ParameterValueError for parameters outside the
+        # supported domain edges (excluding edges)
+        invalid_params = find_invalid_scalar_params(paramdomains)
+        for invalid_param, invalid_edges in invalid_params.items():
+            for invalid_edge in invalid_edges:
+                if invalid_edge is None:
+                    continue
 
-            point = valid_params.copy()
-            point[invalid_param] = invalid_edge
-            with pytest.raises(ParameterValueError):
-                pymc_icdf(**point)
-                pytest.fail(f"test_params={point}")
+                point = valid_params.copy()
+                point[invalid_param] = invalid_edge
+                with pytest.raises(ParameterValueError):
+                    pymc_icdf(**point)
+                    pytest.fail(f"test_params={point}")
 
     # Test that values below 0 or above 1 evaluate to nan
     invalid_values = find_invalid_scalar_params({"q": domain})["q"]

tests/distributions/test_continuous.py

@@ -28,7 +28,7 @@
 
 from pymc.distributions.continuous import Normal, Uniform, get_tau_sigma, interpolated
 from pymc.distributions.dist_math import clipped_beta_rvs
-from pymc.logprob.abstract import logcdf
+from pymc.logprob.abstract import icdf, logcdf
 from pymc.logprob.joint_logprob import logp
 from pymc.logprob.utils import ParameterValueError
 from pymc.pytensorf import floatX
@@ -176,13 +176,21 @@ def test_uniform(self):
             lambda value, lower, upper: st.uniform.logcdf(value, lower, upper - lower),
             skip_paramdomain_outside_edge_test=True,
         )
+        check_icdf(
+            pm.Uniform,
+            {"lower": -Rplusunif, "upper": Rplusunif},
+            lambda q, lower, upper: st.uniform.ppf(q=q, loc=lower, scale=upper - lower),
+            skip_paramdomain_outside_edge_test=True,
+        )
         # Custom logp / logcdf check for invalid parameters
         invalid_dist = pm.Uniform.dist(lower=1, upper=0)
         with pytensor.config.change_flags(mode=Mode("py")):
             with pytest.raises(ParameterValueError):
                 logp(invalid_dist, np.array(0.5)).eval()
             with pytest.raises(ParameterValueError):
                 logcdf(invalid_dist, np.array(0.5)).eval()
+            with pytest.raises(ParameterValueError):
+                icdf(invalid_dist, np.array(0.5)).eval()
 
     def test_triangular(self):
         check_logp(
@@ -2277,41 +2285,3 @@ def dist(cls, **kwargs):
                     extra_args={"rng": pytensor.shared(rng)},
                     ref_rand=ref_rand,
                 )
-
-
-class TestICDF:
-    @pytest.mark.parametrize(
-        "dist_params, obs, size",
-        [
-            ((0, 1), np.array([-0.5, 0, 0.3, 0.5, 1, 1.5], dtype=np.float64), ()),
-            ((-1, 20), np.array([-0.5, 0, 0.3, 0.5, 1, 1.5], dtype=np.float64), ()),
-            ((-1, 20), np.array([-0.5, 0, 0.3, 0.5, 1, 1.5], dtype=np.float64), (2, 3)),
-        ],
-    )
-    def test_normal_icdf(self, dist_params, obs, size):
-        dist_params_at, obs_at, size_at = create_pytensor_params(dist_params, obs, size)
-        dist_params = dict(zip(dist_params_at, dist_params))
-
-        x = Normal.dist(*dist_params_at, size=size_at)
-
-        scipy_logprob_tester(x, obs, dist_params, test_fn=st.norm.ppf, test="icdf")
-
-    @pytest.mark.parametrize(
-        "dist_params, obs, size",
-        [
-            ((-5, 4), np.array([0.0, 0.2, 0.4, 0.6, 0.8, 1.0], dtype=np.float64), ()),
-            ((-1, 2), np.array([0.0, 0.2, 0.4, 0.6, 0.8, 1.0], dtype=np.float64), ()),
-            ((0, 10), np.array([0.0, 0.2, 0.4, 0.6, 0.8, 1.0], dtype=np.float64), ()),
-        ],
-    )
-    def test_uniform_icdf(self, dist_params, obs, size):
-        dist_params_at, obs_at, size_at = create_pytensor_params(dist_params, obs, size)
-        dist_params = dict(zip(dist_params_at, dist_params))
-        x = Uniform.dist(*dist_params_at)
-        scipy_logprob_tester(
-            x,
-            obs,
-            dist_params,
-            test_fn=lambda val, lower, upper: st.uniform.ppf(val, lower, upper - lower),
-            test="icdf",
-        )