Merge branch 'devel' of github.com:SheffieldML/GPy into ties

GPy/core/model.py

@@ -287,7 +287,7 @@ class Model(Parameterized):
             # just check the global ratio
             dx = np.zeros(x.shape)
             dx[transformed_index] = step * (np.sign(np.random.uniform(-1, 1, transformed_index.size)) if transformed_index.size != 2 else 1.)
-
+            
             # evaulate around the point x
             f1 = self._objective(x + dx)
             f2 = self._objective(x - dx)

GPy/core/parameterization/parameter_core.py

@@ -13,7 +13,7 @@ Observable Pattern for patameterization
 
 """
 
-from transformations import Logexp, NegativeLogexp, Logistic, __fixed__, FIXED, UNFIXED
+from transformations import Transformation,Logexp, NegativeLogexp, Logistic, __fixed__, FIXED, UNFIXED
 import numpy as np
 import re
 import logging
@@ -573,7 +573,8 @@ class Indexable(Nameable, Observable):
         Constrain the parameter to the given
         :py:class:`GPy.core.transformations.Transformation`.
         """
-        self.param_array[...] = transform.initialize(self.param_array)
+        if isinstance(transform, Transformation):
+            self.param_array[...] = transform.initialize(self.param_array)
         reconstrained = self.unconstrain()
         added = self._add_to_index_operations(self.constraints, reconstrained, transform, warning)
         self.notify_observers(self, None if trigger_parent else -np.inf)
@@ -649,7 +650,7 @@ class Indexable(Nameable, Observable):
         """
         Helper preventing copy code.
         This adds the given what (transformation, prior etc) to parameter index operations which.
-        revonstrained are reconstrained indices.
+        reconstrained are reconstrained indices.
         warn when reconstraining parameters if warning is True.
         TODO: find out which parameters have changed specifically
         """

GPy/core/parameterization/variational.py

@@ -7,7 +7,7 @@ Created on 6 Nov 2013
 import numpy as np
 from parameterized import Parameterized
 from param import Param
-from transformations import Logexp, Logistic
+from transformations import Logexp, Logistic,__fixed__
 
 class VariationalPrior(Parameterized):
     def __init__(self, name='latent space', **kw):
@@ -35,12 +35,15 @@ class NormalPrior(VariationalPrior):
 
 class SpikeAndSlabPrior(VariationalPrior):
     def __init__(self, pi=None, learnPi=False, variance = 1.0, name='SpikeAndSlabPrior', **kw):
-        super(VariationalPrior, self).__init__(name=name, **kw)
-        self.pi = Param('pi', pi, Logistic(1e-10,1.-1e-10))
+        super(SpikeAndSlabPrior, self).__init__(name=name, **kw)        
         self.variance = Param('variance',variance)
         self.learnPi = learnPi
         if learnPi:
-            self.add_parameters(self.pi)
+            self.pi = Param('Pi', pi, Logistic(1e-10,1.-1e-10))
+        else:
+            self.pi = Param('Pi', pi, __fixed__)
+        self.add_parameter(self.pi)
+
 
     def KL_divergence(self, variational_posterior):
         mu = variational_posterior.mean

GPy/core/sparse_gp_mpi.py

@@ -46,7 +46,7 @@ class SparseGP_MPI(SparseGP):
         self.add_parameter(self.X, index=0)
         if variational_prior is not None:
             self.add_parameter(variational_prior)
-        self.X.fix()
+#         self.X.fix()
 
         self.mpi_comm = mpi_comm
         # Manage the data (Y) division
@@ -67,6 +67,9 @@ class SparseGP_MPI(SparseGP):
             del dc['N_range']
             del dc['N_list']
             del dc['Y_local']
+        if 'normalizer' not in dc:
+            dc['normalizer'] = None
+            dc['Y_normalized'] = dc['Y']
         return dc
 
     #=====================================================

GPy/inference/latent_function_inference/var_dtc_parallel.py

@@ -170,12 +170,14 @@ class VarDTC_minibatch(LatentFunctionInference):
         
         Kmm = kern.K(Z).copy()
         diag.add(Kmm, self.const_jitter)
-        checkFullRank(Kmm)
+        r1 = checkFullRank(Kmm,name='Kmm')
         Lm = jitchol(Kmm)
         
         LmInvPsi2LmInvT = backsub_both_sides(Lm,psi2_full,transpose='right')
         Lambda = np.eye(Kmm.shape[0])+LmInvPsi2LmInvT
-        checkFullRank(Lambda)
+        r2 = checkFullRank(Lambda,name='Lambda')
+        if (not r1) or (not r2):
+            raise
         LL = jitchol(Lambda)
         LL = np.dot(Lm,LL)
         b,_ = dtrtrs(LL, psi1Y_full.T)
@@ -339,7 +341,13 @@ def update_gradients(model, mpi_comm=None):
         Y = model.Y_local
         X = model.X[model.N_range[0]:model.N_range[1]]
 
-    model._log_marginal_likelihood, dL_dKmm, model.posterior = model.inference_method.inference_likelihood(model.kern, X, model.Z, model.likelihood, Y)
+    try:
+        model._log_marginal_likelihood, dL_dKmm, model.posterior = model.inference_method.inference_likelihood(model.kern, X, model.Z, model.likelihood, Y)
+    except Exception:
+        if model.mpi_comm is None or model.mpi_comm.rank==0:
+            import time
+            model.pickle('model_'+str(int(time.time()))+'.pickle')
+        raise
     
     het_noise = model.likelihood.variance.size > 1
     

GPy/models/ss_gplvm.py

@@ -70,8 +70,8 @@ class SSGPLVM(SparseGP_MPI):
         X = SpikeAndSlabPosterior(X, X_variance, gamma)
                 
         super(SSGPLVM,self).__init__(X, Y, Z, kernel, likelihood, variational_prior=self.variational_prior, inference_method=inference_method, name=name, mpi_comm=mpi_comm, normalizer=normalizer, **kwargs)
-        self.X.unfix()
-        self.X.variance.constrain_positive()
+#         self.X.unfix()
+#         self.X.variance.constrain_positive()
                 
         if self.group_spike:
             [self.X.gamma[:,i].tie('tieGamma'+str(i)) for i in xrange(self.X.gamma.shape[1])] # Tie columns together

GPy/util/debug.py

@@ -26,7 +26,7 @@ def checkFullRank(m, tol=1e-10, name=None, force_check=False):
         print 'The size of '+name+'is too big to check (>=10000)!'
         return True
     
-    s = np.linalg.eigvals(m)
+    s = np.real(np.linalg.eigvals(m))
     
     if s.min()/s.max()<tol:
         print name+' is close to singlar!'