From e4868d4e233a47d044ad9c0df6c804412801f6b2 Mon Sep 17 00:00:00 2001
From: Maria Fernanda Morales <65073126+mfmo45@users.noreply.github.com>
Date: Mon, 2 Sep 2024 14:22:58 +0200
Subject: [PATCH] Created a GP-Skl class, tested
---
.../example_analytical_function_gp.py | 309 ++++++++++++
src/bayesvalidrox/__init__.py | 4 +-
.../gaussian_process_sklearn.py | 466 ++++++++++++++++++
3 files changed, 778 insertions(+), 1 deletion(-)
create mode 100644 examples/analytical-function/example_analytical_function_gp.py
create mode 100644 src/bayesvalidrox/surrogate_models/gaussian_process_sklearn.py
diff --git a/examples/analytical-function/example_analytical_function_gp.py b/examples/analytical-function/example_analytical_function_gp.py
new file mode 100644
index 000000000..9788ac701
--- /dev/null
+++ b/examples/analytical-function/example_analytical_function_gp.py
@@ -0,0 +1,309 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+"""
+This test shows a surrogate-assisted Bayesian calibration of a time dependent
+ analytical function.
+
+Author: Farid Mohammadi, M.Sc.
+E-Mail: farid.mohammadi@iws.uni-stuttgart.de
+Department of Hydromechanics and Modelling of Hydrosystems (LH2)
+Institute for Modelling Hydraulic and Environmental Systems (IWS), University
+of Stuttgart, www.iws.uni-stuttgart.de/lh2/
+Pfaffenwaldring 61
+70569 Stuttgart
+
+Created on Fri Aug 9 2019
+
+"""
+
+import numpy as np
+import pandas as pd
+import sys
+import joblib
+import matplotlib
+
+matplotlib.use('agg')
+
+# import bayesvalidrox
+# Add BayesValidRox path
+sys.path.append("../../src/")
+print(sys.path)
+
+from bayesvalidrox import PyLinkForwardModel, Input, ExpDesigns, GPESkl, PostProcessing, BayesInference, Discrepancy, \
+ Engine
+
+if __name__ == "__main__":
+
+ # Number of parameters
+ ndim = 10 # 2, 10
+
+ # =====================================================
+ # ============= COMPUTATIONAL MODEL ================
+ # =====================================================
+ Model = PyLinkForwardModel()
+
+ Model.link_type = 'Function'
+ Model.py_file = 'analytical_function'
+ Model.name = 'AnalyticFunc'
+
+ Model.Output.names = ['Z']
+
+ # For Bayesian inversion synthetic data with X=[0,0]
+ Model.observations = {}
+ Model.observations['Time [s]'] = np.arange(0, 10, 1.) / 9
+ Model.observations['Z'] = np.repeat([2.], 10)
+
+ # For Checking with the MonteCarlo refrence
+ Model.mc_reference = {}
+ Model.mc_reference['Time [s]'] = np.arange(0, 10, 1.) / 9
+ Model.mc_reference['mean'] = np.load(f"data/mean_{ndim}.npy")
+ Model.mc_reference['std'] = np.load(f"data/std_{ndim}.npy")
+
+ # =====================================================
+ # ========= PROBABILISTIC INPUT MODEL ==============
+ # =====================================================
+ # Define the uncertain parameters with their mean and
+ # standard deviation
+ Inputs = Input()
+
+ # Assuming dependent input variables
+ # Inputs.Rosenblatt = True
+
+ for i in range(ndim):
+ Inputs.add_marginals()
+ Inputs.Marginals[i].name = "$\\theta_{" + str(i + 1) + "}$"
+ Inputs.Marginals[i].dist_type = 'uniform'
+ Inputs.Marginals[i].parameters = [-5, 5]
+
+ # arbitrary polynomial chaos
+ # inputParams = np.load('data/InputParameters_{}.npy'.format(ndim))
+ # for i in range(ndim):
+ # Inputs.add_marginals()
+ # Inputs.Marginals[i].name = f'$X_{i+1}$'
+ # Inputs.Marginals[i].input_data = inputParams[:, i]
+
+ # =====================================================
+ # ========== DEFINITION OF THE METAMODEL ============
+ # =====================================================
+ MetaModelOpts = GPESkl(Inputs) # , Model)
+
+ # Select if you want to preserve the spatial/temporal depencencies
+ # MetaModelOpts.dim_red_method = 'PCA'
+ # MetaModelOpts.var_pca_threshold = 99.999
+ # MetaModelOpts.n_pca_components = 10
+
+ # ToDo: This should not be an option for GPEs
+ # Select your metamodel method
+ # 1) PCE (Polynomial Chaos Expansion) 2) aPCE (arbitrary PCE)
+ # 3) GPE (Gaussian Process Emulator)
+ MetaModelOpts.meta_model_type = 'GPE'
+
+ # ------------------------------------------------
+ # ------------- GPE Specification ----------------
+ # ------------------------------------------------
+ # Select the solver for solving for the GP Hyperparameters using the ML approach
+ # ToDo: Remove this as a user-defined parameter, since only one is available?
