swap pareto alpha and scale (#153)

examples/baseball.py

@@ -108,7 +108,7 @@ def partially_pooled(at_bats, hits=None):
     """
     num_players = at_bats.shape[0]
     m = sample("m", dist.Uniform(np.array([0.]), np.array([1.])))
-    kappa = sample("kappa", dist.Pareto(np.array([1.]), np.array([1.5])))
+    kappa = sample("kappa", dist.Pareto(np.array([1.5])))
     shape = np.shape(kappa)[:np.ndim(kappa) - 1] + (num_players,)
     phi_prior = dist.Beta(np.broadcast_to(m * kappa, shape),
                           np.broadcast_to((1 - m) * kappa, shape))

numpyro/distributions/continuous.py

@@ -503,9 +503,7 @@ class Pareto(TransformedDistribution):
     arg_constraints = {'alpha': constraints.positive, 'scale': constraints.positive}
     support = constraints.real
 
-    # FIXME: should we use `concentration` and `scale` for consistence with other distributions
-    # tensorflow uses `concentration` and `scale` though
-    def __init__(self, scale, alpha, validate_args=None):
+    def __init__(self, alpha, scale=1., validate_args=None):
         batch_shape = lax.broadcast_shapes(np.shape(scale), np.shape(alpha))
         self.scale, self.alpha = np.broadcast_to(scale, batch_shape), np.broadcast_to(alpha, batch_shape)
         base_dist = Exponential(self.alpha)

test/test_distributions.py

@@ -53,7 +53,7 @@ def __new__(cls, jax_dist, *params):
     dist.MultinomialLogits: lambda logits, total_count: osp.multinomial(n=total_count,
                                                                         p=_to_probs_multinom(logits)),
     dist.Normal: lambda loc, scale: osp.norm(loc=loc, scale=scale),
-    dist.Pareto: lambda scale, alpha: osp.pareto(alpha, scale=scale),
+    dist.Pareto: lambda alpha, scale: osp.pareto(alpha, scale=scale),
     dist.Poisson: lambda rate: osp.poisson(rate),
     dist.StudentT: lambda df, loc, scale: osp.t(df=df, loc=loc, scale=scale),
     dist.Uniform: lambda a, b: osp.uniform(a, b - a),