More comments

Author: nspope

tests/test_approximations.py

@@ -70,7 +70,6 @@ def pdf(t_i, t_j, a_i, b_i, a_j, b_j, y, mu):
             )
 
     def test_sufficient_statistics(self, pars):
-        print((pars[5] - pars[3]) / (pars[1] + pars[5]))
         logconst, t_i, ln_t_i, t_j, ln_t_j = approx.sufficient_statistics(*pars)
         ck_normconst = scipy.integrate.dblquad(
             lambda ti, tj: self.pdf(ti, tj, *pars), 0, np.inf, lambda tj: tj, np.inf
@@ -184,5 +183,4 @@ def test_approximate_gamma(self, k):
             lambda x: conditional_coalescent_pdf(x, self.n, k),
             lambda x: scipy.stats.gamma.logpdf(x, alpha_1, scale=1 / beta_1),
         )
-        print(kl_1, kl_0)
         assert kl_1 < kl_0

tests/test_hypergeo.py

@@ -90,15 +90,13 @@ def test_2f1(self, pars):
         f, s, *_ = hypergeo._hyp2f1_taylor_series(*pars)
         check = self._2f1_validate(*pars)
         assert s == mpmath.sign(check)
-        print(f, mpmath.log(mpmath.fabs(check)))
         assert np.isclose(f, float(mpmath.log(mpmath.fabs(check))))
 
     def test_2f1_grad(self, pars):
         _, _, *grad = hypergeo._hyp2f1_taylor_series(*pars)
         grad = grad[:-1]
         offset = self._2f1_validate(*pars)
         check = self._2f1_grad_validate(*pars, offset=offset)
-        print(grad, check)
         assert np.allclose(grad, check)
 
 
@@ -142,7 +140,6 @@ def test_2f1(self, pars):
         f, s, *_ = hypergeo._hyp2f1_recurrence(*pars)
         check = self._2f1_validate(*pars)
         assert s == mpmath.sign(check)
-        print(f, mpmath.log(mpmath.fabs(check)))
         assert np.isclose(f, float(mpmath.log(mpmath.fabs(check))))
 
     def test_2f1_grad(self, pars):
@@ -152,7 +149,6 @@ def test_2f1_grad(self, pars):
         offset = self._2f1_validate(*pars)
         check = self._2f1_grad_validate(*pars, offset=offset)
         check[1] = 0.0  # integer parameter has no gradient
-        print(grad, check)
         assert np.allclose(grad, check)
 
 
@@ -235,7 +231,6 @@ def test_2f1(self, muts, hyp2f1_func, pars):
         f, s, *_ = hyp2f1_func(*pars)
         assert s > 0
         check = float(mpmath.log(self._2f1_validate(*pars)))
-        print(f, check)
         assert np.isclose(f, check)
 
     def test_2f1_grad(self, muts, hyp2f1_func, pars):
@@ -244,5 +239,4 @@ def test_2f1_grad(self, muts, hyp2f1_func, pars):
         assert s > 0
         offset = self._2f1_validate(*pars)
         check = self._2f1_grad_validate(*pars, offset=offset)
-        print(grad, check)
         assert np.allclose(grad, check)

tsdate/core.py

@@ -1000,10 +1000,11 @@ def propagate(self, *, edges, progress=None):
                 continue
             if edge.parent in self.fixednodes:
                 raise ValueError("Internal nodes can not be fixed in EP algorithm")
+            # Get the edge likelihood (Poisson) in terms of a gamma distribution
             edge_lik = self.lik.to_gamma(edge, natural=True)
             # Get the cavity posteriors: that is, the rest of the approximation
-            # without the factor for this edge. This only involves the variational
-            # parameters for the parent and child on the edge.
+            # without the factor for this edge. This only involves updating the
+            # variational parameters for the parent and child on the edge.
             parent_cavity = self.lik.ratio(
                 self.posterior[edge.parent], self.parent_message[edge.id]
             )
@@ -1019,14 +1020,16 @@ def propagate(self, *, edges, progress=None):
                 self.posterior[edge.child],
             ) = approx.gamma_projection(*parent_cavity, *child_cavity, *edge_lik)
             # Get the messages: that is, the multiplicative difference between
-            # the target and cavity posteriors.
+            # the target and cavity posteriors. This only involves updating the
+            # variational parameters for the parent and child on the edge.
             self.parent_message[edge.id] = self.lik.ratio(
                 self.posterior[edge.parent], parent_cavity
             )
             self.child_message[edge.id] = self.lik.ratio(
                 self.posterior[edge.child], child_cavity
             )
-            # Get the approximation to the marginal likelihood
+            # Get the contribution to the (approximate) marginal likelihood from
+            # the edge.
             # TODO not complete
             self.factor_norm[edge.id] = norm_const