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

GPy/core/gp.py

@@ -31,7 +31,7 @@ class GP(Model):
         super(GP, self).__init__(name)
 
         assert X.ndim == 2
-        if isinstance(X, ObservableArray) or isinstance(X, VariationalPosterior):
+        if isinstance(X, (ObservableArray, VariationalPosterior)):
             self.X = X
         else: self.X = ObservableArray(X)
 
@@ -224,13 +224,9 @@ class GP(Model):
                 self.kern,
                 self.likelihood,
                 self.output_dim,
-                self._Xoffset,
-                self._Xscale,
                 ]
 
     def _setstate(self, state):
-        self._Xscale = state.pop()
-        self._Xoffset = state.pop()
         self.output_dim = state.pop()
         self.likelihood = state.pop()
         self.kern = state.pop()

GPy/core/model.py

@@ -60,20 +60,6 @@ class Model(Parameterized):
         self.priors = state.pop()
         Parameterized._setstate(self, state)
 
-    def randomize(self):
-        """
-        Randomize the model.
-        Make this draw from the prior if one exists, else draw from N(0,1)
-        """
-        # first take care of all parameters (from N(0,1))
-        # x = self._get_params_transformed()
-        x = np.random.randn(self.size_transformed)
-        x = self._untransform_params(x)
-        # now draw from prior where possible
-        [np.put(x, ind, p.rvs(ind.size)) for p, ind in self.priors.iteritems() if not p is None]
-        self._set_params(x)
-        # self._set_params_transformed(self._get_params_transformed()) # makes sure all of the tied parameters get the same init (since there's only one prior object...)
-
     def optimize_restarts(self, num_restarts=10, robust=False, verbose=True, parallel=False, num_processes=None, **kwargs):
         """
         Perform random restarts of the model, and set the model to the best
@@ -240,6 +226,11 @@ class Model(Parameterized):
 
         TODO: valid args
         """
+        if self.is_fixed:
+            raise RuntimeError, "Cannot optimize, when everything is fixed"
+        if self.size == 0:
+            raise RuntimeError, "Model without parameters cannot be minimized"
+        
         if optimizer is None:
             optimizer = self.preferred_optimizer
 
@@ -279,7 +270,7 @@ class Model(Parameterized):
            and numerical gradients is within <tolerance> of unity.
         """
 
-        x = self._get_params_transformed().copy()
+        x = self._get_params_transformed()
 
         if not verbose:
             # make sure only to test the selected parameters
@@ -297,7 +288,7 @@ class Model(Parameterized):
                     return
 
             # just check the global ratio
-            dx = np.zeros_like(x)
+            dx = np.zeros(x.shape)
             dx[transformed_index] = step * np.sign(np.random.uniform(-1, 1, transformed_index.size))
 
             # evaulate around the point x
@@ -308,9 +299,8 @@ class Model(Parameterized):
             dx = dx[transformed_index]
             gradient = gradient[transformed_index]
 
-            numerical_gradient = (f1 - f2) / (2 * dx)
             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() < tolerance)
+            return (np.abs(1. - global_ratio) < tolerance)
         else:
             # check the gradient of each parameter individually, and do some pretty printing
             try:

GPy/core/parameterization/array_core.py

@@ -6,19 +6,6 @@ __updated__ = '2013-12-16'
 import numpy as np
 from parameter_core import Observable
 
-class ParamList(list):
-    """
-    List to store ndarray-likes in.
-    It will look for 'is' instead of calling __eq__ on each element.
-    """
-    def __contains__(self, other):
-        for el in self:
-            if el is other:
-                return True
-        return False
-
-    pass
-
 class ObservableArray(np.ndarray, Observable):
     """
     An ndarray which reports changes to its observers.
@@ -62,10 +49,11 @@ class ObservableArray(np.ndarray, Observable):
     def __setitem__(self, s, val):
         if self._s_not_empty(s):
             super(ObservableArray, self).__setitem__(s, val)
-            self._notify_observers()
+            self._notify_observers(self[s])
                 
     def __getslice__(self, start, stop):
         return self.__getitem__(slice(start, stop))
+    
     def __setslice__(self, start, stop, val):
         return self.__setitem__(slice(start, stop), val)
 

GPy/core/parameterization/index_operations.py

@@ -5,47 +5,7 @@ Created on Oct 2, 2013
 '''
 import numpy
 from numpy.lib.function_base import vectorize
-from param import Param
-from collections import defaultdict
-
-class ParamDict(defaultdict):
-    def __init__(self):
-        """
-        Default will be self._default, if not set otherwise
-        """
-        defaultdict.__init__(self, self.default_factory)
-        
-    def __getitem__(self, key):
-        try:
-            return defaultdict.__getitem__(self, key)
-        except KeyError:
-            for a in self.iterkeys():
-                if numpy.all(a==key) and a._parent_index_==key._parent_index_:
-                    return defaultdict.__getitem__(self, a)
-            raise        
-        
-    def __contains__(self, key):
-        if defaultdict.__contains__(self, key):
-            return True
-        for a in self.iterkeys():
-            if numpy.all(a==key) and a._parent_index_==key._parent_index_:
-                return True
-        return False
-
-    def __setitem__(self, key, value):
-        if isinstance(key, Param):
-            for a in self.iterkeys():
-                if numpy.all(a==key) and a._parent_index_==key._parent_index_:
-                    return super(ParamDict, self).__setitem__(a, value)
-        defaultdict.__setitem__(self, key, value)
-
-class SetDict(ParamDict):
-    def default_factory(self):
-        return set()
-
-class IntArrayDict(ParamDict):
-    def default_factory(self):
-        return numpy.int_([])
+from lists_and_dicts import IntArrayDict
 
 class ParameterIndexOperations(object):
     '''
@@ -194,9 +154,13 @@ class ParameterIndexOperationsView(object):
 
 
     def shift_right(self, start, size):
-        raise NotImplementedError, 'Shifting only supported in original ParamIndexOperations'
-    
+        self._param_index_ops.shift_right(start+self._offset, size)
 
+    def shift_left(self, start, size):
+        self._param_index_ops.shift_left(start+self._offset, size)
+        self._offset -= size
+        self._size -= size
+            
     def clear(self):
         for i, ind in self.items():
             self._param_index_ops.remove(i, ind+self._offset)
@@ -232,9 +196,7 @@ class ParameterIndexOperationsView(object):
 
     def __getitem__(self, prop):
         ind = self._filter_index(self._param_index_ops[prop])
-        if ind.size > 0:
-            return ind
-        raise KeyError, prop
+        return ind
     
     def __str__(self, *args, **kwargs):
         import pprint

GPy/core/parameterization/lists_and_dicts.py

@@ -0,0 +1,35 @@
+'''
+Created on 27 Feb 2014
+
+@author: maxz
+'''
+
+from collections import defaultdict
+class DefaultArrayDict(defaultdict):
+    def __init__(self):
+        """
+        Default will be self._default, if not set otherwise
+        """
+        defaultdict.__init__(self, self.default_factory)
+
+class SetDict(DefaultArrayDict):
+    def default_factory(self):
+        return set()
+
+class IntArrayDict(DefaultArrayDict):
+    def default_factory(self):
+        import numpy as np
+        return np.int_([])
+
+class ArrayList(list):
+    """
+    List to store ndarray-likes in.
+    It will look for 'is' instead of calling __eq__ on each element.
+    """
+    def __contains__(self, other):
+        for el in self:
+            if el is other:
+                return True
+        return False
+
+    pass

GPy/core/parameterization/param.py

@@ -3,8 +3,8 @@
 
 import itertools
 import numpy
-from parameter_core import Constrainable, Gradcheckable, Indexable, Parentable, adjust_name_for_printing
-from array_core import ObservableArray, ParamList
+from parameter_core import OptimizationHandlable, Gradcheckable, adjust_name_for_printing
+from array_core import ObservableArray
 
 ###### printing
 __constraints_name__ = "Constraint"
@@ -15,7 +15,7 @@ __precision__ = numpy.get_printoptions()['precision'] # numpy printing precision
 __print_threshold__ = 5
 ######
 
-class Param(Constrainable, ObservableArray, Gradcheckable):
+class Param(OptimizationHandlable, ObservableArray, Gradcheckable):
     """
     Parameter object for GPy models.
 
@@ -50,7 +50,7 @@ class Param(Constrainable, ObservableArray, Gradcheckable):
         obj._realsize_ = obj.size
         obj._realndim_ = obj.ndim
         obj._updated_ = False
-        from index_operations import SetDict
+        from lists_and_dicts import SetDict
         obj._tied_to_me_ = SetDict()
         obj._tied_to_ = []
         obj._original_ = True
@@ -148,8 +148,11 @@ class Param(Constrainable, ObservableArray, Gradcheckable):
     #===========================================================================
     # get/set parameters
     #===========================================================================
-    def _set_params(self, param, update=True):
+    def _set_params(self, param, trigger_parent=True):
         self.flat = param
+        if trigger_parent: min_priority = None
+        else: min_priority = -numpy.inf
+        self._notify_observers(None, min_priority)
 
     def _get_params(self):
         return self.flat
@@ -172,11 +175,9 @@ class Param(Constrainable, ObservableArray, Gradcheckable):
         try: new_arr._current_slice_ = s; new_arr._original_ = self.base is new_arr.base
         except AttributeError: pass  # returning 0d array or float, double etc
         return new_arr
+    
     def __setitem__(self, s, val):
         super(Param, self).__setitem__(s, val)
-        if self.has_parent():
-            self._direct_parent_._notify_parameters_changed()
-        #self._notify_observers()
         
     #===========================================================================
     # Index Operations:
@@ -204,6 +205,7 @@ class Param(Constrainable, ObservableArray, Gradcheckable):
         ind = self._indices(slice_index)
         if ind.ndim < 2: ind = ind[:, None]
         return numpy.asarray(numpy.apply_along_axis(lambda x: numpy.sum(extended_realshape * x), 1, ind), dtype=int)
+    
     def _expand_index(self, slice_index=None):
         # this calculates the full indexing arrays from the slicing objects given by get_item for _real..._ attributes
         # it basically translates slices to their respective index arrays and turns negative indices around
@@ -230,7 +232,8 @@ class Param(Constrainable, ObservableArray, Gradcheckable):
     #===========================================================================
     @property
     def is_fixed(self):
-        return self._highest_parent_._is_fixed(self)
+        from transformations import __fixed__
+        return self.constraints[__fixed__].size == self.size
     #def round(self, decimals=0, out=None):
     #    view = super(Param, self).round(decimals, out).view(Param)
     #    view.__array_finalize__(self)
@@ -244,7 +247,8 @@ class Param(Constrainable, ObservableArray, Gradcheckable):
     #===========================================================================
     @property
     def _description_str(self):
-        if self.size <= 1: return ["%f" % self]
+        if self.size <= 1: 
+            return [str(numpy.take(self, 0))]
         else: return [str(self.shape)]
     def parameter_names(self, add_self=False, adjust_for_printing=False):
         if adjust_for_printing:
@@ -267,7 +271,7 @@ class Param(Constrainable, ObservableArray, Gradcheckable):
         return [t._short() for t in self._tied_to_] or ['']
     def __repr__(self, *args, **kwargs):
         name = "\033[1m{x:s}\033[0;0m:\n".format(
-                            x=self.hirarchy_name())
+                            x=self.hierarchy_name())
         return name + super(Param, self).__repr__(*args, **kwargs)
     def _ties_for(self, rav_index):
         # size = sum(p.size for p in self._tied_to_)
@@ -301,12 +305,12 @@ class Param(Constrainable, ObservableArray, Gradcheckable):
         gen = map(lambda x: " ".join(map(str, x)), gen)
         return reduce(lambda a, b:max(a, len(b)), gen, len(header))
     def _max_len_values(self):
-        return reduce(lambda a, b:max(a, len("{x:=.{0}g}".format(__precision__, x=b))), self.flat, len(self.hirarchy_name()))
+        return reduce(lambda a, b:max(a, len("{x:=.{0}g}".format(__precision__, x=b))), self.flat, len(self.hierarchy_name()))
     def _max_len_index(self, ind):
         return reduce(lambda a, b:max(a, len(str(b))), ind, len(__index_name__))
     def _short(self):
         # short string to print
-        name = self.hirarchy_name()
+        name = self.hierarchy_name()
         if self._realsize_ < 2:
             return name
         ind = self._indices()
@@ -329,8 +333,8 @@ class Param(Constrainable, ObservableArray, Gradcheckable):
         if lp is None: lp = self._max_len_names(prirs, __tie_name__)
         sep = '-'
         header_format = "  {i:{5}^{2}s}  |  \033[1m{x:{5}^{1}s}\033[0;0m  |  {c:{5}^{0}s}  |  {p:{5}^{4}s}  |  {t:{5}^{3}s}"
-        if only_name: header = header_format.format(lc, lx, li, lt, lp, ' ', x=self.hirarchy_name(), c=sep*lc, i=sep*li, t=sep*lt, p=sep*lp)  # nice header for printing
-        else: header = header_format.format(lc, lx, li, lt, lp, ' ', x=self.hirarchy_name(), c=__constraints_name__, i=__index_name__, t=__tie_name__, p=__priors_name__)  # nice header for printing
+        if only_name: header = header_format.format(lc, lx, li, lt, lp, ' ', x=self.hierarchy_name(), c=sep*lc, i=sep*li, t=sep*lt, p=sep*lp)  # nice header for printing
+        else: header = header_format.format(lc, lx, li, lt, lp, ' ', x=self.hierarchy_name(), c=__constraints_name__, i=__index_name__, t=__tie_name__, p=__priors_name__)  # nice header for printing
         if not ties: ties = itertools.cycle([''])
         return "\n".join([header] + ["  {i!s:^{3}s}  |  {x: >{1}.{2}g}  |  {c:^{0}s}  |  {p:^{5}s}  |  {t:^{4}s}  ".format(lc, lx, __precision__, li, lt, lp, x=x, c=" ".join(map(str, c)), p=" ".join(map(str, p)), t=(t or ''), i=i) for i, x, c, t, p in itertools.izip(indices, vals, constr_matrix, ties, prirs)])  # return all the constraints with right indices
         # except: return super(Param, self).__str__()
@@ -345,7 +349,8 @@ class ParamConcatenation(object):
         See :py:class:`GPy.core.parameter.Param` for more details on constraining.
         """
         # self.params = params
-        self.params = ParamList([])
+        from lists_and_dicts import ArrayList
+        self.params = ArrayList([])
         for p in params:
             for p in p.flattened_parameters:
                 if p not in self.params:
@@ -353,6 +358,21 @@ class ParamConcatenation(object):
         self._param_sizes = [p.size for p in self.params]
         startstops = numpy.cumsum([0] + self._param_sizes)
         self._param_slices_ = [slice(start, stop) for start,stop in zip(startstops, startstops[1:])]
+        
+        parents = dict()
+        for p in self.params:
+            if p.has_parent():
+                parent = p._direct_parent_
+                level = 0
+                while parent is not None:
+                    if parent in parents:
+                        parents[parent] = max(level, parents[parent])
+                    else:
+                        parents[parent] = level
+                    level += 1
+                    parent = parent._direct_parent_
+        import operator
+        self.parents = map(lambda x: x[0], sorted(parents.iteritems(), key=operator.itemgetter(1)))
     #===========================================================================
     # Get/set items, enable broadcasting
     #===========================================================================
@@ -366,24 +386,26 @@ class ParamConcatenation(object):
             val = val._vals()
         ind = numpy.zeros(sum(self._param_sizes), dtype=bool); ind[s] = True;
         vals = self._vals(); vals[s] = val; del val
-        [numpy.place(p, ind[ps], vals[ps]) and update and p._notify_observers()
+        [numpy.place(p, ind[ps], vals[ps])
          for p, ps in zip(self.params, self._param_slices_)]
+        if update:
+            self.update_all_params()
     def _vals(self):
         return numpy.hstack([p._get_params() for p in self.params])
     #===========================================================================
     # parameter operations:
     #===========================================================================
     def update_all_params(self):
-        for p in self.params:
-            p._notify_observers()
-
+        for par in self.parents:
+            par._notify_observers(-numpy.inf)
+        
     def constrain(self, constraint, warning=True):
-        [param.constrain(constraint, update=False) for param in self.params]
+        [param.constrain(constraint, trigger_parent=False) for param in self.params]
         self.update_all_params()
     constrain.__doc__ = Param.constrain.__doc__
 
     def constrain_positive(self, warning=True):
-        [param.constrain_positive(warning, update=False) for param in self.params]
+        [param.constrain_positive(warning, trigger_parent=False) for param in self.params]
         self.update_all_params()
     constrain_positive.__doc__ = Param.constrain_positive.__doc__
 
@@ -393,12 +415,12 @@ class ParamConcatenation(object):
     fix = constrain_fixed
 
     def constrain_negative(self, warning=True):
-        [param.constrain_negative(warning, update=False) for param in self.params]
+        [param.constrain_negative(warning, trigger_parent=False) for param in self.params]
         self.update_all_params()
     constrain_negative.__doc__ = Param.constrain_negative.__doc__
 
     def constrain_bounded(self, lower, upper, warning=True):
-        [param.constrain_bounded(lower, upper, warning, update=False) for param in self.params]
+        [param.constrain_bounded(lower, upper, warning, trigger_parent=False) for param in self.params]
         self.update_all_params()
     constrain_bounded.__doc__ = Param.constrain_bounded.__doc__
 