+ # 1)LBFGS: only option for Scikit Learn
+ MetaModelOpts._gpe_reg_method = 'LBFGS'
+
+ # Kernel options ----------------------------
+ # Loop over different Kernels:
+ # 1) True to loop over the different kernel types and select the best one
+ MetaModelOpts._autoSelect = False
+
+ # Select Kernel type:
+ # 1) RBF: Gaussian/squared exponential kernel
+ # 2) Matern Kernel
+ # 3) RQ: Rational Quadratic kernel
+ MetaModelOpts._kernel_type = 'RBF'
+
+ MetaModelOpts._kernel_isotropy = False # Kernel isotropy
+ MetaModelOpts._nugget = 1e-9 # Set regularization parameter (constant)
+ MetaModelOpts._kernel_noise = False # Optimize regularization parameter
+
+ # Bootstraping
+ # 1) normal 2) fast
+ # MetaModelOpts.bootstrap_method = 'fast' # ToDo: Remove this as a user-defined parameter, not used for GP?
+ MetaModelOpts.n_bootstrap_itrs = 1
+
+ # Print summary of the regression results
+ # MetaModelOpts.verbose = True
+
+ # ------------------------------------------------
+ # ------ Experimental Design Configuration -------
+ # ------------------------------------------------
+ ExpDesign = ExpDesigns(Inputs)
+
+ # One-shot (normal) or Sequential Adaptive (sequential) Design
+ ExpDesign.method = 'sequential'
+ ExpDesign.n_init_samples = 140 # 00#3*ndim
+
+ # Sampling methods
+ # 1) random 2) latin_hypercube 3) sobol 4) halton 5) hammersley
+ # 6) chebyshev(FT) 7) grid(FT) 8)user
+ ExpDesign.sampling_method = 'sobol'
+
+ # Provide the experimental design object with a hdf5 file
+ # ExpDesign.hdf5_file = 'ExpDesign_AnalyticFunc.hdf5'
+
+ # Set the sampling parameters
+ ExpDesign.n_new_samples = 100
+ ExpDesign.n_max_samples = 141 # 150 # sum of init + sequential
+ ExpDesign.mod_LOO_threshold = 1e-16
+
+ # ExpDesign.adapt_verbose = True
+ # 1) None 2) 'equal' 3)'epsilon-decreasing' 4) 'adaptive'
+ ExpDesign.tradeoff_scheme = None
+ # MetaModelOpts.ExpDesign.n_replication = 5
+ # -------- Exploration ------
+ # 1)'Voronoi' 2)'random' 3)'latin_hypercube' 4)'LOOCV' 5)'dual annealing'
+ ExpDesign.explore_method = 'random'
+
+ # Use when 'dual annealing' chosen
+ ExpDesign.max_func_itr = 1000
+
+ # Use when 'Voronoi' or 'random' or 'latin_hypercube' chosen
+ ExpDesign.n_canddidate = 1000
+ ExpDesign.n_cand_groups = 4
+
+ # -------- Exploitation ------
+ # 1)'BayesOptDesign' 2)'BayesActDesign' 3)'VarOptDesign' 4)'alphabetic'
+ # 5)'Space-filling'
+ ExpDesign.exploit_method = 'Space-filling'
+ ExpDesign.exploit_method = 'BayesActDesign'
+ ExpDesign.util_func = 'DKL'
+
+ # BayesOptDesign/BayesActDesign -> when data is available
+ # 1) MI (Mutual information) 2) ALC (Active learning McKay)
+ # 2)DKL (Kullback-Leibler Divergence) 3)DPP (D-Posterior-percision)
+ # 4)APP (A-Posterior-percision) # ['DKL', 'BME', 'infEntropy']
+ # ExpDesign.util_func = 'DKL'
+
+ # BayesActDesign -> when data is available
+ # 1) BME (Bayesian model evidence) 2) infEntropy (Information entropy)
+ # 2)DKL (Kullback-Leibler Divergence)
+ # ExpDesign.util_func = 'DKL'
+
+ # VarBasedOptDesign -> when data is not available
+ # 1)ALM 2)EIGF, 3)LOOCV
+ # or a combination as a list
+ # ExpDesign.util_func = 'EIGF'
+
+ # alphabetic
+ # 1)D-Opt (D-Optimality) 2)A-Opt (A-Optimality)
+ # 3)K-Opt (K-Optimality) or a combination as a list
+ # ExpDesign.util_func = 'D-Opt'
+
+ # Defining the measurement error, if it's known a priori
+ obsData = pd.DataFrame(Model.observations, columns=Model.Output.names)
+ DiscrepancyOpts = Discrepancy('')
+ DiscrepancyOpts.type = 'Gaussian'
+ DiscrepancyOpts.parameters = obsData ** 2
+ MetaModelOpts.Discrepancy = DiscrepancyOpts
+
+ # Plot the posterior snapshots for SeqDesign
+ ExpDesign.post_snapshot = False
+ ExpDesign.step_snapshot = 1
+ ExpDesign.max_a_post = [0] * ndim
+
