apunkt/GPy
1
0
Fork 0
mirror of https://github.com/SheffieldML/GPy.git synced 2026-05-07 02:52:40 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions

fix the dL_dK symmetric issue for linear kernel and set dL_dK in the kernel test to be random.

Browse source
This commit is contained in:
Zhenwen Dai 2015-10-22 17:57:03 +01:00
parent e18de2b2c0
commit ee5a562c67
2 changed files with 12 additions and 7 deletions
Download patch file Download diff file Expand all files Collapse all files

13
GPy/kern/src/linear.py
View file

@ -73,14 +73,15 @@ class Linear(Kern):
return np.sum(self.variances * np.square(X), -1)
def update_gradients_full(self, dL_dK, X, X2=None):
if X2 is None: dL_dK = (dL_dK+dL_dK.T)/2
if self.ARD:
if X2 is None:
#self.variances.gradient = np.array([np.sum(dL_dK * tdot(X[:, i:i + 1])) for i in range(self.input_dim)])
self.variances.gradient = np.einsum('ij,iq,jq->q', dL_dK, X, X)
self.variances.gradient = (dL_dK.dot(X)*X).sum(0) #np.einsum('ij,iq,jq->q', dL_dK, X, X)
else:
#product = X[:, None, :] * X2[None, :, :]
#self.variances.gradient = (dL_dK[:, :, None] * product).sum(0).sum(0)
self.variances.gradient = np.einsum('ij,iq,jq->q', dL_dK, X, X2)
self.variances.gradient = (dL_dK.dot(X2)*X).sum(0) #np.einsum('ij,iq,jq->q', dL_dK, X, X2)
else:
self.variances.gradient = np.sum(self._dot_product(X, X2) * dL_dK)
@ -93,13 +94,15 @@ class Linear(Kern):
def gradients_X(self, dL_dK, X, X2=None):
if X2 is None: dL_dK = (dL_dK+dL_dK.T)/2
if X2 is None:
return np.einsum('jq,q,ij->iq', X, 2*self.variances, dL_dK)
return dL_dK.dot(X)*(2*self.variances) #np.einsum('jq,q,ij->iq', X, 2*self.variances, dL_dK)
else:
#return (((X2[None,:, :] * self.variances)) * dL_dK[:, :, None]).sum(1)
return np.einsum('jq,q,ij->iq', X2, self.variances, dL_dK)
return dL_dK.dot(X2)*self.variances #np.einsum('jq,q,ij->iq', X2, self.variances, dL_dK)
def gradients_XX(self, dL_dK, X, X2=None):
if X2 is None: dL_dK = (dL_dK+dL_dK.T)/2
if X2 is None:
return 2*np.ones(X.shape)*self.variances
else:
@ -162,6 +165,7 @@ class LinearFull(Kern):
return np.einsum('ij,jk,lk->il', X, P, X if X2 is None else X2)
def update_gradients_full(self, dL_dK, X, X2=None):
if X2 is None: dL_dK = (dL_dK+dL_dK.T)/2
self.kappa.gradient = np.einsum('ij,ik,kj->j', X, dL_dK, X if X2 is None else X2)
self.W.gradient = np.einsum('ij,kl,ik,lm->jm', X, X if X2 is None else X2, dL_dK, self.W)
self.W.gradient += np.einsum('ij,kl,ik,jm->lm', X, X if X2 is None else X2, dL_dK, self.W)
@ -175,6 +179,7 @@ class LinearFull(Kern):
self.W.gradient = 2.*np.einsum('ij,ik,jl,i->kl', X, X, self.W, dL_dKdiag)
def gradients_X(self, dL_dK, X, X2=None):
if X2 is None: dL_dK = (dL_dK+dL_dK.T)/2
P = np.dot(self.W, self.W.T) + np.diag(self.kappa)
if X2 is None:
return 2.*np.einsum('ij,jk,kl->il', dL_dK, X, P)