Allow DiscreteUniform distribution to be used with CategoricalGibbsMetropolis step method.

RELEASE-NOTES.md

@@ -10,6 +10,7 @@
 - `GLM.from_formula` and `LinearComponent.from_formula` can extract variables from the calling scope. Customizable via the new `eval_env` argument. Fixing #3382.
 - Added the `distributions.shape_utils` module with functions used to help broadcast samples drawn from distributions using the `size` keyword argument.
 - Used `numpy.vectorize` in `distributions.distribution._compile_theano_function`. This enables `sample_prior_predictive` and `sample_posterior_predictive` to ask for tuples of samples instead of just integers. This fixes issue #3422.
+- Allow DiscreteUniform distribution to be used with CategoricalGibbsMetropolis step method.
 
 ### Maintenance
 - Fixed an issue in `model_graph` that caused construction of the graph of the model for rendering to hang: replaced a search over the powerset of the nodes with a breadth-first search over the nodes. Fix for #3458.

pymc3/step_methods/metropolis.py

@@ -400,9 +400,15 @@ def __init__(self, vars, proposal='uniform', order='random', model=None):
                 k = draw_values([distr.k])[0]
             elif isinstance(distr, pm.Bernoulli) or (v.dtype in pm.bool_types):
                 k = 2
+            elif isinstance(distr, pm.DiscreteUniform):
+                k = draw_values([distr.upper])[0] + 1
+                if np.all(draw_values([distr.lower])):
+                    raise ValueError('Parameter lower must be 0 to use DiscreteUniform' +
+                                     'with CategoricalGibbsMetropolis.')
+
             else:
                 raise ValueError('All variables must be categorical or binary' +
-                                 'for CategoricalGibbsMetropolis')
+                                 'or DiscreteUniform for CategoricalGibbsMetropolis')
             start = len(dimcats)
             dimcats += [(dim, k) for dim in range(start, start + v.dsize)]
 
@@ -489,6 +495,8 @@ def competence(var):
             return Competence.COMPATIBLE
         elif isinstance(distribution, pm.Bernoulli) or (var.dtype in pm.bool_types):
             return Competence.COMPATIBLE
+        elif isinstance(distribution, pm.DiscreteUniform) and draw_values([distribution.lower])[0] == 0:
+            return Competence.COMPATIBLE
         return Competence.INCOMPATIBLE
 
 

pymc3/tests/test_step.py

@@ -36,6 +36,8 @@
     Categorical,
     Beta,
     HalfNormal,
+    DiscreteUniform,
+    Poisson,
 )
 
 from numpy.testing import assert_array_almost_equal
@@ -1047,3 +1049,27 @@ def test_sampler_stats(self):
             ]
         )
         assert (trace.model_logp == model_logp_).all()
+
+
+class TestDiscreteUniform:
+
+    def test_DiscreteUniform(self, lower=0, upper=200000000,
+                             obs=5000000, draws=20000):
+        """Test that DiscreteUniform distribution can use
+         CategoricalGibbsMetropolis step method."""
+        obs = theano.shared(obs)
+        with Model() as model2:
+
+            x = DiscreteUniform('x', lower, upper - 1)
+            sfs_obs = Poisson('sfs_obs', mu=x, observed=obs)
+
+        with model2:
+
+            step = CategoricalGibbsMetropolis([x])
+            trace = sample(draws, tune=0, step=step)
+        return trace
+
+    def test_bad_lower(self):
+        with pytest.raises(ValueError):
+            self.test_DiscreteUniform(lower=1, upper=200000001,
+                                      obs=5000000, draws=20000)