GPy/core/parameterization/parameter_core.py

@@ -2,35 +2,70 @@
 # Licensed under the BSD 3-clause license (see LICENSE.txt)
 
 from transformations import Transformation, Logexp, NegativeLogexp, Logistic, __fixed__, FIXED, UNFIXED
+import numpy as np
 
 __updated__ = '2013-12-16'
 
+class HierarchyError(Exception):
+    """
+    Gets thrown when something is wrong with the parameter hierarchy
+    """
+
 def adjust_name_for_printing(name):
     if name is not None:
         return name.replace(" ", "_").replace(".", "_").replace("-", "").replace("+", "").replace("!", "").replace("*", "").replace("/", "")
     return ''
 
 class Observable(object):
+    _updated = True
     def __init__(self, *args, **kwargs):
-        from collections import defaultdict
-        self._observer_callables_ = defaultdict(list)
-
-    def add_observer(self, observer, callble):
-        self._observer_callables_[observer].append(callble)
+        self._observer_callables_ = []
 
+    def add_observer(self, observer, callble, priority=0):
+        self._insert_sorted(priority, observer, callble)
+    
     def remove_observer(self, observer, callble=None):
-        if observer in self._observer_callables_:
-            if callble is None:
-                del self._observer_callables_[observer]
-            elif callble in self._observer_callables_[observer]:
-                self._observer_callables_[observer].remove(callble)
-                if len(self._observer_callables_[observer]) == 0:
-                    self.remove_observer(observer)
-
-    def _notify_observers(self):
-        [[callble(self) for callble in callables]
-         for callables in self._observer_callables_.itervalues()]
+        to_remove = []
+        for p, obs, clble in self._observer_callables_:
+            if callble is not None:
+                if (obs == observer) and (callble == clble):
+                    to_remove.append((p, obs, clble))
+            else:
+                if obs is observer:
+                    to_remove.append((p, obs, clble))
+        for r in to_remove:
+            self._observer_callables_.remove(r)
+                
+    def _notify_observers(self, which=None, min_priority=None):
+        """
+        Notifies all observers. Which is the element, which kicked off this 
+        notification loop.
+        
+        NOTE: notifies only observers with priority p > min_priority!
+                                                    ^^^^^^^^^^^^^^^^
+        
+        :param which: object, which started this notification loop
+        :param min_priority: only notify observers with priority > min_priority
+                             if min_priority is None, notify all observers in order
+        """
+        if which is None:
+            which = self
+        if min_priority is None:
+            [callble(which) for _, _, callble in self._observer_callables_]
+        else:
+            for p, _, callble in self._observer_callables_:
+                if p <= min_priority:
+                    break
+                callble(which)
 
+    def _insert_sorted(self, p, o, c):
+        ins = 0
+        for pr, _, _ in self._observer_callables_:
+            if p > pr:
+                break
+            ins += 1
+        self._observer_callables_.insert(ins, (p, o, c))
+        
 class Pickleable(object):
     def _getstate(self):
         """
@@ -95,11 +130,11 @@ class Nameable(Parentable):
         self._name = name
         if self.has_parent():
             self._direct_parent_._name_changed(self, from_name)
-    def hirarchy_name(self, adjust_for_printing=True):
+    def hierarchy_name(self, adjust_for_printing=True):
         if adjust_for_printing: adjust = lambda x: adjust_name_for_printing(x)
         else: adjust = lambda x: x
         if self.has_parent():
-            return self._direct_parent_.hirarchy_name() + "." + adjust(self.name)
+            return self._direct_parent_.hierarchy_name() + "." + adjust(self.name)
         return adjust(self.name)
 
 
@@ -156,7 +191,7 @@ class Constrainable(Nameable, Indexable):
     #===========================================================================
     # Fixing Parameters:
     #===========================================================================
-    def constrain_fixed(self, value=None, warning=True):
+    def constrain_fixed(self, value=None, warning=True, trigger_parent=True):
         """
         Constrain this paramter to be fixed to the current value it carries.
 
@@ -164,7 +199,7 @@ class Constrainable(Nameable, Indexable):
         """
         if value is not None:
             self[:] = value
-        self.constrain(__fixed__, warning=warning)
+        self.constrain(__fixed__, warning=warning, trigger_parent=trigger_parent)
         rav_i = self._highest_parent_._raveled_index_for(self)
         self._highest_parent_._set_fixed(rav_i)
     fix = constrain_fixed
@@ -178,20 +213,17 @@ class Constrainable(Nameable, Indexable):
     unfix = unconstrain_fixed
     
     def _set_fixed(self, index):
-        import numpy as np
         if not self._has_fixes(): self._fixes_ = np.ones(self.size, dtype=bool)
         self._fixes_[index] = FIXED
         if np.all(self._fixes_): self._fixes_ = None  # ==UNFIXED
     
     def _set_unfixed(self, index):
-        import numpy as np
         if not self._has_fixes(): self._fixes_ = np.ones(self.size, dtype=bool)
         # rav_i = self._raveled_index_for(param)[index]
         self._fixes_[index] = UNFIXED
         if np.all(self._fixes_): self._fixes_ = None  # ==UNFIXED
 
     def _connect_fixes(self):
-        import numpy as np
         fixed_indices = self.constraints[__fixed__]
         if fixed_indices.size > 0:
             self._fixes_ = np.ones(self.size, dtype=bool) * UNFIXED
@@ -205,9 +237,9 @@ class Constrainable(Nameable, Indexable):
     #===========================================================================
     # Prior Operations
     #===========================================================================
-    def set_prior(self, prior, warning=True, update=True):
+    def set_prior(self, prior, warning=True, trigger_parent=True):
         repriorized = self.unset_priors()
-        self._add_to_index_operations(self.priors, repriorized, prior, warning, update)
+        self._add_to_index_operations(self.priors, repriorized, prior, warning)
     
     def unset_priors(self, *priors):
         return self._remove_from_index_operations(self.priors, priors)
@@ -221,7 +253,6 @@ class Constrainable(Nameable, Indexable):
     
     def _log_prior_gradients(self):
         """evaluate the gradients of the priors"""
-        import numpy as np
         if self.priors.size > 0:
             x = self._get_params()
             ret = np.zeros(x.size)
@@ -233,7 +264,7 @@ class Constrainable(Nameable, Indexable):
     # Constrain operations -> done
     #===========================================================================
 
-    def constrain(self, transform, warning=True, update=True):
+    def constrain(self, transform, warning=True, trigger_parent=True):
         """
         :param transform: the :py:class:`GPy.core.transformations.Transformation`
                           to constrain the this parameter to.
@@ -243,9 +274,9 @@ class Constrainable(Nameable, Indexable):
         :py:class:`GPy.core.transformations.Transformation`.
         """
         if isinstance(transform, Transformation):
-            self._set_params(transform.initialize(self._get_params()), update=False)
+            self._set_params(transform.initialize(self._get_params()), trigger_parent=trigger_parent)
         reconstrained = self.unconstrain()
-        self._add_to_index_operations(self.constraints, reconstrained, transform, warning, update)
+        self._add_to_index_operations(self.constraints, reconstrained, transform, warning)
 
     def unconstrain(self, *transforms):
         """
@@ -256,30 +287,30 @@ class Constrainable(Nameable, Indexable):
         """
         return self._remove_from_index_operations(self.constraints, transforms)
     
-    def constrain_positive(self, warning=True, update=True):
+    def constrain_positive(self, warning=True, trigger_parent=True):
         """
         :param warning: print a warning if re-constraining parameters.
 
         Constrain this parameter to the default positive constraint.
         """
-        self.constrain(Logexp(), warning=warning, update=update)
+        self.constrain(Logexp(), warning=warning, trigger_parent=trigger_parent)
 
-    def constrain_negative(self, warning=True, update=True):
+    def constrain_negative(self, warning=True, trigger_parent=True):
         """
         :param warning: print a warning if re-constraining parameters.
 
         Constrain this parameter to the default negative constraint.
         """
-        self.constrain(NegativeLogexp(), warning=warning, update=update)
+        self.constrain(NegativeLogexp(), warning=warning, trigger_parent=trigger_parent)
 
-    def constrain_bounded(self, lower, upper, warning=True, update=True):
+    def constrain_bounded(self, lower, upper, warning=True, trigger_parent=True):
         """
         :param lower, upper: the limits to bound this parameter to
         :param warning: print a warning if re-constraining parameters.
 
         Constrain this parameter to lie within the given range.
         """
-        self.constrain(Logistic(lower, upper), warning=warning, update=update)
+        self.constrain(Logistic(lower, upper), warning=warning, trigger_parent=trigger_parent)
 
     def unconstrain_positive(self):
         """
@@ -309,17 +340,15 @@ class Constrainable(Nameable, Indexable):
         for p in self._parameters_:
             p._parent_changed(parent)
 
-    def _add_to_index_operations(self, which, reconstrained, transform, warning, update):
+    def _add_to_index_operations(self, which, reconstrained, transform, warning):
         if warning and reconstrained.size > 0:
+            # TODO: figure out which parameters have changed and only print those
             print "WARNING: reconstraining parameters {}".format(self.parameter_names() or self.name)
         which.add(transform, self._raveled_index())
-        if update:
-            self._notify_observers()
 
     def _remove_from_index_operations(self, which, transforms):
         if len(transforms) == 0:
             transforms = which.properties()
-        import numpy as np
         removed = np.empty((0,), dtype=int)
         for t in transforms:
             unconstrained = which.remove(t, self._raveled_index())
@@ -329,12 +358,72 @@ class Constrainable(Nameable, Indexable):
         
         return removed
 
+class OptimizationHandlable(Constrainable, Observable):
+    def _get_params_transformed(self):
+        # transformed parameters (apply transformation rules)
+        p = self._get_params()
+        [np.put(p, ind, c.finv(p[ind])) for c, ind in self.constraints.iteritems() if c != __fixed__]
+        if self._has_fixes():
+            return p[self._fixes_]
+        return p
+    
+    def _set_params_transformed(self, p):
+        # inverse apply transformations for parameters and set the resulting parameters
+        self._set_params(self._untransform_params(p))
+    
+    def _size_transformed(self):
+        return self.size - self.constraints[__fixed__].size
+    
+    def _untransform_params(self, p):
+        p = p.copy()
+        if self._has_fixes(): tmp = self._get_params(); tmp[self._fixes_] = p; p = tmp; del tmp
+        [np.put(p, ind, c.f(p[ind])) for c, ind in self.constraints.iteritems() if c != __fixed__]
+        return p
+    
+    def _get_params(self):
+        # don't overwrite this anymore!
+        if not self.size:
+            return np.empty(shape=(0,), dtype=np.float64)
+        return np.hstack([x._get_params() for x in self._parameters_ if x.size > 0])
 
-class Parameterizable(Constrainable, Observable):
+    def _set_params(self, params, trigger_parent=True):
+        # don't overwrite this anymore!
+        raise NotImplementedError, "This needs to be implemented in Param and Parametrizable"
+    
+    #===========================================================================
+    # Optimization handles:
+    #===========================================================================
+    def _get_param_names(self):
+        n = np.array([p.hierarchy_name() + '[' + str(i) + ']' for p in self.flattened_parameters for i in p._indices()])
+        return n
+    def _get_param_names_transformed(self):
+        n = self._get_param_names()
+        if self._has_fixes():
+            return n[self._fixes_]
+        return n
+
+    #===========================================================================
+    # Randomizeable
+    #===========================================================================
+    def randomize(self):
+        """
+        Randomize the model.
+        Make this draw from the prior if one exists, else draw from N(0,1)
+        """
+        # first take care of all parameters (from N(0,1))
+        # x = self._get_params_transformed()
+        x = np.random.randn(self._size_transformed())
+        x = self._untransform_params(x)
+        # now draw from prior where possible
+        [np.put(x, ind, p.rvs(ind.size)) for p, ind in self.priors.iteritems() if not p is None]
+        self._set_params(x)
+        # self._set_params_transformed(self._get_params_transformed()) # makes sure all of the tied parameters get the same init (since there's only one prior object...)
+
+class Parameterizable(OptimizationHandlable):
     def __init__(self, *args, **kwargs):
         super(Parameterizable, self).__init__(*args, **kwargs)
-        from GPy.core.parameterization.array_core import ParamList
-        _parameters_ = ParamList()
+        from GPy.core.parameterization.lists_and_dicts import ArrayList
+        _parameters_ = ArrayList()
         self._added_names_ = set()
     
     def parameter_names(self, add_self=False, adjust_for_printing=False, recursive=True):
@@ -357,7 +446,7 @@ class Parameterizable(Constrainable, Observable):
                 if pname in self._added_names_:
                     del self.__dict__[pname]
                     self._add_parameter_name(param)
-        else:
+        elif pname not in dir(self):
             self.__dict__[pname] = param
             self._added_names_.add(pname)
             
@@ -377,42 +466,147 @@ class Parameterizable(Constrainable, Observable):
         import itertools
         [p._collect_gradient(target[s]) for p, s in itertools.izip(self._parameters_, self._param_slices_)]
 
+    def _set_params(self, params, trigger_parent=True):
+        import itertools
+        [p._set_params(params[s], trigger_parent=False) for p, s in itertools.izip(self._parameters_, self._param_slices_)]
+        if trigger_parent: min_priority = None
+        else: min_priority = -np.inf
+        self._notify_observers(None, min_priority)
+
     def _set_gradient(self, g):
         import itertools
         [p._set_gradient(g[s]) for p, s in itertools.izip(self._parameters_, self._param_slices_)]
+        
+    def add_parameter(self, param, index=None):
+        """
+        :param parameters:  the parameters to add
+        :type parameters:   list of or one :py:class:`GPy.core.param.Param`
+        :param [index]:     index of where to put parameters
 
-    def _get_params(self):
-        import numpy as np
-        # don't overwrite this anymore!
-        if not self.size:
-            return np.empty(shape=(0,), dtype=np.float64)
-        return np.hstack([x._get_params() for x in self._parameters_ if x.size > 0])
 
-    def _set_params(self, params, update=True):
-        # don't overwrite this anymore!
-        import itertools
-        [p._set_params(params[s]) for p, s in itertools.izip(self._parameters_, self._param_slices_)]
-        self._notify_parameters_changed()
+        Add all parameters to this param class, you can insert parameters
+        at any given index using the :func:`list.insert` syntax
+        """
+        # if param.has_parent():
+        #    raise AttributeError, "parameter {} already in another model, create new object (or copy) for adding".format(param._short())
+        if param in self._parameters_ and index is not None:
+            self.remove_parameter(param)
+            self.add_parameter(param, index)
+        elif param not in self._parameters_:
+            if param.has_parent():
+                parent = param._direct_parent_
+                while parent is not None:
+                    if parent is self:
+                        raise HierarchyError, "You cannot add a parameter twice into the hirarchy"
+                    parent = parent._direct_parent_
+                param._direct_parent_.remove_parameter(param)
+            # make sure the size is set
+            if index is None:
+                self.constraints.update(param.constraints, self.size)
+                self.priors.update(param.priors, self.size)
+                self._parameters_.append(param)
+            else:
+                start = sum(p.size for p in self._parameters_[:index])
+                self.constraints.shift_right(start, param.size)
+                self.priors.shift_right(start, param.size)
+                self.constraints.update(param.constraints, start)
+                self.priors.update(param.priors, start)
+                self._parameters_.insert(index, param)
+            
+            param.add_observer(self, self._pass_through_notify_observers, -np.inf)
+            
+            self.size += param.size
 
