Added shape inferring in LKJCholeskyCov

pymc3/distributions/multivariate.py

@@ -1040,6 +1040,11 @@ def _lkj_normalizing_constant(eta, n):
 
 
 class _LKJCholeskyCovRV(RandomVariable):
+    def _shape_from_params(self, dist_params, **kwargs):
+        n = dist_params[1]
+        dist_shape = ((n * (n + 1)) // 2,)
+        return dist_shape
+
     def __init__(self, *args, sd_dist=None, **kwargs):
         self.sd_dist = sd_dist
         self._print_name = _print_name = ("_LKJCholeskyCov", "\\operatorname{_LKJCholeskyCov}")
@@ -1152,7 +1157,7 @@ def dist(cls, eta, n, sd_dist, *args, **kwargs):
 
         cls.rv_op = _LKJCholeskyCovRV(
             "_lkjcholeskycov",
-            0,
+            2,
             (0, 0, 0),
             "floatX",
             sd_dist=sd_dist,

pymc3/tests/test_distributions.py

@@ -3140,14 +3140,13 @@ def test_lkjcholeskycov():
     dist_shape = ((D * (D + 1)) // 2,)
 
     with pm.Model() as model:
-        sd_dist = pm.HalfCauchy.dist(beta=2.5, size=(D))
+        sd_dist = pm.HalfCauchy.dist(beta=2.5)
         packedL = pm.LKJCholeskyCov("packedL", eta=2, n=D, sd_dist=sd_dist)
 
+    with model:
+        prior = pm.sample()
+
     pt = np.random.random(120)
     pt = {"packedL_cholesky-cov-packed__": pt}
     logp = model.fastlogp(pt)
-
-    with model:
-        prior = pm.sample_prior_predictive(5)
-
     assert prior["packedL"].shape == (samples,) + dist_shape