def temporal_kernel(self):
kernel = White(variance=self.model_config['noise_inner'])
m_inds = list(range(self.m))
# Initialize a non-linear kernels over inputs
if self.model_config['input_nonlinear']:
scales = [self.model_config['scale']
] * self.m if self.model_config[
'scale_tie'] else self.model_config['scale']
if self.model_config['rq']:
kernel += RationalQuadratic(active_dims=m_inds,
variance=1.0,
lengthscales=scales,
alpha=1e-2)
else:
kernel += SquaredExponential(active_dims=m_inds,
variance=1.0,
lengthscales=scales)
# Add a periodic kernel over inputs
# Decay?????
if self.model_config['per']:
scales = [self.model_config['per_scale']] * self.m
periods = [self.model_config['per_period']] * self.m
base_kernel = SquaredExponential(active_dims=m_inds,
variance=1.0,
lengthscales=scales)
kernel += Periodic(base_kernel, period=periods)
# Add a linear kernel over inputs
if self.model_config['input_linear']:
variances = [self.model_config['input_linear_scale']] * self.m
kernel += LinearKernel(active_dims=m_inds, variance=variances)
return kernel
def boosting_kernel(self):
kernels = []
if self.model_config['nonlinear']:
if self.model_config['rq']:
kernels.append(
RationalQuadratic(
variance=1.0,
lengthscales=self.model_config['nonlinear_scale'],
alpha=1e-2))
else:
kernels.append(
SquaredExponential(
variance=1.0,
lengthscales=self.model_config['nonlinear_scale']))
if self.model_config['linear']:
kernels.append(
LinearKernel(variance=self.model_config['linear_scale']))
kernel = kernels[0]
for kerneli in kernels[1:]:
kernel += kerneli
return kernel
def _kernels_generator(self):
def _determine_indicies(m, pi, markov):
# Build in the Markov structure: juggle with the indices of the outputs.
p_last = pi - 1 # Index of last output that is given as input.
p_start = 0 if markov is None else max(p_last - (markov - 1), 0)
p_num = p_last - p_start + 1
# Determine the indices corresponding to the outputs and inputs.
m_inds = list(range(m))
p_inds = list(range(m + p_start, m + p_last + 1))
return m_inds, p_inds, p_num
kernels = []
for pi in range(self.num_outputs):
m_inds, p_inds, p_num = _determine_indicies(
self.m, pi, self.model_config['markov'])
# Construct inner-layers noise kernel
kernel = White(variance=self.model_config['noise_inner'])
# Initialize a non-linear kernels over inputs
#if pi==0:
scales = [self.model_config['scale']
] * self.m if self.model_config[
'scale_tie'] else self.model_config['scale']
if self.model_config['rq']:
kernel += RationalQuadratic(active_dims=m_inds,
variance=1.0,
lengthscales=scales,
alpha=1e-2)
else:
kernel += SquaredExponential(active_dims=m_inds,
variance=1.0,
lengthscales=scales)
# Add a periodic kernel over inputs
# Decay?????
if self.model_config['per']:
scales = [self.model_config['per_scale']] * self.m
periods = [self.model_config['per_period']] * self.m
base_kernel = SquaredExponential(active_dims=m_inds,
variance=1.0,
lengthscales=scales)
kernel += Periodic(base_kernel, period=periods)
# Add a linear kernel over inputs
if self.model_config['input_linear']:
variances = [self.model_config['input_linear_scale']] * self.m
kernel += LinearKernel(active_dims=m_inds, variance=variances)
# Add a linear kernel over outputs
if self.model_config['linear'] and pi > 0:
variances = [self.model_config['linear_scale']] * p_num
kernel += LinearKernel(active_dims=p_inds, variance=variances)
# Add a non-linear kernel over outputs
if self.model_config['nonlinear'] and pi > 0:
if self.model_config['nonlinear_dependent']:
active_dims = m_inds.extend(p_inds)
scales = [self.model_config['scale']] * self.m
scales.extend([self.model_config['nonlinear_scale']] *
p_num)
else:
active_dims = p_inds
scales = [self.model_config['nonlinear_scale']] * p_num
if self.model_config['rq']:
kernel += RationalQuadratic(active_dims=active_dims,
variance=1.0,
lengthscales=scales,
alpha=1e-2)
else:
kernel += SquaredExponential(active_dims=active_dims,
variance=1.0,
lengthscales=scales)
kernels.append(kernel)
return kernels