+            self._connect_parameters()
+            self._notify_parent_change()
+            self._connect_fixes()
+        else:
+            raise RuntimeError, """Parameter exists already added and no copy made"""
+
+
+    def add_parameters(self, *parameters):
+        """
+        convenience method for adding several
+        parameters without gradient specification
+        """
+        [self.add_parameter(p) for p in parameters]
+
+    def remove_parameter(self, param):
+        """
+        :param param: param object to remove from being a parameter of this parameterized object.
+        """
+        if not param in self._parameters_:
+            raise RuntimeError, "Parameter {} does not belong to this object, remove parameters directly from their respective parents".format(param._short())
+        
+        start = sum([p.size for p in self._parameters_[:param._parent_index_]])
+        self._remove_parameter_name(param)
+        self.size -= param.size
+        del self._parameters_[param._parent_index_]
+        
+        param._disconnect_parent()
+        param.remove_observer(self, self._pass_through_notify_observers)
+        self.constraints.shift_left(start, param.size)
+        
+        self._connect_fixes()
+        self._connect_parameters()
+        self._notify_parent_change()
+        
+        parent = self._direct_parent_
+        while parent is not None:
+            parent._connect_fixes()
+            parent._connect_parameters()
+            parent._notify_parent_change()
+            parent = parent._direct_parent_
+        
+    def _connect_parameters(self):
+        # connect parameterlist to this parameterized object
+        # This just sets up the right connection for the params objects
+        # to be used as parameters
+        # it also sets the constraints for each parameter to the constraints 
+        # of their respective parents 
+        if not hasattr(self, "_parameters_") or len(self._parameters_) < 1:
+            # no parameters for this class
+            return
+        sizes = [0]
+        self._param_slices_ = []
+        for i, p in enumerate(self._parameters_):
+            p._direct_parent_ = self
+            p._parent_index_ = i
+            sizes.append(p.size + sizes[-1])
+            self._param_slices_.append(slice(sizes[-2], sizes[-1]))
+            self._add_parameter_name(p)
+
+    #===========================================================================
+    # notification system
+    #===========================================================================
+    def _parameters_changed_notification(self, which):
+        self.parameters_changed()
+    def _pass_through_notify_observers(self, which):
+        self._notify_observers(which)
+    
+    #===========================================================================
+    # TODO: not working yet
+    #===========================================================================
     def copy(self):
         """Returns a (deep) copy of the current model"""
         import copy
         from .index_operations import ParameterIndexOperations, ParameterIndexOperationsView
-        from .array_core import ParamList
+        from .lists_and_dicts import ArrayList
 
         dc = dict()
         for k, v in self.__dict__.iteritems():
-            if k not in ['_direct_parent_', '_parameters_', '_parent_index_'] + self.parameter_names():
+            if k not in ['_direct_parent_', '_parameters_', '_parent_index_', '_observer_callables_'] + self.parameter_names():
                 if isinstance(v, (Constrainable, ParameterIndexOperations, ParameterIndexOperationsView)):
                     dc[k] = v.copy()
                 else:
                     dc[k] = copy.deepcopy(v)
             if k == '_parameters_':
                 params = [p.copy() for p in v]
-        
+            
         dc['_direct_parent_'] = None
         dc['_parent_index_'] = None
-        dc['_parameters_'] = ParamList()
+        dc['_observer_callables_'] = []
+        dc['_parameters_'] = ArrayList()
         dc['constraints'].clear()
         dc['priors'].clear()
         dc['size'] = 0
@@ -421,15 +615,10 @@ class Parameterizable(Constrainable, Observable):
         s.__dict__ = dc
         
         for p in params:
+            import ipdb;ipdb.set_trace()
             s.add_parameter(p)
         
         return s
-
-    def _notify_parameters_changed(self):
-        self.parameters_changed()
-        self._notify_observers()
-        if self.has_parent():
-            self._direct_parent_._notify_parameters_changed()
         
     def parameters_changed(self):
         """

GPy/core/parameterization/parameterized.py

@@ -7,9 +7,9 @@ import cPickle
 import itertools
 from re import compile, _pattern_type
 from param import ParamConcatenation
-from parameter_core import Constrainable, Pickleable, Parentable, Observable, Parameterizable, adjust_name_for_printing, Gradcheckable
+from parameter_core import Pickleable, Parameterizable, adjust_name_for_printing, Gradcheckable
 from transformations import __fixed__
-from array_core import ParamList
+from lists_and_dicts import ArrayList
 
 class Parameterized(Parameterizable, Pickleable, Gradcheckable):
     """
@@ -56,8 +56,9 @@ class Parameterized(Parameterizable, Pickleable, Gradcheckable):
     def __init__(self, name=None, *a, **kw):
         super(Parameterized, self).__init__(name=name, parent=None, parent_index=None, *a, **kw)
         self._in_init_ = True
-        self._parameters_ = ParamList()
+        self._parameters_ = ArrayList()
         self.size = sum(p.size for p in self._parameters_)
+        self.add_observer(self, self._parameters_changed_notification, -100)
         if not self._has_fixes():
             self._fixes_ = None
         self._param_slices_ = []
@@ -65,7 +66,7 @@ class Parameterized(Parameterizable, Pickleable, Gradcheckable):
         del self._in_init_
 
     def build_pydot(self, G=None):
-        import pydot
+        import pydot  # @UnresolvedImport
         iamroot = False
         if G is None:
             G = pydot.Dot(graph_type='digraph')
@@ -87,88 +88,6 @@ class Parameterized(Parameterizable, Pickleable, Gradcheckable):
             return G
         return node
 
-
-    def add_parameter(self, param, index=None):
-        """
-        :param parameters:  the parameters to add
-        :type parameters:   list of or one :py:class:`GPy.core.param.Param`
-        :param [index]:     index of where to put parameters
-
-
-        Add all parameters to this param class, you can insert parameters
-        at any given index using the :func:`list.insert` syntax
-        """
-        # if param.has_parent():
-        #    raise AttributeError, "parameter {} already in another model, create new object (or copy) for adding".format(param._short())
-        if param in self._parameters_ and index is not None:
-            self.remove_parameter(param)
-            self.add_parameter(param, index)
-        elif param not in self._parameters_:
-            # make sure the size is set
-            if index is None:
-                self.constraints.update(param.constraints, self.size)
-                self.priors.update(param.priors, self.size)
-                self._parameters_.append(param)
-            else:
-                start = sum(p.size for p in self._parameters_[:index])
-                self.constraints.shift_right(start, param.size)
-                self.priors.shift_right(start, param.size)
-                self.constraints.update(param.constraints, start)
-                self.priors.update(param.priors, start)
-                self._parameters_.insert(index, param)
-            self.size += param.size
-        else:
-            raise RuntimeError, """Parameter exists already added and no copy made"""
-        self._connect_parameters()
-        self._notify_parent_change()
-        self._connect_fixes()
-
-
-    def add_parameters(self, *parameters):
-        """
-        convenience method for adding several
-        parameters without gradient specification
-        """
-        [self.add_parameter(p) for p in parameters]
-
-    def remove_parameter(self, param):
-        """
-        :param param: param object to remove from being a parameter of this parameterized object.
-        """
-        if not param in self._parameters_:
-            raise RuntimeError, "Parameter {} does not belong to this object, remove parameters directly from their respective parents".format(param._short())
-        
-        start = sum([p.size for p in self._parameters_[:param._parent_index_]])
-        self._remove_parameter_name(param)
-        self.size -= param.size
-        del self._parameters_[param._parent_index_]
-        
-        param._disconnect_parent()
-        param.remove_observer(self, self._notify_parameters_changed)
-        self.constraints.shift_left(start, param.size)
-        self._connect_fixes()
-        self._connect_parameters()
-        self._notify_parent_change()
-        
-        
-    def _connect_parameters(self):
-        # connect parameterlist to this parameterized object
-        # This just sets up the right connection for the params objects
-        # to be used as parameters
-        # it also sets the constraints for each parameter to the constraints 
-        # of their respective parents 
-        if not hasattr(self, "_parameters_") or len(self._parameters_) < 1:
-            # no parameters for this class
-            return
-        sizes = [0]
-        self._param_slices_ = []
-        for i, p in enumerate(self._parameters_):
-            p._direct_parent_ = self
-            p._parent_index_ = i
-            sizes.append(p.size + sizes[-1])
-            self._param_slices_.append(slice(sizes[-2], sizes[-1]))
-            self._add_parameter_name(p)
-
     #===========================================================================
     # Pickling operations
     #===========================================================================
@@ -185,6 +104,11 @@ class Parameterized(Parameterizable, Pickleable, Gradcheckable):
         else:
             cPickle.dump(self, f, protocol)
 
+    def copy(self):
+        c = super(Parameterized, self).copy()
+        c.add_observer(c, c._parameters_changed_notification, -100)
+        return c
+    
     def __getstate__(self):
         if self._has_get_set_state():
             return self._getstate()
@@ -237,42 +161,7 @@ class Parameterized(Parameterizable, Pickleable, Gradcheckable):
                 g[self._offset_for(p) + numpy.array(list(i))] += g[self._raveled_index_for(t)]
         if self._has_fixes(): return g[self._fixes_]
         return g
-    #===========================================================================
-    # Optimization handles:
-    #===========================================================================
-    def _get_param_names(self):
-        n = numpy.array([p.hirarchy_name() + '[' + str(i) + ']' for p in self.flattened_parameters for i in p._indices()])
-        return n
-    def _get_param_names_transformed(self):
-        n = self._get_param_names()
-        if self._has_fixes():
-            return n[self._fixes_]
-        return n
-    def _get_params_transformed(self):
-        # transformed parameters (apply transformation rules)
-        p = self._get_params()
-        [numpy.put(p, ind, c.finv(p[ind])) for c, ind in self.constraints.iteritems() if c != __fixed__]
-        if self._has_fixes():
-            return p[self._fixes_]
-        return p
-    def _set_params_transformed(self, p):
-        # inverse apply transformations for parameters and set the resulting parameters
-        self._set_params(self._untransform_params(p))
-    def _untransform_params(self, p):
-        p = p.copy()
-        if self._has_fixes(): tmp = self._get_params(); tmp[self._fixes_] = p; p = tmp; del tmp
-        [numpy.put(p, ind, c.f(p[ind])) for c, ind in self.constraints.iteritems() if c != __fixed__]
-        return p
-    #===========================================================================
-    # Indexable Handling
-    #===========================================================================
-    def _backtranslate_index(self, param, ind):
-        # translate an index in parameterized indexing into the index of param
-        ind = ind - self._offset_for(param)
-        ind = ind[ind >= 0]
-        internal_offset = param._internal_offset()
-        ind = ind[ind < param.size + internal_offset]
-        return ind
+
     def _offset_for(self, param):
         # get the offset in the parameterized index array for param
         if param.has_parent():
@@ -297,34 +186,22 @@ class Parameterized(Parameterizable, Pickleable, Gradcheckable):
         this is not in the global view of things!
         """
         return numpy.r_[:self.size]
-    #===========================================================================
-    # Fixing parameters:
-    #===========================================================================
-    def _fixes_for(self, param):
-        if self._has_fixes():
-            return self._fixes_[self._raveled_index_for(param)]
-        return numpy.ones(self.size, dtype=bool)[self._raveled_index_for(param)]
+    
     #===========================================================================
     # Convenience for fixed, tied checking of param:
     #===========================================================================
-    def fixed_indices(self):
-        return np.array([x.is_fixed for x in self._parameters_])
-    def _is_fixed(self, param):
-        # returns if the whole param is fixed
-        if not self._has_fixes():
-            return False
-        return not self._fixes_[self._raveled_index_for(param)].any()
-        # return not self._fixes_[self._offset_for(param): self._offset_for(param)+param._realsize_].any()
     @property
     def is_fixed(self):
         for p in self._parameters_:
             if not p.is_fixed: return False
         return True
+
     def _get_original(self, param):
         # if advanced indexing is activated it happens that the array is a copy
         # you can retrieve the original param through this method, by passing
         # the copy here
         return self._parameters_[param._parent_index_]
+    
     #===========================================================================
     # Get/set parameters:
     #===========================================================================
@@ -352,9 +229,13 @@ class Parameterized(Parameterizable, Pickleable, Gradcheckable):
         return ParamConcatenation(paramlist)
     
     def __setitem__(self, name, value, paramlist=None):
-        try: param = self.__getitem__(name, paramlist)
-        except AttributeError as a: raise a
-        param[:] = value
+        if isinstance(name, slice):
+            self[''][name] = value
+        else:
+            try: param = self.__getitem__(name, paramlist)
+            except AttributeError as a: raise a
+            param[:] = value
+        
     def __setattr__(self, name, val):
         # override the default behaviour, if setting a param, so broadcasting can by used        
         if hasattr(self, '_parameters_'):
@@ -365,7 +246,7 @@ class Parameterized(Parameterizable, Pickleable, Gradcheckable):
     # Printing:
     #===========================================================================
     def _short(self):
-        return self.hirarchy_name()
+        return self.hierarchy_name()
     @property
     def flattened_parameters(self):
         return [xi for x in self._parameters_ for xi in x.flattened_parameters]
@@ -373,11 +254,6 @@ class Parameterized(Parameterizable, Pickleable, Gradcheckable):
     def _parameter_sizes_(self):
         return [x.size for x in self._parameters_]
     @property
-    def size_transformed(self):
-        if self._has_fixes():
-            return sum(self._fixes_)
-        return self.size
-    @property
     def parameter_shapes(self):
         return [xi for x in self._parameters_ for xi in x.parameter_shapes]
     @property
@@ -404,7 +280,7 @@ class Parameterized(Parameterizable, Pickleable, Gradcheckable):
         cl = max([len(str(x)) if x else 0 for x in constrs + ["Constraint"]])
         tl = max([len(str(x)) if x else 0 for x in ts + ["Tied to"]])
         pl = max([len(str(x)) if x else 0 for x in prirs + ["Prior"]])
-        format_spec = "  \033[1m{{name:<{0}s}}\033[0;0m  |  {{desc:^{1}s}}  |  {{const:^{2}s}}  |  {{pri:^{3}s}}  |  {{t:^{4}s}}".format(nl, sl, cl, pl, tl)
+        format_spec = "  \033[1m{{name:<{0}s}}\033[0;0m  |  {{desc:>{1}s}}  |  {{const:^{2}s}}  |  {{pri:^{3}s}}  |  {{t:^{4}s}}".format(nl, sl, cl, pl, tl)
         to_print = []
         for n, d, c, t, p in itertools.izip(names, desc, constrs, ts, prirs):
             to_print.append(format_spec.format(name=n, desc=d, const=c, t=t, pri=p))

GPy/core/parameterization/priors.py

@@ -64,6 +64,36 @@ class Gaussian(Prior):
         return np.random.randn(n) * self.sigma + self.mu
 
 
+class Uniform(Prior):
+    domain = _REAL
+    _instances = []
+    def __new__(cls, lower, upper): # Singleton:
+        if cls._instances:
+            cls._instances[:] = [instance for instance in cls._instances if instance()]
+            for instance in cls._instances:
+                if instance().lower == lower and instance().upper == upper:
+                    return instance()
+        o = super(Prior, cls).__new__(cls, lower, upper)
+        cls._instances.append(weakref.ref(o))
+        return cls._instances[-1]()
+    
+    def __init__(self, lower, upper):
+        self.lower = float(lower)
+        self.upper = float(upper)
+    
+    def __str__(self):
+        return "[" + str(np.round(self.lower)) + ', ' + str(np.round(self.upper)) + ']'
+
+    def lnpdf(self, x):
+        region = (x>=self.lower) * (x<=self.upper)
+        return region
+
+    def lnpdf_grad(self, x):
+        return np.zeros(x.shape)
+
+    def rvs(self, n):
+        return np.random.uniform(self.lower, self.upper, size=n)
+    
 class LogGaussian(Prior):
     """
     Implementation of the univariate *log*-Gaussian probability function, coupled with random variables.

GPy/core/parameterization/transformations.py

@@ -6,8 +6,11 @@ import numpy as np
 from domains import _POSITIVE,_NEGATIVE, _BOUNDED
 import weakref
 
+import sys
+#_lim_val = -np.log(sys.float_info.epsilon)
+
 _exp_lim_val = np.finfo(np.float64).max
-_lim_val = np.log(_exp_lim_val)#-np.log(sys.float_info.epsilon)
+_lim_val = np.log(_exp_lim_val)#
 
 #===============================================================================
 # Fixing constants
@@ -35,7 +38,6 @@ class Transformation(object):
         """ produce a sensible initial value for f(x)"""
         raise NotImplementedError
     def plot(self, xlabel=r'transformed $\theta$', ylabel=r'$\theta$', axes=None, *args,**kw):
-        import sys
         assert "matplotlib" in sys.modules, "matplotlib package has not been imported."
         import matplotlib.pyplot as plt
         from ...plotting.matplot_dep import base_plots
@@ -52,7 +54,7 @@ class Transformation(object):
 class Logexp(Transformation):
     domain = _POSITIVE
     def f(self, x):
-        return np.where(x>_lim_val, x, np.log(1. + np.exp(np.clip(x, -np.inf, _lim_val))))
+        return np.where(x>_lim_val, x, np.log(1. + np.exp(np.clip(x, -_lim_val, _lim_val))))
         #raises overflow warning: return np.where(x>_lim_val, x, np.log(1. + np.exp(x)))
     def finv(self, f):
         return np.where(f>_lim_val, f, np.log(np.exp(f) - 1.))

