def toy_spline_kpl_corr2(kernel_sigma, regu, tasks_correl):
# Spline dict
locs_bounds = np.array(
[toy_data_spline.FREQS[0], toy_data_spline.FREQS[-1]])
freqs_bounds = toy_data_spline.FREQS[0], toy_data_spline.FREQS[-1]
width = toy_data_spline.WIDTH
domain = np.expand_dims(np.array(
[freqs_bounds[0] - width / 2, freqs_bounds[1] + width / 2]),
axis=0)
func_dict = basis.BSplineUniscaleBasis(domain,
freqs_bounds[1],
locs_bounds,
width=width,
add_constant=False)
# Scalar kernel
gauss_ker = kernels.GaussianScalarKernel(kernel_sigma, normalize=False)
# Operator valued kernel matrix
output_matrix_params = {"omega": tasks_correl}
# output_matrices = configs_generation.subconfigs_combinations("chain_graph", output_matrix_params)
output_matrices = configs_generation.subconfigs_combinations(
"all_related", output_matrix_params)
params = {
"kernel": gauss_ker,
"regu": regu,
"B": output_matrices,
"basis_out": func_dict,
"center_output": False
}
configs = configs_generation.configs_combinations(params)
# Create list of regressors from that config
regs = [kproj_learning.SeperableKPL(**config) for config in configs]
return configs, regs
def toy_spline_kpl2(kernel_sigma, regu):
# Spline dict
locs_bounds = np.array(
[toy_data_spline.FREQS[0], toy_data_spline.FREQS[-1]])
freqs_bounds = toy_data_spline.FREQS[0], toy_data_spline.FREQS[-1]
width = toy_data_spline.WIDTH
domain = np.expand_dims(np.array(
[freqs_bounds[0] - width / 2, freqs_bounds[1] + width / 2]),
axis=0)
func_dict = basis.BSplineUniscaleBasis(domain,
freqs_bounds[1],
locs_bounds,
width=width,
add_constant=False)
# Scalar kernel
gauss_ker = kernels.GaussianScalarKernel(kernel_sigma, normalize=False)
# Operator valued kernel matrix
B = np.eye(func_dict.n_basis)
regs = [
kproj_learning.SeperableKPL(r,
gauss_ker,
B,
func_dict,
center_output=False) for r in regu
]
configs = configs_generation.configs_combinations({"regu": regu})
return configs, regs
def toy_spline_kpl_corr(kernel_sigma, regu, tasks_correl):
# Spline dict
locs_bounds = np.array(
[toy_data_spline.BOUNDS_FREQS[0], toy_data_spline.BOUNDS_FREQS[1]])
domain = toy_data_spline.DOM_OUTPUT
func_dict = basis.BSplineUniscaleBasis(domain,
toy_data_spline.BOUNDS_FREQS[-1],
locs_bounds,
width=toy_data_spline.WIDTH,
add_constant=False)
# Scalar kernel
gauss_ker = kernels.GaussianScalarKernel(kernel_sigma, normalize=False)
# Operator valued kernel matrix
output_matrix_params = {"omega": tasks_correl, "dim": func_dict.n_basis}
# output_matrices = configs_generation.subconfigs_combinations("chain_graph", output_matrix_params)
output_matrices = configs_generation.subconfigs_combinations(
"neighbors_correl", output_matrix_params)
params = {
"kernel": gauss_ker,
"regu": regu,
"B": output_matrices,
"basis_out": func_dict,
"center_output": False
}
configs = configs_generation.configs_combinations(params)
# Create list of regressors from that config
regs = [kproj_learning.SeperableKPL(**config) for config in configs]
return configs, regs
def dti_wavs_kpl(kernel_sigma,
regu,
center_output=True,
decrease_base=1,
pywt_name="db",
moments=2,
n_dilat=4,
init_dilat=1.0,
translat=1.0,
dilat=2,
approx_level=5,
add_constant=True,
domain=np.array([[0, 1]]),
locs_bounds=np.array([[0, 1]])):
# Wavelets output basses
output_basis_params = {
"pywt_name": pywt_name,
"moments": moments,
"init_dilat": init_dilat,
"translat": translat,
"dilat": dilat,
"n_dilat": n_dilat,
"approx_level": approx_level,
"add_constant": add_constant,
"locs_bounds": locs_bounds,
"domain": domain
}
output_bases = configs_generation.subconfigs_combinations(
