From 5f92ff67854fba7cec9db36bb80d3cea4ce82b60 Mon Sep 17 00:00:00 2001
From: James Hensman <james.hensman@gmail.com>
Date: Fri, 30 Nov 2012 10:31:02 +0000
Subject: [PATCH] GPLVM demo working

---
 GPy/kern/kern.py    |  6 +++---
 GPy/kern/rbf.py     | 12 +++++++-----
 GPy/kern/white.py   |  2 +-
 GPy/models/GPLVM.py |  7 ++-----
 4 files changed, 13 insertions(+), 14 deletions(-)

diff --git a/GPy/kern/kern.py b/GPy/kern/kern.py
index cb2ef197..947fb47e 100644
--- a/GPy/kern/kern.py
+++ b/GPy/kern/kern.py
@@ -170,12 +170,12 @@ class kern(parameterised):
         [p.dK_dtheta(partial,X[s1,i_s],X2[s2,i_s],target[ps]) for p,i_s,ps,s1,s2 in zip(self.parts, self.input_slices, self.param_slices, slices1, slices2)]
         return target
 
-    def dK_dX(self,X,X2=None,slices1=None,slices2=None):
+    def dK_dX(self,partial,X,X2=None,slices1=None,slices2=None):
         if X2 is None:
             X2 = X
         slices1, slices2 = self._process_slices(slices1,slices2)
-        target = np.zeros((X2.shape[0],X.shape[0],X.shape[1]))
-        [p.dK_dX(X[s1],X2[s2],target[s2,s1,:]) for p,ps,s1,s2 in zip(self.parts, self.param_slices,slices1,slices2)]
+        target = np.zeros_like(X)
+        [p.dK_dX(partial,X[s1],X2[s2],target[s1,:]) for p,ps,s1,s2 in zip(self.parts, self.param_slices,slices1,slices2)]
         return target
 
     def Kdiag(self,X,slices=None):
diff --git a/GPy/kern/rbf.py b/GPy/kern/rbf.py
index 1a8dedd1..087285d1 100644
--- a/GPy/kern/rbf.py
+++ b/GPy/kern/rbf.py
@@ -6,7 +6,7 @@ from kernpart import kernpart
 import numpy as np
 import hashlib
 
-class rbf(kernpart): 
+class rbf(kernpart):
     """
     Radial Basis Function kernel, aka squared-exponential or Gaussian kernel.
 
@@ -55,13 +55,15 @@ class rbf(kernpart):
         target[0] += np.sum(self._K_dvar*partial)
         target[1] += np.sum(self._K_dvar*self.variance*self._K_dist2/self.lengthscale*partial)
 
-    def dKdiag_dtheta(self,X,target):
-        target[0] += partial
+    def dKdiag_dtheta(self,partial,X,target):
+        #NB: derivative of diagonal elements wrt lengthscale is 0
+        target[0] += np.sum(partial)
 
-    def dK_dX(self,X,X2,target):
+    def dK_dX(self,partial,X,X2,target):
         self._K_computations(X,X2)
         _K_dist = X[:,None,:]-X2[None,:,:]
-        target += np.transpose(-self.variance*self._K_dvar[:,:,np.newaxis]*_K_dist/self.lengthscale2,(1,0,2))
+        dK_dX = np.transpose(-self.variance*self._K_dvar[:,:,np.newaxis]*_K_dist/self.lengthscale2,(1,0,2))
+        target += np.sum(dK_dX*partial[:,:,None],1)
 
     def dKdiag_dX(self,X,target):
         pass
diff --git a/GPy/kern/white.py b/GPy/kern/white.py
index 1215e101..beb99416 100644
--- a/GPy/kern/white.py
+++ b/GPy/kern/white.py
@@ -45,7 +45,7 @@ class white(kernpart):
     def dKdiag_dtheta(self,X,target):
         np.add(target[:,0],1.,target[:,0])
 
-    def dK_dX(self,X,X2,target):
+    def dK_dX(self,partial,X,X2,target):
         pass
 
     def dKdiag_dX(self,X,target):
diff --git a/GPy/models/GPLVM.py b/GPy/models/GPLVM.py
index db678d84..8f67fe3f 100644
--- a/GPy/models/GPLVM.py
+++ b/GPy/models/GPLVM.py
@@ -46,11 +46,8 @@ class GPLVM(GP_regression):
     def log_likelihood_gradients(self):
         dL_dK = self.dL_dK()
 
-        dK_dtheta = self.kern.dK_dtheta(self.X)
-        dL_dtheta = (dK_dtheta*dL_dK[:,:,None]).sum(0).sum(0)
-
-        dK_dX = self.kern.dK_dX(self.X)
-        dL_dX = 2.*np.sum(dL_dK[:,:,None]*dK_dX,0)
+        dL_dtheta = self.kern.dK_dtheta(dL_dK,self.X)
+        dL_dX = 2*self.kern.dK_dX(dL_dK,self.X)
 
         return np.hstack((dL_dX.flatten(),dL_dtheta))