def _collect_test_stats(test_function: str, N: int, noise_std: float, random: bool, gps: Tuple[str, ...], M: int):
store = store_path(test_function, N, noise_std, random, M)
for k in range(K):
fold = data.Fold(store, k)
for gp in gps:
gp_path = fold.dir / gp
frame = _test_stats(k, gp_path)
def predict_roms(M: int, N: int, random: bool, noisy: bool):
store_name = 'sin.u1.'
CDF_scale = 2 * pi
CDF_loc = pi
function_with_parameters = function.CallableWithParameters(
function=function.ishigami, parameters={
'a': 0,
'b': 0
})
store_name = store_name + '{0:d}.{1:d}'.format(N, M)
store_name += '.random' if random else '.rom'
noise_std = 0.0001 if noisy else 0
store = data.Store(store_dir(store_name, noise_std, CDF_scale),
data.Store.InitMode.READ_META_ONLY)
fold = data.Fold(store, 0)
rom = model.gpy_.ROM.from_ROM(fold=fold, name='rom', suffix='.test.full')
model_theta = rom.sobol.parameters_read.Theta
data_theta = function.linear_matrix_from_meta(store)
pre_function_with_parameters = (function.CallableWithParameters(
function=function.linear, parameters={'matrix': data_theta})
if random else None)
test_store = scalar_function_of_normal(
store_name=store_name + "\\test",
N=N,
M=M,
X_std=1.0,
noise_std=noise_std,
CDF_scale=CDF_scale,
CDF_loc=CDF_loc,
pre_function_with_parameters=pre_function_with_parameters,
function_with_parameters=function_with_parameters)
fold.set_test_data(df=test_store.data.df)
rom.sobol.gp.test()
result = matmul(model_theta, data_theta.T)
print(result)
def _collect_std(test_function: str, N: int, noise_std: float, random: bool, M: int):
store = data.Store(store_path(test_function, N, noise_std, random, M))
destination = store.dir / "results"
shutil.rmtree(destination, ignore_errors=True)
destination.mkdir(mode=0o777, parents=True, exist_ok=False)
result = 0.0
for k in range(K):
fold = data.Fold(store, k)
result += fold.standard.df.iloc[-1, -1]/K
savetxt(fname=(destination / "std.csv"), X=atleast_2d(result), delimiter=",")
def _collect_test_stats(M, N, function_name, random, noisy):
noisy_str = NORMAL_CDF_DIR if noisy else NOISELESS_DIR
source_store = store_dir(M, N, function_name, random, noisy)
for k in range(FOLDS):
fold = data.Fold(source_store, k)
gp_dir = fold.dir / "ard"
frame = data.Frame(
gp_dir / "test_stats.csv",
_test_stats(k,
data.Frame(gp_dir / "__test__.csv").df.copy()))
def _collect_test_stats(test_function: str,
N: int,
noise_std: float,
random: bool,
gp: str,
M: int = 5):
source_store = store_path(test_function, N, noise_std, random, M)
for k in range(K):
fold = data.Fold(source_store, k)
gp_dir = fold.dir / gp
frame = data.Frame(
gp_dir / "test_stats.csv",
_test_stats(k,
data.Frame(gp_dir / "__test__.csv").df.copy()))
def test_random(test_fuction: str, N: int, noise_std: float, M: int = 5, K: int = 2 ):
random = True
name = 'rbf'
gp_optimizer_options = {'optimizer': 'bfgs', 'max_iters': 5000, 'gtol': 1E-16}
kernel_parameters = model.gpy_.Kernel.ExponentialQuadratic.Parameters(lengthscale=full((1, 1), 2.5**(M/5), dtype=float))
parameters = model.gpy_.GP.DEFAULT_PARAMETERS._replace(kernel=kernel_parameters, e_floor=1E-6, e=0.003)
CDF_scale = 2 * pi
CDF_loc = pi
if test_fuction == 'sin.1':
function_with_parameters = function.CallableWithParameters(function.ishigami, parameters={'a': 0.0, 'b': 0.0})
elif test_fuction == 'sin.2':
function_with_parameters = function.CallableWithParameters(function.ishigami, parameters={'a': 2.0, 'b': 0.0})
elif test_fuction == 'ishigami':
function_with_parameters = function.callable_with_parameters(function.ishigami)
else:
CDF_scale = 1.0
CDF_loc = 0.0
function_with_parameters = function.CallableWithParameters(function.sobol_g,
parameters={'m_very_important': 2, 'm_important': 3, 'm_unimportant': 4})
store_name = test_fuction + '.{0:d}.{1:.3f}.{2:d}'.format(M, noise_std, N)
if random:
lin_trans = ortho_group.rvs(M)
pre_function_with_parameters = function.CallableWithParameters(function=function.linear, parameters={'matrix': lin_trans})
store_name += '.random'
else:
pre_function_with_parameters = None
store_name += '.rom'
store_name = BASE_PATH / store_name
store = scalar_function_of_normal(store_name=store_name, N=N, M=M, X_std=1.0, noise_std=noise_std, CDF_scale=CDF_scale, CDF_loc=CDF_loc,
pre_function_with_parameters=pre_function_with_parameters,
function_with_parameters=function_with_parameters)
# lin_trans = transpose(lin_trans)
savetxt(store.dir / "InverseRotation.csv", lin_trans)
data.Fold.into_K_folds(parent=store, K=K, shuffled_before_folding=False, standard=data.Store.Standard.mean_and_std,
replace_empty_test_with_data_=True)