GPy/core/sparse_gp.py

@@ -85,11 +85,11 @@ class SparseGP(GP):
             self.Z.gradient = self.kern.gradients_X(self.grad_dict['dL_dKmm'], self.Z)
             self.Z.gradient += self.kern.gradients_X(self.grad_dict['dL_dKnm'].T, self.Z, self.X)
 
-    def _raw_predict(self, Xnew, X_variance_new=None, full_cov=False):
+    def _raw_predict(self, Xnew, full_cov=False):
         """
         Make a prediction for the latent function values
         """
-        if X_variance_new is None:
+        if not isinstance(Xnew, VariationalPosterior):
             Kx = self.kern.K(self.Z, Xnew)
             mu = np.dot(Kx.T, self.posterior.woodbury_vector)
             if full_cov:
@@ -100,13 +100,13 @@ class SparseGP(GP):
                 Kxx = self.kern.Kdiag(Xnew)
                 var = (Kxx - np.sum(np.dot(np.atleast_3d(self.posterior.woodbury_inv).T, Kx) * Kx[None,:,:], 1)).T
         else:
-            Kx = self.kern.psi1(self.Z, Xnew, X_variance_new)
-            mu = np.dot(Kx, self.Cpsi1V)
+            Kx = self.kern.psi1(self.Z, Xnew)
+            mu = np.dot(Kx, self.posterior.woodbury_vector)
             if full_cov:
                 raise NotImplementedError, "TODO"
             else:
-                Kxx = self.kern.psi0(self.Z, Xnew, X_variance_new)
-                psi2 = self.kern.psi2(self.Z, Xnew, X_variance_new)
+                Kxx = self.kern.psi0(self.Z, Xnew)
+                psi2 = self.kern.psi2(self.Z, Xnew)
                 var = Kxx - np.sum(np.sum(psi2 * Kmmi_LmiBLmi[None, :, :], 1), 1)
         return mu, var
 
@@ -114,14 +114,12 @@ class SparseGP(GP):
     def _getstate(self):
         """
         Get the current state of the class,
-        here just all the indices, rest can get recomputed
         """
-        return GP._getstate(self) + [self.Z,
-                self.num_inducing,
-                self.X_variance]
+        return GP._getstate(self) + [
+                self.Z,
+                self.num_inducing]
 
     def _setstate(self, state):
-        self.X_variance = state.pop()
         self.num_inducing = state.pop()
         self.Z = state.pop()
         GP._setstate(self, state)

GPy/examples/dimensionality_reduction.py

@@ -187,10 +187,10 @@ def _simulate_sincos(D1, D2, D3, N, num_inducing, Q, plot_sim=False):
     _np.random.seed(1234)
     
     x = _np.linspace(0, 4 * _np.pi, N)[:, None]
-    s1 = _np.vectorize(lambda x: -_np.sin(_np.exp(x)))
+    s1 = _np.vectorize(lambda x: _np.sin(x))
     s2 = _np.vectorize(lambda x: _np.cos(x)**2)
     s3 = _np.vectorize(lambda x:-_np.exp(-_np.cos(2 * x)))
-    sS = _np.vectorize(lambda x: x*_np.sin(x))
+    sS = _np.vectorize(lambda x: _np.cos(x))
 
     s1 = s1(x)
     s2 = s2(x)
@@ -202,7 +202,7 @@ def _simulate_sincos(D1, D2, D3, N, num_inducing, Q, plot_sim=False):
     s3 -= s3.mean(); s3 /= s3.std(0)
     sS -= sS.mean(); sS /= sS.std(0)
 
-    S1 = _np.hstack([s1, s2, sS])
+    S1 = _np.hstack([s1, sS])
     S2 = _np.hstack([s2, s3, sS])
     S3 = _np.hstack([s3, sS])
 
@@ -270,7 +270,7 @@ def bgplvm_simulation(optimize=True, verbose=1,
     from GPy import kern
     from GPy.models import BayesianGPLVM
 
-    D1, D2, D3, N, num_inducing, Q = 13, 5, 8, 45, 5, 9
+    D1, D2, D3, N, num_inducing, Q = 13, 5, 8, 45, 3, 9
     _, _, Ylist = _simulate_sincos(D1, D2, D3, N, num_inducing, Q, plot_sim)
     Y = Ylist[0]
     k = kern.Linear(Q, ARD=True)# + kern.white(Q, _np.exp(-2)) # + kern.bias(Q)
@@ -294,7 +294,7 @@ def bgplvm_simulation_missing_data(optimize=True, verbose=1,
     from GPy.models import BayesianGPLVM
     from GPy.inference.latent_function_inference.var_dtc import VarDTCMissingData
 
-    D1, D2, D3, N, num_inducing, Q = 13, 5, 8, 45, 5, 9
+    D1, D2, D3, N, num_inducing, Q = 13, 5, 8, 45, 7, 9
     _, _, Ylist = _simulate_sincos(D1, D2, D3, N, num_inducing, Q, plot_sim)
     Y = Ylist[0]
     k = kern.Linear(Q, ARD=True)# + kern.white(Q, _np.exp(-2)) # + kern.bias(Q)

GPy/inference/latent_function_inference/var_dtc.py

@@ -60,8 +60,7 @@ class VarDTC(object):
         _, output_dim = Y.shape
 
         #see whether we've got a different noise variance for each datum
-        beta = 1./np.squeeze(likelihood.variance)
-
+        beta = 1./np.fmax(likelihood.variance, 1e-6)
         # VVT_factor is a matrix such that tdot(VVT_factor) = VVT...this is for efficiency!
         #self.YYTfactor = self.get_YYTfactor(Y)
         #VVT_factor = self.get_VVTfactor(self.YYTfactor, beta)
@@ -214,7 +213,7 @@ class VarDTCMissingData(object):
             psi2_all = None
 
         Ys, traces = self._Y(Y)
-        beta_all = 1./likelihood.variance
+        beta_all = 1./np.fmax(likelihood.variance, 1e-6)
         het_noise = beta_all.size != 1
 
         import itertools

GPy/kern/_src/kern.py

@@ -73,7 +73,7 @@ class Kern(Parameterized):
         See GPy.plotting.matplot_dep.plot
         """
         assert "matplotlib" in sys.modules, "matplotlib package has not been imported."
-        from ..plotting.matplot_dep import kernel_plots
+        from ...plotting.matplot_dep import kernel_plots
         kernel_plots.plot(self,*args)
 
     def plot_ARD(self, *args, **kw):
@@ -112,10 +112,12 @@ class Kern(Parameterized):
         """
         assert isinstance(other, Kern), "only kernels can be added to kernels..."
         from add import Add
-        return Add([self, other], tensor)
-
-    def __call__(self, X, X2=None):
-        return self.K(X, X2)
+        kernels = []
+        if not tensor and isinstance(self, Add): kernels.extend(self._parameters_)
+        else: kernels.append(self)
+        if not tensor and isinstance(other, Add): kernels.extend(other._parameters_)
+        else: kernels.append(other)
+        return Add(kernels, tensor)
 
     def __mul__(self, other):
         """ Here we overload the '*' operator. See self.prod for more information"""

GPy/kern/_src/rbf.py

@@ -7,6 +7,8 @@ from scipy import weave
 from ...util.misc import param_to_array
 from stationary import Stationary
 from GPy.util.caching import Cache_this
+from ...core.parameterization import variational
+from rbf_psi_comp import ssrbf_psi_comp
 
 class RBF(Stationary):
     """
@@ -18,7 +20,7 @@ class RBF(Stationary):
 
     """
 
-    def __init__(self, input_dim, variance=1., lengthscale=None, ARD=False, name='RBF'):
+    def __init__(self, input_dim, variance=1., lengthscale=None, ARD=False, name='rbf'):
         super(RBF, self).__init__(input_dim, variance, lengthscale, ARD, name)
         self.weave_options = {}
 
@@ -36,76 +38,140 @@ class RBF(Stationary):
         return self.Kdiag(variational_posterior.mean)
 
     def psi1(self, Z, variational_posterior):
-        _, _, _, psi1 = self._psi1computations(Z, variational_posterior)
+        if isinstance(variational_posterior, variational.SpikeAndSlabPosterior):
+            psi1, _, _, _, _, _, _ = ssrbf_psi_comp._psi1computations(self.variance, self.lengthscale, Z, variational_posterior.mean, variational_posterior.variance, variational_posterior.binary_prob)
+        else:
+            _, _, _, psi1 = self._psi1computations(Z, variational_posterior)
         return psi1
 
     def psi2(self, Z, variational_posterior):
-        _, _, _, _, _, psi2 = self._psi2computations(Z, variational_posterior)
+        if isinstance(variational_posterior, variational.SpikeAndSlabPosterior):
+            psi2, _, _, _, _, _, _ = ssrbf_psi_comp._psi2computations(self.variance, self.lengthscale, Z, variational_posterior.mean, variational_posterior.variance, variational_posterior.binary_prob)
+        else:
+            _, _, _, _, psi2 = self._psi2computations(Z, variational_posterior)
         return psi2
 
     def update_gradients_expectations(self, dL_dpsi0, dL_dpsi1, dL_dpsi2, Z, variational_posterior):
-        l2 = self.lengthscale **2
+        # Spike-and-Slab GPLVM
+        if isinstance(variational_posterior, variational.SpikeAndSlabPosterior):
+            _, _dpsi1_dvariance, _, _, _, _, _dpsi1_dlengthscale = ssrbf_psi_comp._psi1computations(self.variance, self.lengthscale, Z, variational_posterior.mean, variational_posterior.variance, variational_posterior.binary_prob)
+            _, _dpsi2_dvariance, _, _, _, _, _dpsi2_dlengthscale = ssrbf_psi_comp._psi2computations(self.variance, self.lengthscale, Z, variational_posterior.mean, variational_posterior.variance, variational_posterior.binary_prob)
+    
+            #contributions from psi0:
+            self.variance.gradient = np.sum(dL_dpsi0)
+    
+            #from psi1
+            self.variance.gradient += np.sum(dL_dpsi1 * _dpsi1_dvariance)
+            self.lengthscale.gradient = (dL_dpsi1[:,:,None]*_dpsi1_dlengthscale).reshape(-1,self.input_dim).sum(axis=0) 
+        
+    
+            #from psi2
+            self.variance.gradient += (dL_dpsi2 * _dpsi2_dvariance).sum()
+            self.lengthscale.gradient += (dL_dpsi2[:,:,:,None] * _dpsi2_dlengthscale).reshape(-1,self.input_dim).sum(axis=0)        
+            return
+        
+        elif isinstance(variational_posterior, variational.NormalPosterior):
+        
+            l2 = self.lengthscale **2
 
-        #contributions from psi0:
-        self.variance.gradient = np.sum(dL_dpsi0)
-        self.lengthscale.gradient = 0.
+            #contributions from psi0:
+            self.variance.gradient = np.sum(dL_dpsi0)
+            self.lengthscale.gradient = 0.
+
+            #from psi1
+            denom, _, dist_sq, psi1 = self._psi1computations(Z, variational_posterior)
+            d_length = psi1[:,:,None] * ((dist_sq - 1.)/(self.lengthscale*denom) +1./self.lengthscale)
+            dpsi1_dlength = d_length * dL_dpsi1[:, :, None]
+            if self.ARD:
+                self.lengthscale.gradient += dpsi1_dlength.sum(0).sum(0)
+            else:
+                self.lengthscale.gradient += dpsi1_dlength.sum()
+            self.variance.gradient += np.sum(dL_dpsi1 * psi1) / self.variance
+
+            #from psi2
+            S = variational_posterior.variance
+            _, Zdist_sq, _, mudist_sq, psi2 = self._psi2computations(Z, variational_posterior)
+
+            if not self.ARD:
+                self.lengthscale.gradient += self._weave_psi2_lengthscale_grads(dL_dpsi2, psi2, Zdist_sq, S, mudist_sq, l2).sum()
+            else:
+                self.lengthscale.gradient += self._weave_psi2_lengthscale_grads(dL_dpsi2, psi2, Zdist_sq, S, mudist_sq, l2)
+
+            self.variance.gradient += 2.*np.sum(dL_dpsi2 * psi2)/self.variance
 
-        #from psi1
-        denom, _, dist_sq, psi1 = self._psi1computations(Z, variational_posterior)
-        d_length = psi1[:,:,None] * ((dist_sq - 1.)/(self.lengthscale*denom) +1./self.lengthscale)
-        dpsi1_dlength = d_length * dL_dpsi1[:, :, None]
-        if not self.ARD:
-            self.lengthscale.gradient += dpsi1_dlength.sum()
         else:
-            self.lengthscale.gradient += dpsi1_dlength.sum(0).sum(0)
-        self.variance.gradient += np.sum(dL_dpsi1 * psi1) / self.variance
-
-        #from psi2
-        S = variational_posterior.variance
-        denom, _, Zdist_sq, _, mudist_sq, psi2 = self._psi2computations(Z, variational_posterior)
-        d_length = 2.*psi2[:, :, :, None] * (Zdist_sq * denom + mudist_sq + S[:, None, None, :] / l2) / (self.lengthscale * denom)
-        #TODO: combine denom and l2 as denom_l2??
-        #TODO: tidy the above!
-        #TODO: tensordot below?
-
-        dpsi2_dlength = d_length * dL_dpsi2[:, :, :, None]
-        if not self.ARD:
-            self.lengthscale.gradient += dpsi2_dlength.sum()
-        else:
-            self.lengthscale.gradient += dpsi2_dlength.sum(0).sum(0).sum(0)
-
-        self.variance.gradient += 2.*np.sum(dL_dpsi2 * psi2)/self.variance
+            raise ValueError, "unknown distriubtion received for psi-statistics"
 
     def gradients_Z_expectations(self, dL_dpsi1, dL_dpsi2, Z, variational_posterior):
-        l2 = self.lengthscale **2
+        # Spike-and-Slab GPLVM
+        if isinstance(variational_posterior, variational.SpikeAndSlabPosterior):
+            _, _, _, _, _, _dpsi1_dZ, _ = ssrbf_psi_comp._psi1computations(self.variance, self.lengthscale, Z, variational_posterior.mean, variational_posterior.variance, variational_posterior.binary_prob)
+            _, _, _, _, _, _dpsi2_dZ, _ = ssrbf_psi_comp._psi2computations(self.variance, self.lengthscale, Z, variational_posterior.mean, variational_posterior.variance, variational_posterior.binary_prob)
+    
+            #psi1
+            grad = (dL_dpsi1[:, :, None] * _dpsi1_dZ).sum(axis=0)
+    
+            #psi2
+            grad += (dL_dpsi2[:, :, :, None] * _dpsi2_dZ).sum(axis=0).sum(axis=1)
+    
+            return grad
 
-        #psi1
-        denom, dist, dist_sq, psi1 = self._psi1computations(Z, variational_posterior)
-        denominator = l2 * denom
-        dpsi1_dZ = -psi1[:, :, None] * (dist / denominator)
-        grad = np.sum(dL_dpsi1[:, :, None] * dpsi1_dZ, 0)
+        elif isinstance(variational_posterior, variational.NormalPosterior):
+            
+            l2 = self.lengthscale **2
 
-        #psi2
-        denom, Zdist, Zdist_sq, mudist, mudist_sq, psi2 = self._psi2computations(Z, variational_posterior)
-        term1 = Zdist / l2 # M, M, Q
-        term2 = mudist / denom / l2 # N, M, M, Q
-        dZ = psi2[:, :, :, None] * (term1[None, :, :, :] + term2) #N,M,M,Q
-        grad += 2*(dL_dpsi2[:, :, :, None] * dZ).sum(0).sum(0)
+            #psi1
+            denom, dist, dist_sq, psi1 = self._psi1computations(Z, variational_posterior)
+            grad = np.einsum('ij,ij,ijk,ijk->jk', dL_dpsi1, psi1, dist, -1./(denom*l2))
 
-        return grad
+            #psi2
+            Zdist, Zdist_sq, mudist, mudist_sq, psi2 = self._psi2computations(Z, variational_posterior)
+            term1 = Zdist / l2 # M, M, Q
+            S = variational_posterior.variance
+            term2 = mudist / (2.*S[:,None,None,:] + l2) # N, M, M, Q
+
+            grad += 2.*np.einsum('ijk,ijk,ijkl->kl', dL_dpsi2, psi2, term1[None,:,:,:] + term2)
+
+            return grad
+        else:
+            raise ValueError, "unknown distriubtion received for psi-statistics"
 
     def gradients_qX_expectations(self, dL_dpsi0, dL_dpsi1, dL_dpsi2, Z, variational_posterior):
-        l2 = self.lengthscale **2
-        #psi1
-        denom, dist, dist_sq, psi1 = self._psi1computations(Z, variational_posterior)
-        tmp = psi1[:, :, None] / l2 / denom
-        grad_mu = np.sum(dL_dpsi1[:, :, None] * tmp * dist, 1)
-        grad_S = np.sum(dL_dpsi1[:, :, None] * 0.5 * tmp * (dist_sq - 1), 1)
-        #psi2
-        denom, Zdist, Zdist_sq, mudist, mudist_sq, psi2 = self._psi2computations(Z, variational_posterior)
-        tmp = psi2[:, :, :, None] / l2 / denom
-        grad_mu += -2.*(dL_dpsi2[:, :, :, None] * tmp * mudist).sum(1).sum(1)
-        grad_S += (dL_dpsi2[:, :, :, None] * tmp * (2.*mudist_sq - 1)).sum(1).sum(1)
+        # Spike-and-Slab GPLVM
+        if isinstance(variational_posterior, variational.SpikeAndSlabPosterior):
+            ndata = variational_posterior.mean.shape[0]
+            
+            _, _, _dpsi1_dgamma, _dpsi1_dmu, _dpsi1_dS, _, _ = ssrbf_psi_comp._psi1computations(self.variance, self.lengthscale, Z, variational_posterior.mean, variational_posterior.variance, variational_posterior.binary_prob)
+            _, _, _dpsi2_dgamma, _dpsi2_dmu, _dpsi2_dS, _, _ = ssrbf_psi_comp._psi2computations(self.variance, self.lengthscale, Z, variational_posterior.mean, variational_posterior.variance, variational_posterior.binary_prob)
+    
+            #psi1
+            grad_mu = (dL_dpsi1[:, :, None] * _dpsi1_dmu).sum(axis=1)
+            grad_S = (dL_dpsi1[:, :, None] * _dpsi1_dS).sum(axis=1)
+            grad_gamma = (dL_dpsi1[:,:,None] * _dpsi1_dgamma).sum(axis=1)
+            #psi2
+            grad_mu += (dL_dpsi2[:, :, :, None] * _dpsi2_dmu).reshape(ndata,-1,self.input_dim).sum(axis=1)
+            grad_S += (dL_dpsi2[:, :, :, None] * _dpsi2_dS).reshape(ndata,-1,self.input_dim).sum(axis=1)
+            grad_gamma += (dL_dpsi2[:,:,:, None] * _dpsi2_dgamma).reshape(ndata,-1,self.input_dim).sum(axis=1)
+            
+            return grad_mu, grad_S, grad_gamma
+
+        elif isinstance(variational_posterior, variational.NormalPosterior):
+        
+            l2 = self.lengthscale **2
+            #psi1
+            denom, dist, dist_sq, psi1 = self._psi1computations(Z, variational_posterior)
+            tmp = psi1[:, :, None] / l2 / denom
+            grad_mu = np.sum(dL_dpsi1[:, :, None] * tmp * dist, 1)
+            grad_S = np.sum(dL_dpsi1[:, :, None] * 0.5 * tmp * (dist_sq - 1), 1)
+            #psi2
+            _, _, mudist, mudist_sq, psi2 = self._psi2computations(Z, variational_posterior)
+            S = variational_posterior.variance
+            tmp = psi2[:, :, :, None] / (2.*S[:,None,None,:] + l2)
+            grad_mu += -2.*np.einsum('ijk,ijkl,ijkl->il', dL_dpsi2, tmp , mudist)
+            grad_S += np.einsum('ijk,ijkl,ijkl->il', dL_dpsi2 , tmp , (2.*mudist_sq - 1))
+
+        else:
+            raise ValueError, "unknown distriubtion received for psi-statistics"
 
         return grad_mu, grad_S
 
@@ -113,61 +179,6 @@ class RBF(Stationary):
     #            Precomputations            #
     #---------------------------------------#
 