+ # For calculation of validation error for SeqDesign
+ prior = np.load(f"data/Prior_{ndim}.npy")
+ prior_outputs = np.load(f"data/origModelOutput_{ndim}.npy")
+ likelihood = np.load(f"data/validLikelihoods_{ndim}.npy")
+ ExpDesign.valid_samples = prior[:500]
+ ExpDesign.valid_model_runs = {'Z': prior_outputs[:500]}
+
+ # Run using the engine
+ engine = Engine(MetaModelOpts, Model, ExpDesign)
+ # engine.train_sequential()
+ engine.train_normal()
+
+ # Load the objects
+ # with open(f"PCEModel_{Model.name}.pkl", "rb") as input:
+ # PCEModel = joblib.load(input)
+
+ # =====================================================
+ # ========= POST PROCESSING OF METAMODELS ===========
+ # =====================================================
+ PostPCE = PostProcessing(engine)
+
+ # Plot to check validation visually.
+ PostPCE.valid_metamodel(n_samples=1)
+
+ # Compute and print RMSE error
+ PostPCE.check_accuracy(n_samples=300)
+
+ # =====================================================
+ # ======== Bayesian inference with Emulator ==========
+ # =====================================================
+ BayesOpts = BayesInference(engine)
+ BayesOpts.emulator = True
+ BayesOpts.plot_post_pred = True
+
+ # BayesOpts.selected_indices = [0, 3, 5, 7, 9]
+ # BME Bootstrap
+ BayesOpts.bootstrap = True
+ BayesOpts.n_bootstrap_itrs = 500
+ BayesOpts.bootstrap_noise = 100
+
+ # Bayesian cross validation
+ BayesOpts.bayes_loocv = True # TODO: test what this does
+
+ # Select the inference method
+ import emcee
+
+ BayesOpts.inference_method = "MCMC"
+ # Set the MCMC parameters passed to self.mcmc_params
+ BayesOpts.mcmc_params = {
+ 'n_steps': 1e4, # 5,
+ 'n_walkers': 30,
+ 'moves': emcee.moves.KDEMove(),
+ 'multiprocessing': False,
+ 'verbose': False
+ }
+
+ # ----- Define the discrepancy model -------
+ obsData = pd.DataFrame(Model.observations, columns=Model.Output.names)
+ BayesOpts.measurement_error = obsData
+
+ # # -- (Option B) --
+ DiscrepancyOpts = Discrepancy('')
+ DiscrepancyOpts.type = 'Gaussian'
+ DiscrepancyOpts.parameters = obsData ** 2
+ BayesOpts.Discrepancy = DiscrepancyOpts
+
+ # -- (Option C) --
+ if 0:
+ DiscOutputOpts = Input()
+ # # # OutputName = 'Z'
+ DiscOutputOpts.add_marginals()
+ DiscOutputOpts.Marginals[0].Nnme = '$\sigma^2_{\epsilon}$'
+ DiscOutputOpts.Marginals[0].dist_type = 'uniform'
+ DiscOutputOpts.Marginals[0].parameters = [0, 10]
+ # BayesOpts.Discrepancy = {'known': DiscrepancyOpts,
+ # 'infer': Discrepancy(DiscOutputOpts)}
+
+ BayesOpts.bias_inputs = {'Z': np.arange(0, 10, 1.).reshape(-1, 1) / 9}
+
+ DiscOutputOpts = Input()
+ # OutputName = 'lambda'
+ DiscOutputOpts.add_marginals()
+ DiscOutputOpts.Marginals[0].name = '$\lambda$'
+ DiscOutputOpts.Marginals[0].dist_type = 'uniform'
+ DiscOutputOpts.Marginals[0].parameters = [0, 1]
+
+ # # OutputName = 'sigma_f'
+ DiscOutputOpts.add_marginals()
+ DiscOutputOpts.Marginals[1].Name = '$\sigma_f$'
+ DiscOutputOpts.Marginals[1].dist_type = 'uniform'
+ DiscOutputOpts.Marginals[1].parameters = [0, 1e-4]
+ # BayesOpts.Discrepancy = Discrepancy(DiscOutputOpts)
+ BayesOpts.Discrepancy = {'known': DiscrepancyOpts,
+ 'infer': Discrepancy(DiscOutputOpts)}
+
+ # Start the calibration/inference
+ Bayes_PCE = BayesOpts.create_inference()
+
+ # Save class objects
+ with open(f'Bayes_{Model.name}.pkl', 'wb') as output:
+ joblib.dump(Bayes_PCE, output, 2)
diff --git a/src/bayesvalidrox/__init__.py b/src/bayesvalidrox/__init__.py
index 55cfa8ed1..a824212ed 100644
--- a/src/bayesvalidrox/__init__.py
+++ b/src/bayesvalidrox/__init__.py
@@ -4,6 +4,7 @@ __version__ = "1.1.0"
from .pylink.pylink import PyLinkForwardModel
from .surrogate_models.surrogate_models import MetaModel
from .surrogate_models.polynomial_chaos import PCE
