fix sparse EP cavity - energy still not working

AaltoML · Jun 18, 2021 · 1d5b2dd · 1d5b2dd
1 parent a65f15a
commit 1d5b2dd
Show file tree

Hide file tree

Showing 2 changed files with 22 additions and 5 deletions.
diff --git a/newt/basemodels.py b/newt/basemodels.py
@@ -382,7 +382,7 @@ def compute_global_pseudo_lik(self):
         return pseudo_y_full, pseudo_var_full
 
     def compute_full_pseudo_lik(self):
-        nat1lik_full, nat2lik_full = vmap(self.compute_full_pseudo_nat)(self.obs_ind)
+        nat1lik_full, nat2lik_full = self.compute_full_pseudo_nat(self.obs_ind)  # TODO: remove obs_ind
         pseudo_var_full = inv_vmap(nat2lik_full + 1e-12 * np.eye(nat2lik_full.shape[1]))
         pseudo_y_full = pseudo_var_full @ nat1lik_full
         return pseudo_y_full, pseudo_var_full
@@ -391,8 +391,8 @@ def compute_full_pseudo_nat(self, batch_ind):
         Kuf = self.kernel(self.Z.value, self.X[batch_ind].reshape(-1, 1))  # only compute log lik for observed values
         Kuu = self.kernel(self.Z.value, self.Z.value)
         Wuf = solve(Kuu, Kuf)  # conditional mapping, Kuu^-1 Kuf
-        nat1lik_full = Wuf @ self.pseudo_likelihood.nat1[batch_ind].reshape(-1, 1)
-        nat2lik_full = Wuf @ np.diag(self.pseudo_likelihood.nat2[batch_ind].reshape(-1)) @ transpose(Wuf)
+        nat1lik_full = Wuf.T[..., None] @ self.pseudo_likelihood.nat1[batch_ind]
+        nat2lik_full = Wuf.T[..., None] @ self.pseudo_likelihood.nat2[batch_ind] @ Wuf.T[:, None]
         return nat1lik_full, nat2lik_full
 
     def compute_kl(self):
@@ -478,6 +478,23 @@ def conditional_posterior_to_data(self, batch_ind=None, post_mean=None, post_cov
                                                                                          self.Z.value)
             return mean_f.reshape(Nbatch, 1, 1), cov_f.reshape(Nbatch, 1, 1)
 
+    def cavity_distribution(self, batch_ind=None, power=1.):
+        """ Compute the power EP cavity for the given data points """
+        if batch_ind is None:
+            batch_ind = np.arange(self.num_data)
+
+        nat1lik_full, nat2lik_full = self.compute_full_pseudo_nat(batch_ind)
+
+        # then compute the cavity
+        cavity_mean, cavity_cov = vmap(compute_cavity, [None, None, 0, 0, None])(
+            self.posterior_mean.value[..., 0],
+            self.posterior_covariance.value,
+            nat1lik_full,
+            nat2lik_full,
+            power
+        )
+        return cavity_mean, cavity_cov
+
 
 class MarkovGP(BaseModel):
     """

diff --git a/newt/inference.py b/newt/inference.py
@@ -276,7 +276,7 @@ def energy(self, batch_ind=None, cubature=None, power=1.):
         """
         if batch_ind is None:
             batch_ind = np.arange(self.num_data)
-            scale = 1
+            scale = 1.
         else:
             scale = self.num_data / batch_ind.shape[0]
 
@@ -318,7 +318,7 @@ def energy(self, batch_ind=None, cubature=None, power=1.):
 
         ep_energy = -(
             lZ_post
-            + 1 / power * (scale * np.nansum(lZ) - np.nansum(lZ_pseudo))
+            + 1. / power * (scale * np.nansum(lZ) - np.nansum(lZ_pseudo))
         )
 
         return ep_energy