-    #TODO: this function is unused, but it will be useful in the stationary class
-    def _dL_dlengthscales_via_K(self, dL_dK, X, X2):
-        """
-        A helper function for update_gradients_* methods
-
-        Computes the derivative of the objective L wrt the lengthscales via
-
-        dL_dl = sum_{i,j}(dL_dK_{ij} dK_dl)
-
-        assumes self._K_computations has just been called.
-
-        This is only valid if self.ARD=True
-        """
-        target = np.zeros(self.input_dim)
-        dvardLdK = self._K_dvar * dL_dK
-        var_len3 = self.variance / np.power(self.lengthscale, 3)
-        if X2 is None:
-            # save computation for the symmetrical case
-            dvardLdK = dvardLdK + dvardLdK.T
-            code = """
-            int q,i,j;
-            double tmp;
-            for(q=0; q<input_dim; q++){
-              tmp = 0;
-              for(i=0; i<num_data; i++){
-                for(j=0; j<i; j++){
-                  tmp += (X(i,q)-X(j,q))*(X(i,q)-X(j,q))*dvardLdK(i,j);
-                }
-              }
-              target(q) += var_len3(q)*tmp;
-            }
-            """
-            num_data, num_inducing, input_dim = X.shape[0], X.shape[0], self.input_dim
-            X, dvardLdK, var_len3 = param_to_array(X, dvardLdK, var_len3)
-            weave.inline(code, arg_names=['num_data', 'num_inducing', 'input_dim', 'X', 'target', 'dvardLdK', 'var_len3'], type_converters=weave.converters.blitz, **self.weave_options)
-        else:
-            code = """
-            int q,i,j;
-            double tmp;
-            for(q=0; q<input_dim; q++){
-              tmp = 0;
-              for(i=0; i<num_data; i++){
-                for(j=0; j<num_inducing; j++){
-                  tmp += (X(i,q)-X2(j,q))*(X(i,q)-X2(j,q))*dvardLdK(i,j);
-                }
-              }
-              target(q) += var_len3(q)*tmp;
-            }
-            """
-            num_data, num_inducing, input_dim = X.shape[0], X2.shape[0], self.input_dim
-            X, X2, dvardLdK, var_len3 = param_to_array(X, X2, dvardLdK, var_len3)
-            weave.inline(code, arg_names=['num_data', 'num_inducing', 'input_dim', 'X', 'X2', 'target', 'dvardLdK', 'var_len3'], type_converters=weave.converters.blitz, **self.weave_options)
-        return target
-
-
     @Cache_this(limit=1)
     def _psi1computations(self, Z, vp):
         mu, S = vp.mean, vp.variance
@@ -180,7 +191,7 @@ class RBF(Stationary):
         return denom, dist, dist_sq, psi1
 
 
-    #@cache_this(ignore_args=(1,))
+    @Cache_this(limit=1, ignore_args=(0,))
     def _Z_distances(self, Z):
         Zhat = 0.5 * (Z[:, None, :] + Z[None, :, :]) # M,M,Q
         Zdist = 0.5 * (Z[:, None, :] - Z[None, :, :]) # M,M,Q
@@ -200,7 +211,6 @@ class RBF(Stationary):
         #allocate memory for the things we want to compute
         mudist = np.empty((N, M, M, Q))
         mudist_sq = np.empty((N, M, M, Q))
-        exponent = np.zeros((N,M,M))
         psi2 = np.empty((N, M, M))
 
         l2 = self.lengthscale **2
@@ -212,7 +222,7 @@ class RBF(Stationary):
         code = """
         double tmp, exponent_tmp;
 
-        //#pragma omp parallel for private(tmp, exponent_tmp)
+        #pragma omp parallel for private(tmp, exponent_tmp)
         for (int n=0; n<N; n++)
         {
             for (int m=0; m<M; m++)
@@ -253,8 +263,48 @@ class RBF(Stationary):
                      arg_names=['N', 'M', 'Q', 'mu', 'Zhat', 'mudist_sq', 'mudist', 'denom_l2', 'Zdist_sq', 'half_log_denom', 'psi2', 'variance_sq'],
                      type_converters=weave.converters.blitz, **self.weave_options)
 
-        return denom, Zdist, Zdist_sq, mudist, mudist_sq, psi2
+        return Zdist, Zdist_sq, mudist, mudist_sq, psi2
 
-    def input_sensitivity(self):
-        if self.ARD: return 1./self.lengthscale
-        else: return (1./self.lengthscale).repeat(self.input_dim)
+    def _weave_psi2_lengthscale_grads(self, dL_dpsi2, psi2, Zdist_sq, S, mudist_sq, l2):
+
+        #here's the einsum equivalent, it's ~3 times slower
+        #return 2.*np.einsum( 'ijk,ijk,ijkl,il->l', dL_dpsi2, psi2, Zdist_sq * (2.*S[:,None,None,:]/l2 + 1.) + mudist_sq + S[:, None, None, :] / l2, 1./(2.*S + l2))*self.lengthscale
+
+        result = np.zeros(self.input_dim)
+        code = """
+        double tmp;
+        for(int q=0; q<Q; q++)
+        {
+            tmp = 0.0;
+            #pragma omp parallel for reduction(+:tmp)
+            for(int n=0; n<N; n++)
+            {
+                for(int m=0; m<M; m++)
+                {
+                    //diag terms
+                    tmp += dL_dpsi2(n,m,m) * psi2(n,m,m) * (Zdist_sq(m,m,q) * (2.0*S(n,q)/l2(q) + 1.0) + mudist_sq(n,m,m,q) + S(n,q)/l2(q)) / (2.0*S(n,q) + l2(q)) ;
+
+                    //off-diag terms
+                    for(int mm=0; mm<m; mm++)
+                    {
+                        tmp += 2.0 * dL_dpsi2(n,m,mm) * psi2(n,m,mm) * (Zdist_sq(m,mm,q) * (2.0*S(n,q)/l2(q) + 1.0) + mudist_sq(n,m,mm,q) + S(n,q)/l2(q)) / (2.0*S(n,q) + l2(q)) ;
+                    }
+                }
+            }
+            result(q) = tmp;
+        }
+
+        """
+        support_code = """
+        #include <omp.h>
+        #include <math.h>
+        """
+        N,Q = S.shape
+        M = psi2.shape[-1]
+
+        S = param_to_array(S)
+        weave.inline(code, support_code=support_code, libraries=['gomp'],
+                     arg_names=['psi2', 'dL_dpsi2', 'N', 'M', 'Q', 'mudist_sq', 'l2', 'Zdist_sq', 'S', 'result'],
+                     type_converters=weave.converters.blitz, **self.weave_options)
+
+        return 2.*result*self.lengthscale

GPy/kern/_src/rbf_psi_comp/__init__.py

@@ -0,0 +1,2 @@
+# Copyright (c) 2012, GPy authors (see AUTHORS.txt).
+# Licensed under the BSD 3-clause license (see LICENSE.txt)

GPy/kern/_src/rbf_psi_comp/ssrbf_psi_comp.py

@@ -0,0 +1,111 @@
+# Copyright (c) 2012, GPy authors (see AUTHORS.txt).
+# Licensed under the BSD 3-clause license (see LICENSE.txt)
+
+"""
+The package for the psi statistics computation
+"""
+
+import numpy as np
+
+def _Z_distances(Z):
+    Zhat = 0.5 * (Z[:, None, :] + Z[None, :, :]) # M,M,Q
+    Zdist = 0.5 * (Z[:, None, :] - Z[None, :, :]) # M,M,Q
+    return Zhat, Zdist
+
+# def _psi1computations(self, Z, vp):
+#     mu, S = vp.mean, vp.variance
+#     l2 = lengthscale **2
+#     denom = S[:, None, :] / l2 + 1. # N,1,Q
+#     dist = Z[None, :, :] - mu[:, None, :] # N,M,Q
+#     dist_sq = np.square(dist) / l2 / denom # N,M,Q
+#     exponent = -0.5 * np.sum(dist_sq + np.log(denom), -1)#N,M
+#     psi1 = self.variance * np.exp(exponent) # N,M
+#     return denom, dist, dist_sq, psi1
+
+def _psi1computations(variance, lengthscale, Z, mu, S, gamma):
+    """
+    Z - MxQ
+    mu - NxQ
+    S - NxQ
+    gamma - NxQ
+    """
+    # here are the "statistics" for psi1 and psi2
+    # Produced intermediate results:
+    # _psi1                NxM
+    # _dpsi1_dvariance     NxM
+    # _dpsi1_dlengthscale  NxMxQ
+    # _dpsi1_dZ            NxMxQ
+    # _dpsi1_dgamma        NxMxQ
+    # _dpsi1_dmu           NxMxQ
+    # _dpsi1_dS            NxMxQ
+    
+    lengthscale2 = np.square(lengthscale)
+
+    # psi1
+    _psi1_denom = S[:, None, :] / lengthscale2 + 1.  # Nx1xQ
+    _psi1_denom_sqrt = np.sqrt(_psi1_denom) #Nx1xQ
+    _psi1_dist = Z[None, :, :] - mu[:, None, :]  # NxMxQ
+    _psi1_dist_sq = np.square(_psi1_dist) / (lengthscale2 * _psi1_denom) # NxMxQ
+    _psi1_common = gamma[:,None,:] / (lengthscale2*_psi1_denom*_psi1_denom_sqrt) #Nx1xQ
+    _psi1_exponent1 = np.log(gamma[:,None,:]) -0.5 * (_psi1_dist_sq + np.log(_psi1_denom)) # NxMxQ
+    _psi1_exponent2 = np.log(1.-gamma[:,None,:]) -0.5 * (np.square(Z[None,:,:])/lengthscale2) # NxMxQ
+    _psi1_exponent = np.log(np.exp(_psi1_exponent1) + np.exp(_psi1_exponent2)) #NxMxQ
+    _psi1_exp_sum = _psi1_exponent.sum(axis=-1) #NxM
+    _psi1_exp_dist_sq = np.exp(-0.5*_psi1_dist_sq) # NxMxQ
+    _psi1_exp_Z = np.exp(-0.5*np.square(Z[None,:,:])/lengthscale2) # 1xMxQ
+    _psi1_q = variance * np.exp(_psi1_exp_sum[:,:,None] - _psi1_exponent) # NxMxQ
+    _psi1 = variance * np.exp(_psi1_exp_sum) # NxM
+    _dpsi1_dvariance = _psi1 / variance # NxM
+    _dpsi1_dgamma = _psi1_q * (_psi1_exp_dist_sq/_psi1_denom_sqrt-_psi1_exp_Z) # NxMxQ
+    _dpsi1_dmu = _psi1_q * (_psi1_exp_dist_sq * _psi1_dist * _psi1_common) # NxMxQ
+    _dpsi1_dS = _psi1_q * (_psi1_exp_dist_sq * _psi1_common * 0.5 * (_psi1_dist_sq - 1.)) # NxMxQ
+    _dpsi1_dZ = _psi1_q * (- _psi1_common * _psi1_dist * _psi1_exp_dist_sq - (1-gamma[:,None,:])/lengthscale2*Z[None,:,:]*_psi1_exp_Z) # NxMxQ
+    _dpsi1_dlengthscale = 2.*lengthscale*_psi1_q * (0.5*_psi1_common*(S[:,None,:]/lengthscale2+_psi1_dist_sq)*_psi1_exp_dist_sq + 0.5*(1-gamma[:,None,:])*np.square(Z[None,:,:]/lengthscale2)*_psi1_exp_Z) # NxMxQ
+
+    return _psi1, _dpsi1_dvariance, _dpsi1_dgamma, _dpsi1_dmu, _dpsi1_dS, _dpsi1_dZ, _dpsi1_dlengthscale
+
+def _psi2computations(variance, lengthscale, Z, mu, S, gamma):
+    """
+    Z - MxQ
+    mu - NxQ
+    S - NxQ
+    gamma - NxQ
+    """
+    # here are the "statistics" for psi1 and psi2
+    # Produced intermediate results:
+    # _psi2                NxMxM
+    # _psi2_dvariance      NxMxM
+    # _psi2_dlengthscale   NxMxMxQ
+    # _psi2_dZ             NxMxMxQ
+    # _psi2_dgamma         NxMxMxQ
+    # _psi2_dmu            NxMxMxQ
+    # _psi2_dS             NxMxMxQ
+    
+    lengthscale2 = np.square(lengthscale)
+    
+    _psi2_Zhat, _psi2_Zdist = _Z_distances(Z)
+    _psi2_Zdist_sq = np.square(_psi2_Zdist / lengthscale) # M,M,Q
+    _psi2_Z_sq_sum = (np.square(Z[:,None,:])+np.square(Z[None,:,:]))/lengthscale2 # MxMxQ
+
+    # psi2
+    _psi2_denom = 2.*S[:, None, None, :] / lengthscale2 + 1. # Nx1x1xQ
+    _psi2_denom_sqrt = np.sqrt(_psi2_denom)
+    _psi2_mudist = mu[:,None,None,:]-_psi2_Zhat #N,M,M,Q
+    _psi2_mudist_sq = np.square(_psi2_mudist)/(lengthscale2*_psi2_denom)
+    _psi2_common = gamma[:,None,None,:]/(lengthscale2 * _psi2_denom * _psi2_denom_sqrt) # Nx1x1xQ
+    _psi2_exponent1 = -_psi2_Zdist_sq -_psi2_mudist_sq -0.5*np.log(_psi2_denom)+np.log(gamma[:,None,None,:]) #N,M,M,Q
+    _psi2_exponent2 = np.log(1.-gamma[:,None,None,:]) - 0.5*(_psi2_Z_sq_sum) # NxMxMxQ
+    _psi2_exponent = np.log(np.exp(_psi2_exponent1) + np.exp(_psi2_exponent2))
+    _psi2_exp_sum = _psi2_exponent.sum(axis=-1) #NxM
+    _psi2_q = np.square(variance) * np.exp(_psi2_exp_sum[:,:,:,None]-_psi2_exponent) # NxMxMxQ 
+    _psi2_exp_dist_sq = np.exp(-_psi2_Zdist_sq -_psi2_mudist_sq) # NxMxMxQ
+    _psi2_exp_Z = np.exp(-0.5*_psi2_Z_sq_sum) # MxMxQ
+    _psi2 = np.square(variance) * np.exp(_psi2_exp_sum) # N,M,M
+    _dpsi2_dvariance = 2. * _psi2/variance # NxMxM
+    _dpsi2_dgamma = _psi2_q * (_psi2_exp_dist_sq/_psi2_denom_sqrt - _psi2_exp_Z) # NxMxMxQ
+    _dpsi2_dmu = _psi2_q * (-2.*_psi2_common*_psi2_mudist * _psi2_exp_dist_sq) # NxMxMxQ
+    _dpsi2_dS = _psi2_q * (_psi2_common * (2.*_psi2_mudist_sq - 1.) * _psi2_exp_dist_sq) # NxMxMxQ
+    _dpsi2_dZ = 2.*_psi2_q * (_psi2_common*(-_psi2_Zdist*_psi2_denom+_psi2_mudist)*_psi2_exp_dist_sq - (1-gamma[:,None,None,:])*Z[:,None,:]/lengthscale2*_psi2_exp_Z) # NxMxMxQ
+    _dpsi2_dlengthscale = 2.*lengthscale* _psi2_q * (_psi2_common*(S[:,None,None,:]/lengthscale2+_psi2_Zdist_sq*_psi2_denom+_psi2_mudist_sq)*_psi2_exp_dist_sq+(1-gamma[:,None,None,:])*_psi2_Z_sq_sum*0.5/lengthscale2*_psi2_exp_Z) # NxMxMxQ
+
+    return _psi2, _dpsi2_dvariance, _dpsi2_dgamma, _dpsi2_dmu, _dpsi2_dS, _dpsi2_dZ, _dpsi2_dlengthscale

