noddling

GPy/inference/latent_function_inference/posterior.py

@@ -25,6 +25,8 @@ class Posterior(object):
         Not all of the above need to be supplied! You *must* supply:
 
           K (for lazy computation)
+          or
+          K_chol (for lazy computation)
 
        You may supply either:
 
@@ -35,11 +37,10 @@ class Posterior(object):
 
           mean
           cov
-          K_chol (for lazy computation)
 
-        Of course, you can supply more than that, but this class will lazily compute all other quantites on demand.
+        Of course, you can supply more than that, but this class will lazily
+        compute all other quantites on demand.
 
-        From the supplied quantities, all of the others will be computed on demand (lazy computation)
         """
         #obligatory
         self._K = K
@@ -49,6 +50,8 @@ class Posterior(object):
         else:
             raise ValueError, "insufficient information to compute the posterior"
 
+        self._K_chol = K_chol
+        self._K = K
         #option 1:
         self._woodbury_chol = woodbury_chol
         self._woodbury_vector = woodbury_vector
@@ -56,7 +59,6 @@ class Posterior(object):
         #option 2:
         self._mean = mean
         self._covariance = cov
-        self._K_chol = K_chol
 
         #compute this lazily
         self._precision = None
@@ -95,6 +97,12 @@ class Posterior(object):
             self._woodbury_vector, _ = dpotrs(self._K_chol, self.mean)
         return self._woodbury_vector
 
+    @property
+    def K_chol(self):
+        if self._K_chol is None:
+            self._K_chol = dportf(self._K)
+        return self._K_chol
+
 
 
 

GPy/inference/latent_function_inference/varDTC.py

@@ -33,6 +33,9 @@ class VarDTC(object):
             #if Y in self.cache, return self.Cache[Y], else stor Y in cache and return L.
             raise NotImplementedError, 'TODO' #TODO
 
+    def get_VVTfactor(self, Y, prec):
+        return Y * prec # TODO chache this, and make it effective
+
     def inference(self, kern, X, X_variance, Z, likelihood, Y):
 
         num_inducing, _ = Z.shape
@@ -79,12 +82,12 @@ class VarDTC(object):
         LB = jitchol(B)
 
         # VVT_factor is a matrix such that tdot(VVT_factor) = VVT...this is for efficiency!
-        psi1Vf = np.dot(psi1.T, likelihood.VVT_factor)
+        VVT_factor = self.get_VVTfactor(Y, likelihood.precision)
+        psi1Vf = np.dot(psi1.T, VVT_factor)
 
         # back substutue C into psi1Vf
         tmp, info1 = dtrtrs(Lm, np.asfortranarray(psi1Vf), lower=1, trans=0)
         _LBi_Lmi_psi1Vf, _ = dtrtrs(LB, np.asfortranarray(tmp), lower=1, trans=0)
-        # tmp, info2 = dpotrs(LB, tmp, lower=1)
         tmp, info2 = dtrtrs(LB, _LBi_Lmi_psi1Vf, lower=1, trans=1)
         Cpsi1Vf, info3 = dtrtrs(Lm, tmp, lower=1, trans=1)
 
@@ -163,3 +166,14 @@ class VarDTC(object):
             partial_for_likelihood += likelihood.precision * (0.5 * np.sum(_A * DBi_plus_BiPBi) - data_fit)
 
 
+        #get sufficient things for posterior prediction
+        if VVT_factor.shape[1] == Y.shape[1]:
+            Cpsi1V = Cpsi1Vf
+        else:
+            raise NotImplementedError #TODO
+
+        #construct a posterior object
+        post = Posterior(woodbury_chol=None, woodbury_vector=Cpsi1V, K=None, mean=None, cov=None, K_chol=None):
+        return
+
+