cmu_mocap() example mostly working except some fiddling with axes for visualization. Also changes to naming of scaling and offset parameters in GP.py and deal with the case where the scale parameter is zero.

GPy/likelihoods/Gaussian.py

@@ -2,19 +2,30 @@ import numpy as np
 from likelihood import likelihood
 
 class Gaussian(likelihood):
+    """
+    Likelihood class for doing Expectation propagation
+
+    :param Y: observed output (Nx1 numpy.darray)
+    ..Note:: Y values allowed depend on the likelihood_function used
+    :param variance : 
+    :param normalize:  whether to normalize the data before computing (predictions will be in original scales)
+    :type normalize: False|True
+    """
     def __init__(self,data,variance=1.,normalize=False):
         self.is_heteroscedastic = False
         self.Nparams = 1
         self.Z = 0. # a correction factor which accounts for the approximation made
         N, self.D = data.shape
 
-        #normaliztion
+        #normalization
         if normalize:
-            self._mean = data.mean(0)[None,:]
-            self._std = data.std(0)[None,:]
+            self._bias = data.mean(0)[None,:]
+            self._scale = data.std(0)[None,:]
+            # Don't scale outputs which have zero variance to zero. 
+            self._scale[np.nonzero(self._scale==0.)] = 1.0e-3
         else:
-            self._mean = np.zeros((1,self.D))
-            self._std = np.ones((1,self.D))
+            self._bias = np.zeros((1,self.D))
+            self._scale = np.ones((1,self.D))
 
         self.set_data(data)
 
@@ -24,7 +35,7 @@ class Gaussian(likelihood):
         self.data = data
         self.N,D = data.shape
         assert D == self.D
-        self.Y = (self.data - self._mean)/self._std
+        self.Y = (self.data - self._bias)/self._scale
         if D > self.N:
             self.YYT = np.dot(self.Y,self.Y.T)
             self.trYYT = np.trace(self.YYT)
@@ -47,19 +58,19 @@ class Gaussian(likelihood):
         """
         Un-normalize the prediction and add the likelihood variance, then return the 5%, 95% interval
         """
-        mean = mu*self._std + self._mean
+        mean = mu*self._scale + self._bias
         if full_cov:
             if self.D >1:
                 raise NotImplementedError, "TODO"
                 #Note. for D>1, we need to re-normalise all the outputs independently.
                 # This will mess up computations of diag(true_var), below.
                 #note that the upper, lower quantiles should be the same shape as mean
-            true_var = (var + np.eye(var.shape[0])*self._variance)*self._std**2
-            _5pc = mean + - 2.*np.sqrt(np.diag(true_var))
+            true_var = (var + np.eye(var.shape[0])*self._variance)*self._scale**2
+            _5pc = mean - 2.*np.sqrt(np.diag(true_var))
             _95pc = mean + 2.*np.sqrt(np.diag(true_var))
         else:
-            true_var = (var + self._variance)*self._std**2
-            _5pc = mean + - 2.*np.sqrt(true_var)
+            true_var = (var + self._variance)*self._scale**2
+            _5pc = mean - 2.*np.sqrt(true_var)
             _95pc = mean + 2.*np.sqrt(true_var)
         return mean, true_var, _5pc, _95pc