GPy/kern/_src/ssrbf.py

@@ -7,6 +7,7 @@ import numpy as np
 from ...util.linalg import tdot
 from ...util.config import *
 from stationary import Stationary
+from rbf_psi_comp import ssrbf_psi_comp
 
 class SSRBF(Stationary):
     """
@@ -54,101 +55,63 @@ class SSRBF(Stationary):
     #             PSI statistics            #
     #---------------------------------------#
     
-    def psi0(self, Z, posterior_variational):
-        ret = np.empty(posterior_variational.mean.shape[0])
+    def psi0(self, Z, variational_posterior):
+        ret = np.empty(variational_posterior.mean.shape[0])
         ret[:] = self.variance
         return ret
 
-    def psi1(self, Z, posterior_variational):
-        self._psi_computations(Z, posterior_variational.mean, posterior_variational.variance, posterior_variational.binary_prob)
-        return self._psi1
+    def psi1(self, Z, variational_posterior):
+        _psi1, _, _, _, _, _, _ = ssrbf_psi_comp._psi1computations(self.variance, self.lengthscale, Z, variational_posterior.mean, variational_posterior.variance, variational_posterior.binary_prob)
+        return _psi1
 
-    def psi2(self, Z, posterior_variational):
-        self._psi_computations(Z, posterior_variational.mean, posterior_variational.variance, posterior_variational.binary_prob)
-        return self._psi2
+    def psi2(self, Z, variational_posterior):
+        _psi2, _, _, _, _, _, _ = ssrbf_psi_comp._psi2computations(self.variance, self.lengthscale, Z, variational_posterior.mean, variational_posterior.variance, variational_posterior.binary_prob)
+        return _psi2
 
-    def dL_dpsi0_dmuSgamma(self, dL_dpsi0, Z, mu, S, gamma, target_mu, target_S, target_gamma):
-        pass
-
-
-    def dL_dpsi1_dmuSgamma(self, dL_dpsi1, Z, mu, S, gamma, target_mu, target_S, target_gamma):
-        self._psi_computations(Z, mu, S, gamma)
-        target_mu += (dL_dpsi1[:, :, None] * self._dpsi1_dmu).sum(axis=1)
-        target_S += (dL_dpsi1[:, :, None] * self._dpsi1_dS).sum(axis=1)
-        target_gamma += (dL_dpsi1[:,:,None] * self._dpsi1_dgamma).sum(axis=1)
-
-
-    def dL_dpsi2_dmuSgamma(self, dL_dpsi2, Z, mu, S, gamma, target_mu, target_S, target_gamma):
-        """Think N,num_inducing,num_inducing,input_dim """
-        self._psi_computations(Z, mu, S, gamma)
-        target_mu += (dL_dpsi2[:, :, :, None] * self._dpsi2_dmu).reshape(mu.shape[0],-1,mu.shape[1]).sum(axis=1)
-        target_S += (dL_dpsi2[:, :, :, None] * self._dpsi2_dS).reshape(S.shape[0],-1,S.shape[1]).sum(axis=1)
-        target_gamma += (dL_dpsi2[:,:,:, None] *self._dpsi2_dgamma).reshape(gamma.shape[0],-1,gamma.shape[1]).sum(axis=1)
-
-    def update_gradients_variational(self, dL_dKmm, dL_dpsi0, dL_dpsi1, dL_dpsi2, Z, posterior_variational):
-        self._psi_computations(Z, posterior_variational.mean, posterior_variational.variance, posterior_variational.binary_prob)
+    def update_gradients_expectations(self, dL_dpsi0, dL_dpsi1, dL_dpsi2, Z, variational_posterior):
+        _, _dpsi1_dvariance, _, _, _, _, _dpsi1_dlengthscale = ssrbf_psi_comp._psi1computations(self.variance, self.lengthscale, Z, variational_posterior.mean, variational_posterior.variance, variational_posterior.binary_prob)
+        _, _dpsi2_dvariance, _, _, _, _, _dpsi2_dlengthscale = ssrbf_psi_comp._psi2computations(self.variance, self.lengthscale, Z, variational_posterior.mean, variational_posterior.variance, variational_posterior.binary_prob)
 
         #contributions from psi0:
         self.variance.gradient = np.sum(dL_dpsi0)
 
         #from psi1
-        self.variance.gradient += np.sum(dL_dpsi1 * self._dpsi1_dvariance)
-        self.lengthscale.gradient = (dL_dpsi1[:,:,None]*self._dpsi1_dlengthscale).reshape(-1,self.input_dim).sum(axis=0) 
+        self.variance.gradient += np.sum(dL_dpsi1 * _dpsi1_dvariance)
+        self.lengthscale.gradient = (dL_dpsi1[:,:,None]*_dpsi1_dlengthscale).reshape(-1,self.input_dim).sum(axis=0) 
     
 
         #from psi2
-        self.variance.gradient += (dL_dpsi2 * self._dpsi2_dvariance).sum()
-        self.lengthscale.gradient += (dL_dpsi2[:,:,:,None] * self._dpsi2_dlengthscale).reshape(-1,self.input_dim).sum(axis=0)
-
-        #from Kmm
-        self._K_computations(Z, None)
-        dvardLdK = self._K_dvar * dL_dKmm
-        var_len3 = self.variance / (np.square(self.lengthscale)*self.lengthscale)
-
-        self.variance.gradient += np.sum(dvardLdK)
-        self.lengthscale.gradient += (np.square(Z[:,None,:]-Z[None,:,:])*dvardLdK[:,:,None]).reshape(-1,self.input_dim).sum(axis=0)*var_len3
+        self.variance.gradient += (dL_dpsi2 * _dpsi2_dvariance).sum()
+        self.lengthscale.gradient += (dL_dpsi2[:,:,:,None] * _dpsi2_dlengthscale).reshape(-1,self.input_dim).sum(axis=0)        
         
-        
-    def gradients_Z_variational(self, dL_dKmm, dL_dpsi0, dL_dpsi1, dL_dpsi2, Z, posterior_variational):
-        self._psi_computations(Z, posterior_variational.mean, posterior_variational.variance, posterior_variational.binary_prob)
+    def gradients_Z_expectations(self, dL_dpsi1, dL_dpsi2, Z, variational_posterior):
+        _, _, _, _, _, _dpsi1_dZ, _ = ssrbf_psi_comp._psi1computations(self.variance, self.lengthscale, Z, variational_posterior.mean, variational_posterior.variance, variational_posterior.binary_prob)
+        _, _, _, _, _, _dpsi2_dZ, _ = ssrbf_psi_comp._psi2computations(self.variance, self.lengthscale, Z, variational_posterior.mean, variational_posterior.variance, variational_posterior.binary_prob)
 
         #psi1
-        grad = (dL_dpsi1[:, :, None] * self._dpsi1_dZ).sum(axis=0)
+        grad = (dL_dpsi1[:, :, None] * _dpsi1_dZ).sum(axis=0)
 
         #psi2
-        grad += (dL_dpsi2[:, :, :, None] * self._dpsi2_dZ).sum(axis=0).sum(axis=1)
-
-        grad += self.gradients_X(dL_dKmm, Z, None)
+        grad += (dL_dpsi2[:, :, :, None] * _dpsi2_dZ).sum(axis=0).sum(axis=1)
 
         return grad
 
-    def gradients_q_variational(self, dL_dKmm, dL_dpsi0, dL_dpsi1, dL_dpsi2, Z, posterior_variational):
-        ndata = posterior_variational.mean.shape[0]
-        self._psi_computations(Z, posterior_variational.mean, posterior_variational.variance, posterior_variational.binary_prob)
+    def gradients_qX_expectations(self, dL_dpsi0, dL_dpsi1, dL_dpsi2, Z, variational_posterior):
+        ndata = variational_posterior.mean.shape[0]
+        
+        _, _, _dpsi1_dgamma, _dpsi1_dmu, _dpsi1_dS, _, _ = ssrbf_psi_comp._psi1computations(self.variance, self.lengthscale, Z, variational_posterior.mean, variational_posterior.variance, variational_posterior.binary_prob)
+        _, _, _dpsi2_dgamma, _dpsi2_dmu, _dpsi2_dS, _, _ = ssrbf_psi_comp._psi2computations(self.variance, self.lengthscale, Z, variational_posterior.mean, variational_posterior.variance, variational_posterior.binary_prob)
+
         #psi1
-        grad_mu = (dL_dpsi1[:, :, None] * self._dpsi1_dmu).sum(axis=1)
-        grad_S = (dL_dpsi1[:, :, None] * self._dpsi1_dS).sum(axis=1)
-        grad_gamma = (dL_dpsi1[:,:,None] * self._dpsi1_dgamma).sum(axis=1)
+        grad_mu = (dL_dpsi1[:, :, None] * _dpsi1_dmu).sum(axis=1)
+        grad_S = (dL_dpsi1[:, :, None] * _dpsi1_dS).sum(axis=1)
+        grad_gamma = (dL_dpsi1[:,:,None] * _dpsi1_dgamma).sum(axis=1)
         #psi2
-        grad_mu += (dL_dpsi2[:, :, :, None] * self._dpsi2_dmu).reshape(ndata,-1,self.input_dim).sum(axis=1)
-        grad_S += (dL_dpsi2[:, :, :, None] * self._dpsi2_dS).reshape(ndata,-1,self.input_dim).sum(axis=1)
-        grad_gamma += (dL_dpsi2[:,:,:, None] *self._dpsi2_dgamma).reshape(ndata,-1,self.input_dim).sum(axis=1)
+        grad_mu += (dL_dpsi2[:, :, :, None] * _dpsi2_dmu).reshape(ndata,-1,self.input_dim).sum(axis=1)
+        grad_S += (dL_dpsi2[:, :, :, None] * _dpsi2_dS).reshape(ndata,-1,self.input_dim).sum(axis=1)
+        grad_gamma += (dL_dpsi2[:,:,:, None] * _dpsi2_dgamma).reshape(ndata,-1,self.input_dim).sum(axis=1)
         
         return grad_mu, grad_S, grad_gamma
-
-    def gradients_X(self, dL_dK, X, X2=None):
-        #if self._X is None or X.base is not self._X.base or X2 is not None:
-        if X2==None:
-            _K_dist = X[:,None,:] - X[None,:,:]
-            _K_dist2 = np.square(_K_dist/self.lengthscale).sum(axis=-1)
-            dK_dX = self.variance*np.exp(-0.5 * self._K_dist2[:,:,None]) * (-2.*_K_dist/np.square(self.lengthscale))
-            dL_dX = (dL_dK[:,:,None] * dK_dX).sum(axis=1)
-        else:
-            _K_dist = X[:,None,:] - X2[None,:,:]
-            _K_dist2 = np.square(_K_dist/self.lengthscale).sum(axis=-1)
-            dK_dX = self.variance*np.exp(-0.5 * self._K_dist2[:,:,None]) * (-_K_dist/np.square(self.lengthscale))
-            dL_dX = (dL_dK[:,:,None] * dK_dX).sum(axis=1)
-        return dL_dX
         
     #---------------------------------------#
     #            Precomputations            #
@@ -174,78 +137,3 @@ class SSRBF(Stationary):
             self._K_dist2 = -2.*np.dot(X, X2.T) + (np.sum(np.square(X), axis=1)[:, None] + np.sum(np.square(X2), axis=1)[None, :])
         self._K_dvar = np.exp(-0.5 * self._K_dist2)
 
-    #@cache_this(1)
-    def _psi_computations(self, Z, mu, S, gamma):
-        """
-        Z - MxQ
-        mu - NxQ
-        S - NxQ
-        gamma - NxQ
-        """
-        # here are the "statistics" for psi1 and psi2
-        # Produced intermediate results:
-        # _psi1                NxM
-        # _dpsi1_dvariance     NxM
-        # _dpsi1_dlengthscale  NxMxQ
-        # _dpsi1_dZ            NxMxQ
-        # _dpsi1_dgamma        NxMxQ
-        # _dpsi1_dmu           NxMxQ
-        # _dpsi1_dS            NxMxQ
-        # _psi2                NxMxM
-        # _psi2_dvariance      NxMxM
-        # _psi2_dlengthscale   NxMxMxQ
-        # _psi2_dZ             NxMxMxQ
-        # _psi2_dgamma         NxMxMxQ
-        # _psi2_dmu            NxMxMxQ
-        # _psi2_dS             NxMxMxQ
-        
-        lengthscale2 = np.square(self.lengthscale)
-                    
-        _psi2_Zhat = 0.5 * (Z[:, None, :] + Z[None, :, :]) # M,M,Q
-        _psi2_Zdist = 0.5 * (Z[:, None, :] - Z[None, :, :]) # M,M,Q
-        _psi2_Zdist_sq = np.square(_psi2_Zdist / self.lengthscale) # M,M,Q
-        _psi2_Z_sq_sum = (np.square(Z[:,None,:])+np.square(Z[None,:,:]))/lengthscale2 # MxMxQ
-
-        # psi1
-        _psi1_denom = S[:, None, :] / lengthscale2 + 1.  # Nx1xQ
-        _psi1_denom_sqrt = np.sqrt(_psi1_denom) #Nx1xQ
-        _psi1_dist = Z[None, :, :] - mu[:, None, :]  # NxMxQ
-        _psi1_dist_sq = np.square(_psi1_dist) / (lengthscale2 * _psi1_denom) # NxMxQ
-        _psi1_common = gamma[:,None,:] / (lengthscale2*_psi1_denom*_psi1_denom_sqrt) #Nx1xQ
-        _psi1_exponent1 = np.log(gamma[:,None,:]) -0.5 * (_psi1_dist_sq + np.log(_psi1_denom)) # NxMxQ
-        _psi1_exponent2 = np.log(1.-gamma[:,None,:]) -0.5 * (np.square(Z[None,:,:])/lengthscale2) # NxMxQ
-        _psi1_exponent = np.log(np.exp(_psi1_exponent1) + np.exp(_psi1_exponent2)) #NxMxQ
-        _psi1_exp_sum = _psi1_exponent.sum(axis=-1) #NxM
-        _psi1_exp_dist_sq = np.exp(-0.5*_psi1_dist_sq) # NxMxQ
-        _psi1_exp_Z = np.exp(-0.5*np.square(Z[None,:,:])/lengthscale2) # 1xMxQ
-        _psi1_q = self.variance * np.exp(_psi1_exp_sum[:,:,None] - _psi1_exponent) # NxMxQ
-        self._psi1 = self.variance * np.exp(_psi1_exp_sum) # NxM
-        self._dpsi1_dvariance = self._psi1 / self.variance # NxM
-        self._dpsi1_dgamma = _psi1_q * (_psi1_exp_dist_sq/_psi1_denom_sqrt-_psi1_exp_Z) # NxMxQ
-        self._dpsi1_dmu = _psi1_q * (_psi1_exp_dist_sq * _psi1_dist * _psi1_common) # NxMxQ
-        self._dpsi1_dS = _psi1_q * (_psi1_exp_dist_sq * _psi1_common * 0.5 * (_psi1_dist_sq - 1.)) # NxMxQ
-        self._dpsi1_dZ = _psi1_q * (- _psi1_common * _psi1_dist * _psi1_exp_dist_sq - (1-gamma[:,None,:])/lengthscale2*Z[None,:,:]*_psi1_exp_Z) # NxMxQ
-        self._dpsi1_dlengthscale = 2.*self.lengthscale*_psi1_q * (0.5*_psi1_common*(S[:,None,:]/lengthscale2+_psi1_dist_sq)*_psi1_exp_dist_sq + 0.5*(1-gamma[:,None,:])*np.square(Z[None,:,:]/lengthscale2)*_psi1_exp_Z) # NxMxQ
-
-
-        # psi2
-        _psi2_denom = 2.*S[:, None, None, :] / lengthscale2 + 1. # Nx1x1xQ
-        _psi2_denom_sqrt = np.sqrt(_psi2_denom)
-        _psi2_mudist = mu[:,None,None,:]-_psi2_Zhat #N,M,M,Q
-        _psi2_mudist_sq = np.square(_psi2_mudist)/(lengthscale2*_psi2_denom)
-        _psi2_common = gamma[:,None,None,:]/(lengthscale2 * _psi2_denom * _psi2_denom_sqrt) # Nx1x1xQ
-        _psi2_exponent1 = -_psi2_Zdist_sq -_psi2_mudist_sq -0.5*np.log(_psi2_denom)+np.log(gamma[:,None,None,:]) #N,M,M,Q
-        _psi2_exponent2 = np.log(1.-gamma[:,None,None,:]) - 0.5*(_psi2_Z_sq_sum) # NxMxMxQ
-        _psi2_exponent = np.log(np.exp(_psi2_exponent1) + np.exp(_psi2_exponent2))
-        _psi2_exp_sum = _psi2_exponent.sum(axis=-1) #NxM
-        _psi2_q = np.square(self.variance) * np.exp(_psi2_exp_sum[:,:,:,None]-_psi2_exponent) # NxMxMxQ 
-        _psi2_exp_dist_sq = np.exp(-_psi2_Zdist_sq -_psi2_mudist_sq) # NxMxMxQ
-        _psi2_exp_Z = np.exp(-0.5*_psi2_Z_sq_sum) # MxMxQ
-        self._psi2 = np.square(self.variance) * np.exp(_psi2_exp_sum) # N,M,M
-        self._dpsi2_dvariance = 2. * self._psi2/self.variance # NxMxM
-        self._dpsi2_dgamma = _psi2_q * (_psi2_exp_dist_sq/_psi2_denom_sqrt - _psi2_exp_Z) # NxMxMxQ
-        self._dpsi2_dmu = _psi2_q * (-2.*_psi2_common*_psi2_mudist * _psi2_exp_dist_sq) # NxMxMxQ
-        self._dpsi2_dS = _psi2_q * (_psi2_common * (2.*_psi2_mudist_sq - 1.) * _psi2_exp_dist_sq) # NxMxMxQ
-        self._dpsi2_dZ = 2.*_psi2_q * (_psi2_common*(-_psi2_Zdist*_psi2_denom+_psi2_mudist)*_psi2_exp_dist_sq - (1-gamma[:,None,None,:])*Z[:,None,:]/lengthscale2*_psi2_exp_Z) # NxMxMxQ
-        self._dpsi2_dlengthscale = 2.*self.lengthscale* _psi2_q * (_psi2_common*(S[:,None,None,:]/lengthscale2+_psi2_Zdist_sq*_psi2_denom+_psi2_mudist_sq)*_psi2_exp_dist_sq+(1-gamma[:,None,None,:])*_psi2_Z_sq_sum*0.5/lengthscale2*_psi2_exp_Z) # NxMxMxQ
-        