"wavelets",
output_basis_params,
exclude_list=["domain", "locs_bounds"])
# Gaussian kernel
kernel = kernels.GaussianScalarKernel(kernel_sigma, normalize=False)
# Penalize power
output_matrix_params = {"decrease_base": decrease_base}
output_matrices = configs_generation.subconfigs_combinations(
"wavelets_pow", output_matrix_params)
# Generate full configs
params = {
"kernel": kernel,
"B": output_matrices,
"basis_out": output_bases,
"regu": regu,
"center_output": center_output
}
configs = configs_generation.configs_combinations(params)
# Create list of regressors from that config
regs = [kproj_learning.SeperableKPL(**config) for config in configs]
return configs, regs
def speech_rffs_kpl(kernel_sigma,
regu,
n_rffs,
rffs_sigma,
seed_rffs,
center_output,
domain=np.array([[0, 1]])):
# FPCA output basis
output_basis_params = {
"n_basis": n_rffs,
"bandwidth": rffs_sigma,
"input_dim": 1,
"domain": domain,
"seed": seed_rffs
}
output_bases = configs_generation.subconfigs_combinations(
"random_fourier", output_basis_params, exclude_list=["domain"])
# Sum of Gaussian kernels
kernel_sigmas = kernel_sigma * np.ones(13)
gauss_kers = [
kernels.GaussianScalarKernel(sig, normalize=False, normalize_dist=True)
for sig in kernel_sigmas
]
multi_ker = kernels.SumOfScalarKernel(gauss_kers, normalize=False)
# Penalize power
output_matrix_params = {}
output_matrices = configs_generation.subconfigs_combinations(
"eye", output_matrix_params)
# Generate full configs
params = {
"kernel": multi_ker,
"B": output_matrices,
"basis_out": output_bases,
"regu": regu,
"center_output": center_output
}
configs = configs_generation.configs_combinations(params)
# Create list of regressors from that config
regs = [kproj_learning.SeperableKPL(**config) for config in configs]
return configs, regs
def speech_fpca_penpow_kpl(kernel_sigma,
regu,
n_fpca,
n_evals_fpca,
decrease_base,
domain=np.array([[0, 1]])):
# FPCA output basis
output_basis_params = {
"n_basis": n_fpca,
"input_dim": 1,
"domain": domain,
"n_evals": n_evals_fpca
}
output_bases = configs_generation.subconfigs_combinations(
"functional_pca", output_basis_params, exclude_list=["domain"])
# Sum of Gaussian kernels
kernel_sigmas = kernel_sigma * np.ones(13)
gauss_kers = [
kernels.GaussianScalarKernel(sig, normalize=False, normalize_dist=True)
for sig in kernel_sigmas
]
multi_ker = kernels.SumOfScalarKernel(gauss_kers, normalize=False)
# Penalize power
output_matrix_params = {"decrease_base": decrease_base}
output_matrices = configs_generation.subconfigs_combinations(
"pow", output_matrix_params)
# Generate full configs
params = {
"kernel": multi_ker,
"B": output_matrices,
"basis_out": output_bases,
"regu": regu,
"center_output": True
}
configs = configs_generation.configs_combinations(params)
# Create list of regressors from that config
regs = [kproj_learning.SeperableKPL(**config) for config in configs]
return configs, regs
def toy_spline_kpl(kernel_sigma, regu):
# Spline dict
locs_bounds = np.array(
[toy_data_spline.BOUNDS_FREQS[0], toy_data_spline.BOUNDS_FREQS[1]])
domain = toy_data_spline.DOM_OUTPUT
func_dict = basis.BSplineUniscaleBasis(domain,
toy_data_spline.BOUNDS_FREQS[-1],
locs_bounds,
width=toy_data_spline.WIDTH,
add_constant=False)
# Scalar kernel
gauss_ker = kernels.GaussianScalarKernel(kernel_sigma, normalize=False)
# Operator valued kernel matrix
B = np.eye(func_dict.n_basis)
regs = [
kproj_learning.SeperableKPL(r,
gauss_ker,
B,
func_dict,
center_output=False) for r in regu
]
configs = configs_generation.configs_combinations({"regu": regu})
return configs, regs