apunkt/GPy
1
0
Fork 0
mirror of https://github.com/SheffieldML/GPy.git synced 2026-04-28 14:26:23 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions

Small changes in svigp

Browse source
This commit is contained in:
Andreas 2013-12-13 14:01:01 +00:00
parent 958e9f7c7a
commit 4d82f30367
1 changed files with 13 additions and 8 deletions
Download patch file Download diff file Expand all files Collapse all files

19
GPy/core/svigp.py
View file

@ -52,7 +52,6 @@ class SVIGP(GPBase):
self.Y = self.likelihood.Y.copy() self.Y = self.likelihood.Y.copy()
self.Z = Z self.Z = Z
self.num_inducing = Z.shape[0] self.num_inducing = Z.shape[0]
self.batchcounter = 0 self.batchcounter = 0
self.epochs = 0 self.epochs = 0
self.iterations = 0 self.iterations = 0
@ -323,7 +322,7 @@ class SVIGP(GPBase):
self._param_trace.append(self._get_params()) self._param_trace.append(self._get_params())
self._ll_trace.append(self.log_likelihood() + self.log_prior()) self._ll_trace.append(self.log_likelihood() + self.log_prior())
#load a batch #load a batch and do the appropriate computations (kernel matrices, etc)
self.load_batch() self.load_batch()
#compute the (stochastic) gradient #compute the (stochastic) gradient
@ -333,7 +332,8 @@ class SVIGP(GPBase):
#compute the steps in all parameters #compute the steps in all parameters
vb_step = self.vb_steplength*natgrads[0] vb_step = self.vb_steplength*natgrads[0]
if (self.epochs>=1):#only move the parameters after the first epoch #only move the parameters after the first epoch and only if the steplength is nonzero
if (self.epochs>=1) and (self.param_steplength > 0):
param_step = self.momentum*param_step + self.param_steplength*grads param_step = self.momentum*param_step + self.param_steplength*grads
else: else:
param_step = 0. param_step = 0.
@ -355,6 +355,8 @@ class SVIGP(GPBase):
if self.epochs > 10: if self.epochs > 10:
self._adapt_steplength() self._adapt_steplength()
self._vb_steplength_trace.append(self.vb_steplength)
self._param_steplength_trace.append(self.param_steplength)
self.iterations += 1 self.iterations += 1
@ -363,17 +365,20 @@ class SVIGP(GPBase):
if self.adapt_vb_steplength: if self.adapt_vb_steplength:
# self._adaptive_vb_steplength() # self._adaptive_vb_steplength()
self._adaptive_vb_steplength_KL() self._adaptive_vb_steplength_KL()
self._vb_steplength_trace.append(self.vb_steplength) #self._vb_steplength_trace.append(self.vb_steplength)
assert self.vb_steplength > 0 assert self.vb_steplength >= 0
if self.adapt_param_steplength: if self.adapt_param_steplength:
self._adaptive_param_steplength() self._adaptive_param_steplength()
# self._adaptive_param_steplength_log() # self._adaptive_param_steplength_log()
# self._adaptive_param_steplength_from_vb() # self._adaptive_param_steplength_from_vb()
self._param_steplength_trace.append(self.param_steplength) #self._param_steplength_trace.append(self.param_steplength)
def _adaptive_param_steplength(self): def _adaptive_param_steplength(self):
decr_factor = 0.02 if hasattr(self, 'adapt_param_steplength_decr'):
decr_factor = self.adapt_param_steplength_decr
else:
decr_factor = 0.02
g_tp = self._transform_gradients(self._log_likelihood_gradients()) g_tp = self._transform_gradients(self._log_likelihood_gradients())
self.gbar_tp = (1-1/self.tau_tp)*self.gbar_tp + 1/self.tau_tp * g_tp self.gbar_tp = (1-1/self.tau_tp)*self.gbar_tp + 1/self.tau_tp * g_tp
self.hbar_tp = (1-1/self.tau_tp)*self.hbar_tp + 1/self.tau_tp * np.dot(g_tp.T, g_tp) self.hbar_tp = (1-1/self.tau_tp)*self.hbar_tp + 1/self.tau_tp * np.dot(g_tp.T, g_tp)