GPy/kern/_src/stationary.py

@@ -12,6 +12,35 @@ from scipy import integrate
 from ...util.caching import Cache_this
 
 class Stationary(Kern):
+    """
+    Stationary kernels (covariance functions).
+
+    Stationary covariance fucntion depend only on r, where r is defined as 
+
+      r = \sqrt{ \sum_{q=1}^Q (x_q - x'_q)^2 }
+
+    The covariance function k(x, x' can then be written k(r). 
+
+    In this implementation, r is scaled by the lengthscales parameter(s):
+
+      r = \sqrt{ \sum_{q=1}^Q \frac{(x_q - x'_q)^2}{\ell_q^2} }. 
+    
+    By default, there's only one lengthscale: seaprate lengthscales for each
+    dimension can be enables by setting ARD=True. 
+
+    To implement a stationary covariance function using this class, one need
+    only define the covariance function k(r), and it derivative. 
+
+      ...
+      def K_of_r(self, r):
+          return foo
+      def dK_dr(self, r):
+          return bar
+
+    The lengthscale(s) and variance parameters are added to the structure automatically. 
+          
+    """
+    
     def __init__(self, input_dim, variance, lengthscale, ARD, name):
         super(Stationary, self).__init__(input_dim, name)
         self.ARD = ARD
@@ -20,11 +49,11 @@ class Stationary(Kern):
                 lengthscale = np.ones(1)
             else:
                 lengthscale = np.asarray(lengthscale)
-                assert lengthscale.size == 1, "Only  lengthscale needed for non-ARD kernel"
+                assert lengthscale.size == 1, "Only 1 lengthscale needed for non-ARD kernel"
         else:
             if lengthscale is not None:
                 lengthscale = np.asarray(lengthscale)
-                assert lengthscale.size in [1, input_dim], "Bad lengthscales"
+                assert lengthscale.size in [1, input_dim], "Bad number of lengthscales"
                 if lengthscale.size != input_dim:
                     lengthscale = np.ones(input_dim)*lengthscale
             else:
@@ -35,31 +64,32 @@ class Stationary(Kern):
         self.add_parameters(self.variance, self.lengthscale)
 
     def K_of_r(self, r):
-        raise NotImplementedError, "implement the covaraiance function as a fn of r to use this class"
+        raise NotImplementedError, "implement the covariance function as a fn of r to use this class"
 
     def dK_dr(self, r):
-        raise NotImplementedError, "implement the covaraiance function as a fn of r to use this class"
+        raise NotImplementedError, "implement derivative of the covariance function wrt r to use this class"
 
     @Cache_this(limit=5, ignore_args=())
     def K(self, X, X2=None):
         r = self._scaled_dist(X, X2)
         return self.K_of_r(r)
 
-    @Cache_this(limit=5, ignore_args=(0,))
-    def _dist(self, X, X2):
-        if X2 is None:
-            X2 = X
-        return X[:, None, :] - X2[None, :, :]
+    @Cache_this(limit=3, ignore_args=())
+    def dK_dr_via_X(self, X, X2):
+        #a convenience function, so we can cache dK_dr
+        return self.dK_dr(self._scaled_dist(X, X2))
 
     @Cache_this(limit=5, ignore_args=(0,))
     def _unscaled_dist(self, X, X2=None):
         """
-        Compute the square distance between each row of X and X2, or between
+        Compute the Euclidean distance between each row of X and X2, or between
         each pair of rows of X if X2 is None.
         """
         if X2 is None:
             Xsq = np.sum(np.square(X),1)
-            return np.sqrt(-2.*tdot(X) + (Xsq[:,None] + Xsq[None,:]))
+            r2 = -2.*tdot(X) + (Xsq[:,None] + Xsq[None,:])
+            util.diag.view(r2)[:,]= 0. # force diagnoal to be zero: sometime numerically a little negative
+            return np.sqrt(r2)
         else:
             X1sq = np.sum(np.square(X),1)
             X2sq = np.sum(np.square(X2),1)
@@ -70,7 +100,7 @@ class Stationary(Kern):
         """
         Efficiently compute the scaled distance, r.
 
-        r = \sum_{q=1}^Q (x_q - x'q)^2/l_q^2
+        r = \sqrt( \sum_{q=1}^Q (x_q - x'q)^2/l_q^2 )
 
         Note that if thre is only one lengthscale, l comes outside the sum. In
         this case we compute the unscaled distance first (in a separate
@@ -84,7 +114,6 @@ class Stationary(Kern):
         else:
             return self._unscaled_dist(X, X2)/self.lengthscale
 
-
     def Kdiag(self, X):
         ret = np.empty(X.shape[0])
         ret[:] = self.variance
@@ -95,20 +124,23 @@ class Stationary(Kern):
         self.lengthscale.gradient = 0.
 
     def update_gradients_full(self, dL_dK, X, X2=None):
-        r = self._scaled_dist(X, X2)
-        K = self.K_of_r(r)
 
-        rinv = self._inv_dist(X, X2)
-        dL_dr = self.dK_dr(r) * dL_dK
+        self.variance.gradient = np.einsum('ij,ij,i', self.K(X, X2), dL_dK, 1./self.variance)
 
+        #now the lengthscale gradient(s)
+        dL_dr = self.dK_dr_via_X(X, X2) * dL_dK
         if self.ARD:
-            x_xl3 = np.square(self._dist(X, X2)) / self.lengthscale**3
-            self.lengthscale.gradient = -((dL_dr*rinv)[:,:,None]*x_xl3).sum(0).sum(0)
+            #rinv = self._inv_dis# this is rather high memory? Should we loop instead?t(X, X2)
+            #d =  X[:, None, :] - X2[None, :, :]
+            #x_xl3 = np.square(d) 
+            #self.lengthscale.gradient = -((dL_dr*rinv)[:,:,None]*x_xl3).sum(0).sum(0)/self.lengthscale**3
+            tmp = dL_dr*self._inv_dist(X, X2)
+            if X2 is None: X2 = X
+            self.lengthscale.gradient = np.array([np.einsum('ij,ij,...', tmp, np.square(X[:,q:q+1] - X2[:,q:q+1].T), -1./self.lengthscale[q]**3) for q in xrange(self.input_dim)])
         else:
-            x_xl3 = np.square(self._dist(X, X2)) / self.lengthscale**3
-            self.lengthscale.gradient = -((dL_dr*rinv)[:,:,None]*x_xl3).sum()
+            r = self._scaled_dist(X, X2)
+            self.lengthscale.gradient = -np.sum(dL_dr*r)/self.lengthscale
 
-        self.variance.gradient = np.sum(K * dL_dK)/self.variance
 
     def _inv_dist(self, X, X2=None):
         """
@@ -116,7 +148,7 @@ class Stationary(Kern):
         diagonal, where we return zero (the distance on the diagonal is zero).
         This term appears in derviatives.
         """
-        dist = self._scaled_dist(X, X2)
+        dist = self._scaled_dist(X, X2).copy()
         if X2 is None:
             nondiag = util.diag.offdiag_view(dist)
             nondiag[:] = 1./nondiag
@@ -128,10 +160,11 @@ class Stationary(Kern):
         """
         Given the derivative of the objective wrt K (dL_dK), compute the derivative wrt X
         """
-        r = self._scaled_dist(X, X2)
         invdist = self._inv_dist(X, X2)
-        dL_dr = self.dK_dr(r) * dL_dK
-        #The high-memory numpy way: ret = np.sum((invdist*dL_dr)[:,:,None]*self._dist(X, X2),1)/self.lengthscale**2
+        dL_dr = self.dK_dr_via_X(X, X2) * dL_dK
+        #The high-memory numpy way:
+        #d =  X[:, None, :] - X2[None, :, :]
+        #ret = np.sum((invdist*dL_dr)[:,:,None]*d,1)/self.lengthscale**2
         #if X2 is None:
             #ret *= 2.
 
@@ -141,7 +174,7 @@ class Stationary(Kern):
             tmp *= 2.
             X2 = X
         ret = np.empty(X.shape, dtype=np.float64)
-        [np.copyto(ret[:,q], np.sum(tmp*(X[:,q][:,None]-X2[:,q][None,:]), 1)) for q in xrange(self.input_dim)]
+        [np.einsum('ij,ij->i', tmp, X[:,q][:,None]-X2[:,q][None,:], out=ret[:,q]) for q in xrange(self.input_dim)]
         ret /= self.lengthscale**2
 
         return ret
@@ -214,7 +247,7 @@ class Matern52(Stationary):
 
     .. math::
 
-       k(r) = \sigma^2 (1 + \sqrt{5} r + \\frac53 r^2) \exp(- \sqrt{5} r) \ \ \ \ \  \\text{ where  } r = \sqrt{\sum_{i=1}^input_dim \\frac{(x_i-y_i)^2}{\ell_i^2} }
+       k(r) = \sigma^2 (1 + \sqrt{5} r + \\frac53 r^2) \exp(- \sqrt{5} r) 
        """
     def __init__(self, input_dim, variance=1., lengthscale=None, ARD=False, name='Mat52'):
         super(Matern52, self).__init__(input_dim, variance, lengthscale, ARD, name)
@@ -225,7 +258,7 @@ class Matern52(Stationary):
     def dK_dr(self, r):
         return self.variance*(10./3*r -5.*r -5.*np.sqrt(5.)/3*r**2)*np.exp(-np.sqrt(5.)*r)
 
-    def Gram_matrix(self,F,F1,F2,F3,lower,upper):
+    def Gram_matrix(self, F, F1, F2, F3, lower, upper):
         """
         Return the Gram matrix of the vector of functions F with respect to the RKHS norm. The use of this function is limited to input_dim=1.
 

GPy/kern/_src/sympykern.py

@@ -76,34 +76,32 @@ class Sympykern(Kern):
             
             self.num_split_params = len(self._sp_theta_i)
             self._split_theta_names = ["%s"%theta.name[:-2] for theta in self._sp_theta_i]
+            # Add split parameters to the model.
             for theta in self._split_theta_names:
+                # TODO: what if user has passed a parameter vector, how should that be stored and interpreted?
                 setattr(self, theta, Param(theta, np.ones(self.output_dim), None))
-                self.add_parameters(getattr(self, theta))
+                self.add_parameter(getattr(self, theta))
 
-                #setattr(self, theta, np.ones(self.output_dim))
             
             self.num_shared_params = len(self._sp_theta)
             for theta_i, theta_j in zip(self._sp_theta_i, self._sp_theta_j):
                 self._sp_kdiag = self._sp_kdiag.subs(theta_j, theta_i)
-            #self.num_params = self.num_shared_params+self.num_split_params*self.output_dim
             
         else:
             self.num_split_params = 0
             self._split_theta_names = []
             self._sp_theta = thetas
             self.num_shared_params = len(self._sp_theta)
-            #self.num_params = self.num_shared_params
 
         # Add parameters to the model.
         for theta in self._sp_theta:
             val = 1.0
+            # TODO: what if user has passed a parameter vector, how should that be stored and interpreted? This is the old way before params class.
             if param is not None:
                 if param.has_key(theta):
                     val = param[theta]
             setattr(self, theta.name, Param(theta.name, val, None))
             self.add_parameters(getattr(self, theta.name))
-        #deal with param            
-        #self._set_params(self._get_params())
 
         # Differentiate with respect to parameters.
         derivative_arguments = self._sp_x + self._sp_theta
@@ -113,7 +111,6 @@ class Sympykern(Kern):
         self.derivatives = {theta.name : sp.diff(self._sp_k,theta).simplify() for theta in derivative_arguments}
         self.diag_derivatives = {theta.name : sp.diff(self._sp_kdiag,theta).simplify() for theta in derivative_arguments}
         
-        
         # This gives the parameters for the arg list.
         self.arg_list = self._sp_x + self._sp_z + self._sp_theta
         self.diag_arg_list = self._sp_x + self._sp_theta
@@ -134,7 +131,7 @@ class Sympykern(Kern):
         return spkern(self._sp_k+other._sp_k)
 
     def _gen_code(self):
-
+        #fn_theano = theano_function([self.arg_lists], [self._sp_k + self.derivatives], dims={x: 1}, dtypes={x_0: 'float64', z_0: 'float64'})
         self._K_function = lambdify(self.arg_list, self._sp_k, 'numpy')
         for key in self.derivatives.keys():
             setattr(self, '_K_diff_' + key, lambdify(self.arg_list, self.derivatives[key], 'numpy'))

GPy/models/bayesian_gplvm.py

@@ -150,37 +150,6 @@ class BayesianGPLVM(SparseGP):
 
         return dim_reduction_plots.plot_steepest_gradient_map(self,*args,**kwargs)
 
-class BayesianGPLVMWithMissingData(BayesianGPLVM):
-    def __init__(self, Y, input_dim, X=None, X_variance=None, init='PCA', num_inducing=10,
-        Z=None, kernel=None, inference_method=None, likelihood=None, name='bayesian gplvm', **kwargs):
-        from ..util.subarray_and_sorting import common_subarrays
-        self.subarrays = common_subarrays(Y)
-        import ipdb;ipdb.set_trace()
-        BayesianGPLVM.__init__(self, Y, input_dim, X=X, X_variance=X_variance, init=init, num_inducing=num_inducing, Z=Z, kernel=kernel, inference_method=inference_method, likelihood=likelihood, name=name, **kwargs)
-
-
-    def parameters_changed(self):
-        super(BayesianGPLVM, self).parameters_changed()
-        self._log_marginal_likelihood -= self.KL_divergence()
-
-        dL_dmu, dL_dS = self.dL_dmuS()
-
-        # dL:
-        self.X.mean.gradient  = dL_dmu
-        self.X.variance.gradient  = dL_dS
-
-        # dKL:
-        self.X.mean.gradient -= self.X.mean
-        self.X.variance.gradient -= (1. - (1. / (self.X.variance))) * 0.5
-
-if __name__ == '__main__':
-    import numpy as np
-    X = np.random.randn(20,2)
-    W = np.linspace(0,1,10)[None,:]
-    Y = (X*W).sum(1)
-    missing = np.random.binomial(1,.1,size=Y.shape)
-
-    pass
 
 def latent_cost_and_grad(mu_S, kern, Z, dL_dpsi0, dL_dpsi1, dL_dpsi2):
     """

GPy/models/sparse_gp_regression.py

@@ -8,7 +8,7 @@ from .. import likelihoods
 from .. import kern
 from ..inference.latent_function_inference import VarDTC
 from ..util.misc import param_to_array
-from ..core.parameterization.variational import VariationalPosterior
+from ..core.parameterization.variational import NormalPosterior
 
 class SparseGPRegression(SparseGP):
     """
@@ -47,7 +47,7 @@ class SparseGPRegression(SparseGP):
         likelihood = likelihoods.Gaussian()
         
         if not (X_variance is None):
-            X = VariationalPosterior(X,X_variance)
+            X = NormalPosterior(X,X_variance)
         
         SparseGP.__init__(self, X, Y, Z, kernel, likelihood, inference_method=VarDTC())
 
@@ -88,7 +88,7 @@ class SparseGPRegressionUncertainInput(SparseGP):
 
         # kern defaults to rbf (plus white for stability)
         if kernel is None:
-            kernel = kern.rbf(input_dim)  + kern.white(input_dim, variance=1e-3)
+            kernel = kern.RBF(input_dim) + kern.White(input_dim, variance=1e-3)
 
         # Z defaults to a subset of the data
         if Z is None:
@@ -99,5 +99,5 @@ class SparseGPRegressionUncertainInput(SparseGP):
 
         likelihood = likelihoods.Gaussian()
 