model.run.GPs(module=model.run.Module.GPY_, name=name, store=store, M_Used=-1, parameters=parameters, optimize=True, test=True, sobol=True,
optimizer_options=gp_optimizer_options)
model.run.GPs(module=model.run.Module.GPY_, name=name, store=store, M_Used=-1, parameters=None, optimize=True, test=True, sobol=True,
optimizer_options=gp_optimizer_options, make_ard=True)
sobol_options = {'semi_norm': model.base.Sobol.SemiNorm.DEFAULT_META, 'N_exploit': 3, 'N_explore': 4096, 'options': {'gtol': 1.0E-16}}
for k in range(K):
sobol = model.gpy_.Sobol.from_GP(data.Fold(store, k), 'rbf.ard', 'rbf.ard.derotated')
for sobol.m in reversed(range(M)):
sobol.Theta_old = lin_trans
sobol.write_parameters(sobol.Parameters(Mu=sobol.Mu, Theta=sobol.Theta, D=sobol.Tensor3AsMatrix(sobol.D), S1=None,
S=sobol.Tensor3AsMatrix(sobol.S)))
def test_random3(test_fuction: str, N: int, noise_std: float, M: int = 5, K: int = 2 ):
random = True
name = 'derotated.rbf'
gp_optimizer_options = {'optimizer': 'bfgs', 'max_iters': 5000, 'gtol': 1E-16}
kernel_parameters = model.gpy_.Kernel.ExponentialQuadratic.Parameters(lengthscale=full((1, 1), 2.5**(M/5), dtype=float))
parameters = model.gpy_.GP.DEFAULT_PARAMETERS._replace(kernel=kernel_parameters, e_floor=1E-6, e=0.003)
store_name = test_fuction + '.{0:d}.{1:.3f}.{2:d}'.format(M, noise_std, N)
store_name += '.random'
store_name = BASE_PATH / store_name
store = data.Store(store_name)
lin_trans = loadtxt(store.dir / "InverseRotation.csv")
for k in range(K):
fold = data.Fold(store, k)
replace_X_with_U(fold, transpose(lin_trans))
model.run.GPs(module=model.run.Module.GPY_, name=name, store=store, M_Used=-1, parameters=parameters, optimize=True, test=True, sobol=True,
optimizer_options=gp_optimizer_options)
model.run.GPs(module=model.run.Module.GPY_, name=name, store=store, M_Used=-1, parameters=None, optimize=True, test=True, sobol=True,
optimizer_options=gp_optimizer_options, make_ard=True)
sobol_options = {'semi_norm': model.base.Sobol.SemiNorm.DEFAULT_META, 'N_exploit': 3, 'N_explore': 4096, 'options': {'gtol': 1.0E-16}}
def _run_test(M, N, function_name, random, noisy):
noisy_str = NORMAL_CDF_DIR if noisy else NOISELESS_DIR
source_store = store_dir(M, N, function_name, random, noisy)
Mu = choose_Mu(function_name)
kernel_parameters = model.gpy_.Kernel.ExponentialQuadratic.Parameters(
lengthscale=full((1, Mu), 0.2, dtype=float))
parameters = model.gpy_.GP.DEFAULT_PARAMETERS._replace(
kernel=kernel_parameters, e_floor=1E-5, e=1E-10)
for k in range(FOLDS):
fold = data.Fold(source_store, k, Mu)
dst = fold.dir / "rom.reduced"
if dst.exists():
shutil.rmtree(dst)
shutil.copytree(src=fold.dir / "rom.optimized", dst=dst)
gp = model.gpy_.GP(fold, "rom.reduced", parameters)
gp.optimize(model.gpy_.GP.DEFAULT_OPTIMIZER_OPTIONS)
frame = data.Frame(gp.dir / "test_stats.csv",
_test_stats(k,
gp.test().df.copy()))
def predict(gb_path: Union[str, Path],
inputs: NP.Array) -> Tuple[NP.Vector, NP.Vector]:
""" Prediction using a GaussianBundle.
Args:
gb_path: Path to a model.GaussianBundle. The extension of this filename is the number of input dimensions M.
An extension of 0 or a missing extension means full order, taking M from the training data.
inputs: An (N,M) numpy array, consisting of N test inputs, each of dimension M.
Returns: A pair (predictive_mean, predictive_std) of (N,1) numpy arrays.
"""
gb_path = Path(gb_path)
gb_name = gb_path.name
Xs_taken = int(gb_path.suffix[1:])
fold_dir = gb_path.parent
k = int(fold_dir.suffix[1:])
fold = data.Fold(fold_dir.parent, k, Xs_taken)
if Xs_taken != inputs.shape[1]:
raise IndexError(
"The gp you have chosen uses {0:d} dimensional inputs".format(
fold.M))
gb = model.gpy_.GaussianBundle(fold, gb_name, reset_log_likelihood=False)
return gb.predict(inputs)
def linear_transformation(model_dir: Path) -> NP.Matrix:
with open(model_dir / "__meta__.json", mode='r') as file:
meta = load(file)
function_with_parameters = meta['origin']['functions_with_parameters'][
0].split("; matrix=")
if len(function_with_parameters) > 1:
function_with_parameters = eval(function_with_parameters[-1][:-1])
return array(function_with_parameters)
else:
return eye(meta['data']['M'], dtype=float)
if __name__ == '__main__':
store = data.Store((BASE_PATH / NOISELESS_DIR) / "sin.u1.5000.5.random",
data.Store.InitMode.READ_META_ONLY)
fold = data.Fold(store, 0)
rom = model.gpy_.ROM.from_ROM(fold=fold, name='rom', suffix='.test.full')
model_theta = rom.sobol.parameters_read.Theta
data_theta = function.linear_matrix_from_meta(store)
print(model_theta)
print(data_theta)
"""
def rename(dir_: Path):
for p in dir_.iterdir():
if p.is_dir():
if p.name == "rom..optimized":
p.replace(p.parent / "rom.optimized")
else:
rename(p)
"""
"""