def _child_gradient(self, param, X1, X2, param_num=None):
""" Computes gradient of kernel w.r.t the param. """
scaled_X1 = self.get_scaled_repr(X1)
scaled_X2 = self.get_scaled_repr(X2)
dist = np.sqrt(dist_squared(scaled_X1, scaled_X2))
if param == 'scale':
unnorm_kernel_vals = self._eval_kernel_values_unnormalised(dist)
return self.hyperparams[
'scale'] * self.norm_constant * unnorm_kernel_vals
elif param == 'same_dim_bandwidths':
dist_dv = -(dist / (self.hyperparams['dim_bandwidths'][0, 0]))
unnorm_kernel_vals = self._eval_grad_kernel_values_unnormalised(
dist, dist_dv)
return self.hyperparams[
'scale'] * self.norm_constant * unnorm_kernel_vals
else:
dim_X1 = np.expand_dims(scaled_X1[:, param_num], axis=1)
dim_X2 = np.expand_dims(scaled_X2[:, param_num], axis=1)
np.fill_diagonal(dist, 1.0)
dist_dv = 1 / dist
np.fill_diagonal(dist, 0.0)
dim_sq = dist_squared(dim_X1, dim_X2)
dist_dv *= -(dim_sq /
(self.hyperparams['dim_bandwidths'][0, param_num]))
unnorm_kernel_vals = self._eval_grad_kernel_values_unnormalised(
dist, dist_dv)
return self.hyperparams[
'scale'] * self.norm_constant * unnorm_kernel_vals
def _child_evaluate(self, X1, X2):
""" Evaluates the SE kernel between X1 and X2 and returns the gram matrix. """
scaled_X1 = self.get_scaled_repr(X1)
scaled_X2 = self.get_scaled_repr(X2)
dist_sq = dist_squared(scaled_X1, scaled_X2)
K = self.hyperparams['scale'] * np.exp(-dist_sq / 2)
return K
def _child_evaluate(self, X1, X2):
""" Evaluates the SE kernel between X1 and X2 and returns the gram matrix. """
scaled_X1 = self.get_scaled_repr(X1)
scaled_X2 = self.get_scaled_repr(X2)
dist = np.sqrt(dist_squared(scaled_X1, scaled_X2))
unnorm_kernel_vals = self._eval_kernel_values_unnormalised(dist)
K = self.hyperparams['scale'] * self.norm_constant * unnorm_kernel_vals
return K
def _child_gradient(self, param, X1, X2, param_num=None):
""" Computes gradient of kernel w.r.t the param """
scaled_X1 = self.get_scaled_repr(X1)
scaled_X2 = self.get_scaled_repr(X2)
dist_sq = dist_squared(scaled_X1, scaled_X2)
if param == 'scale':
return self.hyperparams['scale'] * np.exp(-dist_sq / 2)
elif param == 'same_dim_bandwidths':
dist_sq_dv = dist_sq / (self.hyperparams['dim_bandwidths'][0, 0])
return self.hyperparams['scale'] * (np.multiply(
dist_sq_dv, np.exp(-dist_sq / 2)))
else:
dim_X1 = np.expand_dims(scaled_X1[:, param_num], axis=1)
dim_X2 = np.expand_dims(scaled_X2[:, param_num], axis=1)
dim_sq = dist_squared(dim_X1, dim_X2)
dim_sq = dim_sq / (self.hyperparams['dim_bandwidths'][0,
param_num])
return self.hyperparams['scale'] * (np.multiply(
dim_sq, np.exp(-dist_sq / 2)))
def test_dist_squared(self):
""" Tests the squared distance function. """
self.report('dist_squared')
comp_dist_sq = general_utils.dist_squared(self.X1, self.X2)
assert (self.true_dist_sq == comp_dist_sq).all()