-        SparseGP.__init__(self, X, Y, Z, kernel, likelihood, X_variance=X_variance)
+        SparseGP.__init__(self, X, Y, Z, kernel, likelihood, X_variance=X_variance, inference_method=VarDTC())
         self.ensure_default_constraints()

GPy/models/ss_gplvm.py

@@ -58,7 +58,7 @@ class SSGPLVM(SparseGP):
         super(SSGPLVM, self).parameters_changed()
         self._log_marginal_likelihood -= self.variational_prior.KL_divergence(self.X)
 
-        self.X.mean.gradient, self.X.variance.gradient, self.X.binary_prob.gradient = self.kern.gradients_q_variational(posterior_variational=self.X, Z=self.Z, **self.grad_dict)
+        self.X.mean.gradient, self.X.variance.gradient, self.X.binary_prob.gradient = self.kern.gradients_qX_expectations(variational_posterior=self.X, Z=self.Z, **self.grad_dict)
 
         # update for the KL divergence
         self.variational_prior.update_gradients_KL(self.X)

GPy/plotting/matplot_dep/kernel_plots.py

@@ -106,7 +106,7 @@ def plot(kernel, x=None, plot_limits=None, which_parts='all', resolution=None, *
             raise ValueError, "Bad limits for plotting"
 
         Xnew = np.linspace(xmin, xmax, resolution or 201)[:, None]
-        Kx = kernel.K(Xnew, x, which_parts)
+        Kx = kernel.K(Xnew, x)
         pb.plot(Xnew, Kx, *args, **kwargs)
         pb.xlim(xmin, xmax)
         pb.xlabel("x")

GPy/plotting/matplot_dep/models_plots.py

@@ -56,10 +56,13 @@ def plot_fit(model, plot_limits=None, which_data_rows='all',
     if ax is None:
         fig = pb.figure(num=fignum)
         ax = fig.add_subplot(111)
-
-    X, Y = param_to_array(model.X, model.Y)
-    if hasattr(model, 'has_uncertain_inputs') and model.has_uncertain_inputs(): X_variance = model.X_variance
-
+    
+    if hasattr(model, 'has_uncertain_inputs') and model.has_uncertain_inputs(): 
+        X = model.X.mean
+        X_variance = param_to_array(model.X.variance)
+    else:
+        X = model.X
+    X, Y = param_to_array(X, model.Y)
     if hasattr(model, 'Z'): Z = param_to_array(model.Z)
 
     #work out what the inputs are for plotting (1D or 2D)
@@ -98,10 +101,10 @@ def plot_fit(model, plot_limits=None, which_data_rows='all',
 
 
         #add error bars for uncertain (if input uncertainty is being modelled)
-        #if hasattr(model,"has_uncertain_inputs") and model.has_uncertain_inputs():
-        #    ax.errorbar(X[which_data_rows, free_dims].flatten(), Y[which_data_rows, which_data_ycols].flatten(),
-        #                xerr=2 * np.sqrt(X_variance[which_data_rows, free_dims].flatten()),
-        #                ecolor='k', fmt=None, elinewidth=.5, alpha=.5)
+        if hasattr(model,"has_uncertain_inputs") and model.has_uncertain_inputs():
+            ax.errorbar(X[which_data_rows, free_dims].flatten(), Y[which_data_rows, which_data_ycols].flatten(),
+                        xerr=2 * np.sqrt(X_variance[which_data_rows, free_dims].flatten()),
+                        ecolor='k', fmt=None, elinewidth=.5, alpha=.5)
 
 
         #set the limits of the plot to some sensible values

GPy/testing/kernel_tests.py

@@ -8,13 +8,6 @@ import sys
 
 verbose = True
 
-try:
-    import sympy
-    SYMPY_AVAILABLE=True
-except ImportError:
-    SYMPY_AVAILABLE=False
-
-
 class Kern_check_model(GPy.core.Model):
     """
     This is a dummy model class used as a base class for checking that the
@@ -70,14 +63,11 @@ class Kern_check_dKdiag_dtheta(Kern_check_model):
         Kern_check_model.__init__(self,kernel=kernel,dL_dK=dL_dK, X=X, X2=None)
         self.add_parameter(self.kernel)
 
-    def parameters_changed(self):
-        self.kernel.update_gradients_diag(self.dL_dK, self.X)
-
     def log_likelihood(self):
         return (np.diag(self.dL_dK)*self.kernel.Kdiag(self.X)).sum()
 
     def parameters_changed(self):
-        return self.kernel.update_gradients_diag(np.diag(self.dL_dK), self.X)
+        self.kernel.update_gradients_diag(np.diag(self.dL_dK), self.X)
 
 class Kern_check_dK_dX(Kern_check_model):
     """This class allows gradient checks for the gradient of a kernel with respect to X. """
@@ -99,6 +89,8 @@ class Kern_check_dKdiag_dX(Kern_check_dK_dX):
     def parameters_changed(self):
         self.X.gradient =  self.kernel.gradients_X_diag(self.dL_dK, self.X)
 
+
+
 def kern_test(kern, X=None, X2=None, output_ind=None, verbose=False):
     """
     This function runs on kernels to check the correctness of their
@@ -217,11 +209,15 @@ def kern_test(kern, X=None, X2=None, output_ind=None, verbose=False):
     return pass_checks
 
 
+
 class KernelTestsContinuous(unittest.TestCase):
     def setUp(self):
         self.X = np.random.randn(100,2)
         self.X2 = np.random.randn(110,2)
 
+        continuous_kerns = ['RBF', 'Linear']
+        self.kernclasses = [getattr(GPy.kern, s) for s in continuous_kerns]
+
     def test_Matern32(self):
         k = GPy.kern.Matern32(2)
         self.assertTrue(kern_test(k, X=self.X, X2=self.X2, verbose=verbose))
@@ -234,6 +230,7 @@ class KernelTestsContinuous(unittest.TestCase):
 
 
 
+
 if __name__ == "__main__":
     print "Running unit tests, please be (very) patient..."
     unittest.main()

GPy/testing/observable_tests.py

@@ -0,0 +1,133 @@
+'''
+Created on 27 Feb 2014
+
+@author: maxz
+'''
+import unittest
+from GPy.core.parameterization.parameterized import Parameterized
+from GPy.core.parameterization.param import Param
+import numpy
+
+
+class ParamTestParent(Parameterized):
+    parent_changed_count = 0
+    def parameters_changed(self):
+        self.parent_changed_count += 1
+
+class ParameterizedTest(Parameterized):
+    params_changed_count = 0
+    def parameters_changed(self):
+        self.params_changed_count += 1
+    def _set_params(self, params, trigger_parent=True):
+        Parameterized._set_params(self, params, trigger_parent=trigger_parent)
+
+class Test(unittest.TestCase):
+    
+    def setUp(self):
+        self.parent = ParamTestParent('test parent')
+        self.par = ParameterizedTest('test model')
+        self.par2 = ParameterizedTest('test model 2')
+        self.p = Param('test parameter', numpy.random.normal(1,2,(10,3)))
+
+        self.par.add_parameter(self.p)
+        self.par.add_parameter(Param('test1', numpy.random.normal(0,1,(1,))))
+        self.par.add_parameter(Param('test2', numpy.random.normal(0,1,(1,))))
+
+        self.par2.add_parameter(Param('par2 test1', numpy.random.normal(0,1,(1,))))
+        self.par2.add_parameter(Param('par2 test2', numpy.random.normal(0,1,(1,))))
+
+        self.parent.add_parameter(self.par)
+        self.parent.add_parameter(self.par2)
+        
+        self._observer_triggered = None
+        self._trigger_count = 0
+        self._first = None
+        self._second = None
+        
+    def _trigger(self, which):
+        self._observer_triggered = float(which)
+        self._trigger_count += 1
+        if self._first is not None:
+            self._second = self._trigger
+        else:
+            self._first = self._trigger
+    
+    def _trigger_priority(self, which):
+        if self._first is not None:
+            self._second = self._trigger_priority
+        else:
+            self._first = self._trigger_priority
+            
+    def test_observable(self):
+        self.par.add_observer(self, self._trigger, -1)
+        self.assertEqual(self.par.params_changed_count, 0, 'no params changed yet')
+        self.assertEqual(self.par.params_changed_count, self.parent.parent_changed_count, 'parent should be triggered as often as param')
+        
+        self.p[0,1] = 3 # trigger observers
+        self.assertEqual(self._observer_triggered, 3, 'observer should have triggered')
+        self.assertEqual(self._trigger_count, 1, 'observer should have triggered once')
+        self.assertEqual(self.par.params_changed_count, 1, 'params changed once')
+        self.assertEqual(self.par.params_changed_count, self.parent.parent_changed_count, 'parent should be triggered as often as param')
+
+        self.par.remove_observer(self)
+        self.p[2,1] = 4
+        self.assertEqual(self._observer_triggered, 3, 'observer should not have triggered')
+        self.assertEqual(self._trigger_count, 1, 'observer should have triggered once')
+        self.assertEqual(self.par.params_changed_count, 2, 'params changed second')
+        self.assertEqual(self.par.params_changed_count, self.parent.parent_changed_count, 'parent should be triggered as often as param')
+        
+        self.par.add_observer(self, self._trigger, -1)
+        self.p[2,1] = 4
+        self.assertEqual(self._observer_triggered, 4, 'observer should have triggered')
+        self.assertEqual(self._trigger_count, 2, 'observer should have triggered once')
+        self.assertEqual(self.par.params_changed_count, 3, 'params changed second')
+        self.assertEqual(self.par.params_changed_count, self.parent.parent_changed_count, 'parent should be triggered as often as param')
+        
+        self.par.remove_observer(self, self._trigger)
+        self.p[0,1] = 3
+        self.assertEqual(self._observer_triggered, 4, 'observer should not have triggered')
+        self.assertEqual(self._trigger_count, 2, 'observer should have triggered once')
+        self.assertEqual(self.par.params_changed_count, 4, 'params changed second')
+        self.assertEqual(self.par.params_changed_count, self.parent.parent_changed_count, 'parent should be triggered as often as param')
+
+    def test_set_params(self):
+        self.assertEqual(self.par.params_changed_count, 0, 'no params changed yet')
+        self.par._set_params(numpy.ones(self.par.size))
+        self.assertEqual(self.par.params_changed_count, 1, 'now params changed')
+        self.assertEqual(self.parent.parent_changed_count, self.par.params_changed_count)
+        
+        self.parent._set_params(numpy.ones(self.parent.size) * 2)
+        self.assertEqual(self.par.params_changed_count, 2, 'now params changed')
+        self.assertEqual(self.parent.parent_changed_count, self.par.params_changed_count)
+
+
+    def test_priority_notify(self):
+        self.assertEqual(self.par.params_changed_count, 0)
+        self.par._notify_observers(0, None)
+        self.assertEqual(self.par.params_changed_count, 1)
+        self.assertEqual(self.parent.parent_changed_count, self.par.params_changed_count)
+
+        self.par._notify_observers(0, -numpy.inf)
+        self.assertEqual(self.par.params_changed_count, 2)
+        self.assertEqual(self.parent.parent_changed_count, 1)
+
+    def test_priority(self):
+        self.par.add_observer(self, self._trigger, -1)
+        self.par.add_observer(self, self._trigger_priority, 0)
+        self.par._notify_observers(0)
+        self.assertEqual(self._first, self._trigger_priority, 'priority should be first')
+        self.assertEqual(self._second, self._trigger, 'priority should be first')
+
+        self.par.remove_observer(self)
+        self._first = self._second = None
+        
+        self.par.add_observer(self, self._trigger, 1)
+        self.par.add_observer(self, self._trigger_priority, 0)
+        self.par._notify_observers(0)
+        self.assertEqual(self._first, self._trigger, 'priority should be second')
+        self.assertEqual(self._second, self._trigger_priority, 'priority should be second')
+        
+
+if __name__ == "__main__":
+    #import sys;sys.argv = ['', 'Test.testName']
+    unittest.main()

GPy/testing/parameterized_tests.py

@@ -6,6 +6,7 @@ Created on Feb 13, 2014
 import unittest
 import GPy
 import numpy as np
+from GPy.core.parameterization.parameter_core import HierarchyError
 
 class Test(unittest.TestCase):
 
@@ -65,7 +66,7 @@ class Test(unittest.TestCase):
         self.assertListEqual(self.test1.constraints[Logexp()].tolist(), [0,1])
         
     def test_add_parameter_already_in_hirarchy(self):
-        self.test1.add_parameter(self.white._parameters_[0])
+        self.assertRaises(HierarchyError, self.test1.add_parameter, self.white._parameters_[0])        
         
     def test_default_constraints(self):
         self.assertIs(self.rbf.variance.constraints._param_index_ops, self.rbf.constraints._param_index_ops)

GPy/util/caching.py

@@ -1,6 +1,14 @@
 from ..core.parameterization.parameter_core import Observable
+import itertools
 
 class Cacher(object):
+    """
+
+
+
+
+    """
+
     def __init__(self, operation, limit=5, ignore_args=()):
         self.limit = int(limit)
         self.ignore_args = ignore_args
@@ -10,50 +18,78 @@ class Cacher(object):
         self.inputs_changed = []
 
     def __call__(self, *args):
+        """
+        A wrapper function for self.operation,
+        """
+
+        #ensure that specified arguments are ignored
         if len(self.ignore_args) != 0:
-            ca = [a for i,a in enumerate(args) if i not in self.ignore_args]
+            oa = [a for i,a in enumerate(args) if i not in self.ignore_args]
         else:
-            ca = args
+            oa = args
+
         # this makes sure we only add an observer once, and that None can be in args
-        cached_args = []
-        for a in ca:
-            if (not any(a is ai for ai in cached_args)) and a is not None:
-                cached_args.append(a)
-        if not all([isinstance(arg, Observable) for arg in cached_args]):
-            print cached_args
-            import ipdb;ipdb.set_trace()
+        observable_args = []
+        for a in oa:
+            if (not any(a is ai for ai in observable_args)) and a is not None:
+                observable_args.append(a)
+
+        #make sure that all the found argument really are observable:
+        #otherswise don't cache anything, pass args straight though
+        if not all([isinstance(arg, Observable) for arg in observable_args]):
             return self.operation(*args)
-        
-        if cached_args in self.cached_inputs:
-            i = self.cached_inputs.index(cached_args)
+
+        # TODO: WARNING !!! Cache OFFSWITCH !!! WARNING
+        # return self.operation(*args)
+
+        #if the result is cached, return the cached computation
+        state = [all(a is b for a, b in itertools.izip_longest(args, cached_i)) for cached_i in self.cached_inputs]
+        if any(state):
+            i = state.index(True)
             if self.inputs_changed[i]:
+                #(elements of) the args have changed since we last computed: update
                 self.cached_outputs[i] = self.operation(*args)
                 self.inputs_changed[i] = False
             return self.cached_outputs[i]
         else:
+            #first time we've seen these arguments: compute
+
+            #first make sure the depth limit isn't exceeded
             if len(self.cached_inputs) == self.limit:
                 args_ = self.cached_inputs.pop(0)
-                [a.remove_observer(self, self.on_cache_changed) for a in args_]
+                [a.remove_observer(self, self.on_cache_changed) for a in args_ if a is not None]
                 self.inputs_changed.pop(0)
                 self.cached_outputs.pop(0)
 
-            self.cached_inputs.append(cached_args)
+            #compute
+            self.cached_inputs.append(args)
             self.cached_outputs.append(self.operation(*args))
             self.inputs_changed.append(False)
-            [a.add_observer(self, self.on_cache_changed) for a in cached_args]
-            return self.cached_outputs[-1]
+            [a.add_observer(self, self.on_cache_changed) for a in observable_args]
+            return self.cached_outputs[-1]#Max says return.
 
     def on_cache_changed(self, arg):
+        """
+        A callback funtion, which sets local flags when the elements of some cached inputs change
+
+        this function gets 'hooked up' to the inputs when we cache them, and upon their elements being changed we update here.
+        """
         self.inputs_changed = [any([a is arg for a in args]) or old_ic for args, old_ic in zip(self.cached_inputs, self.inputs_changed)]
 
     def reset(self, obj):
-        [[a.remove_observer(self, self.on_cache_changed) for a in args] for args in self.cached_inputs]
-        [[a.remove_observer(self, self.reset) for a in args] for args in self.cached_inputs]
+        """
+        Totally reset the cache
+        """
+        [[a.remove_observer(self, self.on_cache_changed) for a in args if isinstance(a, Observable)] for args in self.cached_inputs]
+        [[a.remove_observer(self, self.reset) for a in args if isinstance(a, Observable)] for args in self.cached_inputs]
         self.cached_inputs = []
         self.cached_outputs = []
         self.inputs_changed = []
 
 class Cache_this(object):
+    """
+    A decorator which can be applied to bound methods in order to cache them
+    """
     def __init__(self, limit=5, ignore_args=()):
         self.limit = limit
         self.ignore_args = ignore_args
@@ -64,4 +100,5 @@ class Cache_this(object):
                 self.c = Cacher(f, self.limit, ignore_args=self.ignore_args)
             return self.c(*args)
         f_wrap._cacher = self
-        return f_wrap
+        f_wrap.__doc__ = "**cached**\n\n" + (f.__doc__ or "")
+        return f_wrap