diff --git a/GPy/examples/sparse_GP_regression_demo.py b/GPy/examples/sparse_GP_regression_demo.py index e996ce50..13ff47e0 100644 --- a/GPy/examples/sparse_GP_regression_demo.py +++ b/GPy/examples/sparse_GP_regression_demo.py @@ -21,7 +21,7 @@ X = np.random.uniform(-3.,3.,(N,1)) Y = np.sin(X)+np.random.randn(N,1)*0.05 # construct kernel -rbf = GPy.kern.Matern52(1) +rbf = GPy.kern.rbf(1) noise = GPy.kern.white(1) kernel = rbf + noise @@ -37,7 +37,7 @@ m1.constrain_positive('(variance|lengthscale|precision)') #check gradient FIXME unit test please m1.checkgrad() # optimize and plot -m1.optimize('bfgs', messages = 1) +m1.optimize('tnc', messages = 1) m1.plot() # print(m1) diff --git a/GPy/kern/linear.py b/GPy/kern/linear.py index 8f960a9c..d3e3f42c 100644 --- a/GPy/kern/linear.py +++ b/GPy/kern/linear.py @@ -48,6 +48,9 @@ class linear(kernpart): def dK_dX(self,partial,X,X2,target): target += self.variance * np.sum(partial[:,None,:]*X2.T[None,:,:],-1) + def dKdiag_dtheta(self,partial,X,target): + target += np.sum(partial*np.square(X).sum(1)) + # def psi0(self,Z,mu,S,target): # expected = np.square(mu) + S # np.add(target,np.sum(self.variance*expected),target) diff --git a/GPy/kern/rbf.py b/GPy/kern/rbf.py index f899737a..658130b3 100644 --- a/GPy/kern/rbf.py +++ b/GPy/kern/rbf.py @@ -63,7 +63,7 @@ class rbf(kernpart): self._K_computations(X,X2) _K_dist = X[:,None,:]-X2[None,:,:] dK_dX = np.transpose(-self.variance*self._K_dvar[:,:,np.newaxis]*_K_dist/self.lengthscale2,(1,0,2)) - target += np.sum(dK_dX*partial[:,:,None],0) + target += np.sum(dK_dX*partial.T[:,:,None],0) def dKdiag_dX(self,X,target): pass diff --git a/GPy/kern/white.py b/GPy/kern/white.py index 04effabc..df62d4b6 100644 --- a/GPy/kern/white.py +++ b/GPy/kern/white.py @@ -42,13 +42,13 @@ class white(kernpart): if np.all(X==X2): target += np.trace(partial) - def dKdiag_dtheta(self,X,target): - np.add(target[:,0],1.,target[:,0]) + def dKdiag_dtheta(self,partial,X,target): + target += np.sum(partial) def dK_dX(self,partial,X,X2,target): pass - def dKdiag_dX(self,X,target): + def dKdiag_dX(self,partial,X,target): pass def psi0(self,Z,mu,S,target): diff --git a/GPy/models/sparse_GP_regression.py b/GPy/models/sparse_GP_regression.py index ef86afa9..39a38214 100644 --- a/GPy/models/sparse_GP_regression.py +++ b/GPy/models/sparse_GP_regression.py @@ -63,15 +63,15 @@ class sparse_GP_regression(GP_regression): self.psi1 = self.kern.K(self.Z,self.X) self.psi2 = np.dot(self.psi1,self.psi1.T) - self.dKmm_dtheta = self.kern.dK_dtheta(self.Z) - self.dpsi0_dtheta = self.kern.dKdiag_dtheta(self.X).sum(0) - self.dpsi1_dtheta = self.kern.dK_dtheta(self.Z,self.X) - tmp = np.dot(self.psi1, self.dpsi1_dtheta) - self.dpsi2_dtheta = tmp + tmp.transpose(1,0,2) + #self.dKmm_dtheta = self.kern.dK_dtheta(self.Z) + #self.dpsi0_dtheta = self.kern.dKdiag_dtheta(self.X).sum(0) + #self.dpsi1_dtheta = self.kern.dK_dtheta(self.Z,self.X) + #tmp = np.dot(self.psi1, self.dpsi1_dtheta) + #self.dpsi2_dtheta = tmp + tmp.transpose(1,0,2) - self.dpsi1_dZ = self.kern.dK_dX(self.Z,self.X) - self.dpsi2_dZ = np.tensordot(self.psi1,self.dpsi1_dZ,((1),(0)))*2.0 - self.dKmm_dZ = self.kern.dK_dX(self.Z) + #self.dpsi1_dZ = self.kern.dK_dX(self.Z,self.X) + #self.dpsi2_dZ = np.tensordot(self.psi1,self.dpsi1_dZ,((1),(0)))*2.0 + #self.dKmm_dZ = self.kern.dK_dX(self.Z) def _computations(self): # TODO find routine to multiply triangular matrices @@ -92,7 +92,7 @@ class sparse_GP_regression(GP_regression): self.G = mdot(self.LBL_inv, self.psi1YYpsi1, self.LBL_inv.T) # Computes dL_dpsi - self.dL_dpsi0 = - 0.5 * self.D * self.beta + self.dL_dpsi0 = - 0.5 * self.D * self.beta * np.ones(self.N) dC_dpsi1 = (self.LLambdai.T[:,:, None, None] * self.Y) # this is sane. tmp = (dC_dpsi1*self.C[None,:,None,:]).sum(1).sum(-1) self.dL_dpsi1 = self.beta2 * tmp @@ -145,29 +145,34 @@ class sparse_GP_regression(GP_regression): return np.squeeze(dA_dbeta + dB_dbeta + dC_dbeta + dD_dbeta + dE_dbeta) def dL_dtheta(self): - dL_dtheta = (self.dL_dpsi0 * self.dpsi0_dtheta + (self.dL_dpsi1[:,:, None] * self.dpsi1_dtheta).sum(0).sum(0) - + (self.dL_dpsi2[:, :, None] * self.dpsi2_dtheta).sum(0).sum(0) - + (self.dL_dKmm[:, :, None] * self.dKmm_dtheta).sum(0).sum(0)) + #re-cast computations in psi2 back to psi1: + dL_dpsi1 = self.dL_dpsi1 + 2.*np.dot(self.dL_dpsi2,self.psi1) + + dL_dtheta = self.kern.dK_dtheta(self.dL_dKmm,self.Z) + dL_dtheta += self.kern.dK_dtheta(dL_dpsi1,self.Z,self.X) + dL_dtheta += self.kern.dKdiag_dtheta(self.dL_dpsi0, self.X) + return dL_dtheta def dL_dZ(self): - dL_dZ = ((self.dL_dpsi1[:,:,None]*self.dpsi1_dZ.swapaxes(0,1)).sum(1) - + (self.dL_dpsi2[:, :, None] * self.dpsi2_dZ).sum(0) - + 2.0*(self.dL_dKmm[:, :, None] * self.dKmm_dZ).sum(0)) - return dL_dZ + #re-cast computations in psi2 back to psi1: + dL_dpsi1 = self.dL_dpsi1 + 2.*np.dot(self.dL_dpsi2,self.psi1) + dL_dZ = 2.*self.kern.dK_dX(self.dL_dKmm,self.Z,)#factor of two becase of vertical and horizontal 'stripes' in dKmm_dZ + dL_dZ += self.kern.dK_dX(dL_dpsi1,self.Z,self.X) + return dL_dZ def log_likelihood_gradients(self): return np.hstack([self.dL_dZ().flatten(), self.dL_dbeta(), self.dL_dtheta()]) - def _raw_predict(self,_Xnew,slices): + def _raw_predict(self,Xnew,slices): """Internal helper function for making predictions, does not account for normalisation""" - Kx = self.kern.K(self.Z, _Xnew, self.Xslices, slices) - Kxx = self.kern.K(_Xnew, slices1=slices, slices2=slices) + Kx = self.kern.K(self.Z, Xnew) + Kxx = self.kern.K(Xnew) mu = self.beta * mdot(Kx.T, self.LBL_inv, self.psi1Y) - var = Kxx - mdot(Kx.T, (self.Kmmi - self.LBL_inv), Kx) + np.eye(_Xnew.shape[0])/self.beta # TODO: This beta doesn't belong here in the EP case. + var = Kxx - mdot(Kx.T, (self.Kmmi - self.LBL_inv), Kx) + np.eye(Xnew.shape[0])/self.beta # TODO: This beta doesn't belong here in the EP case. return mu,var def plot(self,*args,**kwargs):