Changed kern.py so that X2 is correctly passed as None to the kern parts for dK_dX. Modified several kern parts so that dK_dX is correctly computed (factors of 2 missing). Removed spurious factors of 2 from gplvm, bcgplvm, sparse_gp and fitc code.

GPy/core/fitc.py

@@ -174,7 +174,7 @@ class FITC(SparseGP):
 
     def dL_dZ(self):
         dL_dZ = self.kern.dK_dX(self._dL_dpsi1.T,self.Z,self.X)
-        dL_dZ += 2. * self.kern.dK_dX(self._dL_dKmm,X=self.Z)
+        dL_dZ += self.kern.dK_dX(self._dL_dKmm,X=self.Z)
         dL_dZ += self._dpsi1_dX
         dL_dZ += self._dKmm_dX
         return dL_dZ

GPy/core/model.py

@@ -1,4 +1,4 @@
-# Copyright (c) 2012, GPy authors (see AUTHORS.txt).
+# Copyright (c) 2012, 2013, GPy authors (see AUTHORS.txt).
 # Licensed under the BSD 3-clause license (see LICENSE.txt)
 
 
@@ -456,8 +456,8 @@ class Model(Parameterized):
             gradient = self.objective_function_gradients(x)
 
             numerical_gradient = (f1 - f2) / (2 * dx)
-            global_ratio = (f1 - f2) / (2 * np.dot(dx, gradient))
-
+            global_ratio = (f1 - f2) / (2 * np.dot(dx, np.where(gradient==0, 1e-32, gradient)))
+            
             return (np.abs(1. - global_ratio) < tolerance) or (np.abs(gradient - numerical_gradient).mean() - 1) < tolerance
         else:
             # check the gradient of each parameter individually, and do some pretty printing
@@ -496,7 +496,7 @@ class Model(Parameterized):
                 gradient = self.objective_function_gradients(x)[i]
 
                 numerical_gradient = (f1 - f2) / (2 * step)
-                ratio = (f1 - f2) / (2 * step * gradient)
+                ratio = (f1 - f2) / (2 * step * np.where(gradient==0, 1e-312, gradient))
                 difference = np.abs((f1 - f2) / 2 / step - gradient)
 
                 if (np.abs(1. - ratio) < tolerance) or np.abs(difference) < tolerance:

GPy/core/sparse_gp.py

@@ -240,7 +240,7 @@ class SparseGP(GPBase):
         """
         The derivative of the bound wrt the inducing inputs Z
         """
-        dL_dZ = 2.*self.kern.dK_dX(self.dL_dKmm, self.Z) # factor of two becase of vertical and horizontal 'stripes' in dKmm_dZ
+        dL_dZ = self.kern.dK_dX(self.dL_dKmm, self.Z) 
         if self.has_uncertain_inputs:
             dL_dZ += self.kern.dpsi1_dZ(self.dL_dpsi1, self.Z, self.X, self.X_variance)
             dL_dZ += self.kern.dpsi2_dZ(self.dL_dpsi2, self.Z, self.X, self.X_variance)

GPy/core/transformations.py

@@ -18,9 +18,11 @@ class transformation(object):
     def gradfactor(self, f):
         """ df_dx evaluated at self.f(x)=f"""
         raise NotImplementedError
+
     def initialize(self, f):
         """ produce a sensible initial value for f(x)"""
         raise NotImplementedError
+
     def __str__(self):
         raise NotImplementedError
 
@@ -42,15 +44,13 @@ class logexp(transformation):
 class negative_logexp(transformation):
     domain = NEGATIVE
     def f(self, x):
-        return -logexp.f(x) #np.log(1. + np.exp(x))
+        return -logexp.f(x)
     def finv(self, f):
-        return logexp.finv(-f) #np.log(np.exp(-f) - 1.)
+        return logexp.finv(-f) 
     def gradfactor(self, f):
         return -logexp.gradfactor(-f)
-        #ef = np.exp(-f)
-        #return -(ef - 1.) / ef
     def initialize(self, f):
-        return -logexp.initialize(f) #np.abs(f)
+        return -logexp.initialize(f)
     def __str__(self):
         return '(-ve)'
 
@@ -82,7 +82,6 @@ class logexp_clipped(logexp):
         return '(+ve_c)'
 
 class exponent(transformation):
-    # TODO: can't allow this to go to zero, need to set a lower bound. Similar with negative exponent below. See old MATLAB code.
     domain = POSITIVE
     def f(self, x):
         return np.where(x<lim_val, np.where(x>-lim_val, np.exp(x), np.exp(-lim_val)), np.exp(lim_val))