improve mle for NegativeBinomial (#147)

preliz/distributions/discrete.py

@@ -11,7 +11,7 @@
 
 
 from .distributions import Discrete
-from ..utils.optimization import optimize_matching_moments
+from ..utils.optimization import optimize_matching_moments, optimize_ml
 
 _log = logging.getLogger("preliz")
 
@@ -298,13 +298,7 @@ def _fit_moments(self, mean, sigma):
         self._update(mu, alpha)
 
     def _fit_mle(self, sample):
-        # the upper bound is based on a quick heuristic. The fit will underestimate
-        # the value of n when p is very close to 1.
-        fitted = stats.fit(self.dist, sample, bounds={"n": (1, max(sample) * 2)})
-        if not fitted.success:
-            _log.info("Optimization did not terminate successfully.")
-        mu, alpha = self._from_p_n(fitted.params.p, fitted.params.n)  # pylint: disable=no-member
-        self._update(mu, alpha)
+        optimize_ml(self, sample)
 
 
 class Poisson(Discrete):

preliz/tests/test_distributions.py

@@ -93,7 +93,7 @@ def test_moments(distribution, params):
         (Gumbel, (0, 1)),
         (HalfCauchy, (1,)),
         (HalfNormal, (1,)),
-        (HalfStudent, (10, 1)),
+        (HalfStudent, (100, 1)),
         (InverseGamma, (3, 0.5)),
         (Laplace, (0, 1)),
         (Logistic, (0, 1)),
@@ -127,7 +127,7 @@ def test_mle(distribution, params):
 
     assert_almost_equal(dist.rv_frozen.mean(), dist_.rv_frozen.mean(), 1)
     assert_almost_equal(dist.rv_frozen.std(), dist_.rv_frozen.std(), 1)
-    if dist.name == "student":
+    if dist.name in "student":
         assert_almost_equal(params[1:], dist_.params[1:], 0)
     else:
         assert_almost_equal(params, dist_.params, 0)

preliz/utils/optimization.py

@@ -102,10 +102,14 @@ def func(params, dist, mean, sigma):
 def optimize_ml(dist, sample):
     def negll(params, dist, sample):
         dist._update(*params)
-        return -dist.rv_frozen.logpdf(sample).sum()
+        if dist.kind == "continuous":
+            neg = -dist.rv_frozen.logpdf(sample).sum()
+        else:
+            neg = -dist.rv_frozen.logpmf(sample).sum()
+        return neg
 
-    dist._fit_moments(0, np.std(sample))
-    init_vals = dist.params[::-1]
+    dist._fit_moments(np.mean(sample), np.std(sample))
+    init_vals = dist.params
 
     opt = minimize(negll, x0=init_vals, bounds=dist.params_support, args=(dist, sample))