def _update_hyperstate2(self, X, y, update_hyperstate):
if update_hyperstate:
basis = _basis(X, self.random_matrix, self.bias, self.basis_dim,
self.length_scale, self.signal_sd)
basis = np.expand_dims(basis, axis=1)
y = y[..., np.newaxis]
XX_tiled_new = self.XX_tiled + np.matmul(
np.transpose(basis, [0, 2, 1]), basis)
Xy_tiled_new = self.Xy_tiled + np.matmul(
np.transpose(basis, [0, 2, 1]), y)
XX_tiled, Xy_tiled, Llower_tiled = [], [], []
for XX_old, XX_new, Xy_old, Xy_new, Llower_old in zip(
self.XX_tiled, XX_tiled_new, self.Xy_tiled, Xy_tiled_new,
self.Llower_tiled):
try:
tmp = spla.cholesky(XX_new +
(self.noise_sd / self.prior_sd)**2 *
np.eye(self.basis_dim),
lower=True)
XX_tiled.append(XX_new.copy())
Xy_tiled.append(Xy_new.copy())
Llower_tiled.append(tmp.copy())
except Exception as e:
print(e)
XX_tiled.append(XX_old.copy())
Xy_tiled.append(Xy_old.copy())
Llower_tiled.append(Llower_old.copy())
self.XX_tiled = np.stack(XX_tiled, axis=0)
self.Xy_tiled = np.stack(Xy_tiled, axis=0)
self.Llower_tiled = np.stack(Llower_tiled, axis=0)
def _update_hyperstate(self, X, y, update_hyperstate):
if update_hyperstate:
basis = _basis(X, self.random_matrix, self.bias, self.basis_dim, self.length_scale, self.signal_sd)
self.Llower_tiled = self.Llower_tiled.transpose([0, 2, 1])
assert len(self.Llower_tiled) == len(basis)
for i in range(len(self.Llower_tiled)):
cholupdate(self.Llower_tiled[i], basis[i].copy())
self.Llower_tiled = self.Llower_tiled.transpose([0, 2, 1])
self.Xy_tiled += np.matmul(basis[:, None, :].transpose([0, 2, 1]), y[:, None, :])
def _predict(self, X, update_hyperstate):
if update_hyperstate:
basis = _basis(X, self.random_matrix, self.bias, self.basis_dim, self.length_scale, self.signal_sd)
basis = np.expand_dims(basis, axis=1)
LinvXT = solve_triangular(self.Llower_tiled, np.transpose(basis, [0, 2, 1]))
pred_sigma = np.sum(np.square(LinvXT), axis=1)*self.noise_sd**2+self.noise_sd**2
tmp0 = np.transpose(solve_triangular(self.Llower_tiled, np.transpose(basis, [0, 2, 1])), [0, 2, 1])
tmp1 = solve_triangular(self.Llower_tiled, self.Xy_tiled)
pred_mu = np.matmul(tmp0, tmp1)
pred_mu = np.squeeze(pred_mu, axis=-1)
return pred_mu, pred_sigma
else:
return RegressionWrapper._predict(self, X)