+from .surrogate_models.gaussian_process_sklearn import GPESkl
from .surrogate_models.engine import Engine
from .surrogate_models.inputs import Input
from .surrogate_models.exp_designs import ExpDesigns
@@ -23,5 +24,6 @@ __all__ = [
"ExpDesigns",
"PostProcessing",
"BayesInference",
- "BayesModelComparison"
+ "BayesModelComparison",
+ "GPESkl"
]
diff --git a/src/bayesvalidrox/surrogate_models/gaussian_process_sklearn.py b/src/bayesvalidrox/surrogate_models/gaussian_process_sklearn.py
new file mode 100644
index 000000000..e50cc20ad
--- /dev/null
+++ b/src/bayesvalidrox/surrogate_models/gaussian_process_sklearn.py
@@ -0,0 +1,466 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+"""
+Implementation of metamodel as GPE, using the Scikit-Learn library
+"""
+
+import copy
+import math
+import os
+import warnings
+
+import functools
+import matplotlib.pyplot as plt
+import numpy as np
+import sklearn.gaussian_process.kernels as kernels
+from joblib import Parallel, delayed
+from sklearn.gaussian_process import GaussianProcessRegressor
+from sklearn.preprocessing import MinMaxScaler, StandardScaler
+from tqdm import tqdm
+
+from .surrogate_models import MetaModel, _preprocessing_fit, _bootstrap_fit, _preprocessing_eval, _bootstrap_eval
+
+warnings.filterwarnings("ignore")
+# warnings.simplefilter('default')
+# Load the mplstyle
+# noinspection SpellCheckingInspection
+plt.style.use(os.path.join(os.path.split(__file__)[0],
+ '../', 'bayesvalidrox.mplstyle'))
+
+
+class MySklGPE(GaussianProcessRegressor):
+ """
+ GP ScikitLearn class, to change the default values for maximum iterations and optimization tolerance.
+ """
+ def __init__(self, max_iter=10_000, gtol=1e-6, **kwargs):
+ super().__init__(**kwargs)
+ self.max_iter = max_iter
+ self.gtol = gtol
+
+
+class GPESkl(MetaModel):
+ """
+ GP MetaModel using the Scikit-Learn library
+
+ This class trains a surrogate model of type Gaussian Process Regression.
+ It accepts an input object (input_obj)
+ containing the specification of the distributions for uncertain parameters
+ and a model object with instructions on how to run the computational model.
+
+ Attributes
+ ----------
+ input_obj : obj
+ Input object with the information on the model input parameters.
+ _meta_model_type : str
+ Surrogate model types, in this case GPE.
+ Default is GPE.
+ gpe_reg_method : str
+ GPE regression method to compute the kernel hyperparameters. The following
+ regression methods are available for Scikit-Learn library
+ 1. LBFGS:
+ Default is `LBFGS`.
+ autoSelect: bool
+ Flag to loop through different available Kernels and select the best one based on BME criteria.
+ Default is False.
+ kernel_type: str
+ Type of kernel to use and train for. The following Scikit-Learn kernels are available:
+ 1. RBF: Squared exponential kernel
+ 2. Matern: Matern kernel
+ 3. RQ: rational quadratic kernel
+ Default is `'RBF'` kernel.
+ isotropy: bool
+ Flat to train an isotropic kernel (one length scale for all input parameters) or
+ an anisotropic kernel (a length scale for each dimension). True for isotropy, False for anisotropic kernel
+ Default is True
+ noisy: bool
+ Consider a WhiteKernel for regularization purposes, and optimize for the noise hyperparameter.
+ Default is False
+ nugget: float
+ Constant value added to the Kernel matrix for regularization purposes (not optimized)
+ Default is 1e-9
+
+ bootstrap_method : str
+ Bootstraping method. Options are `'normal'` and `'fast'`. The default
+ is `'fast'`. It means that in each iteration except the first one, only
+ the coefficent are recalculated with the ordinary least square method.
+ n_bootstrap_itrs : int
+ Number of iterations for the bootstrap sampling. The default is `1`.
+ dim_red_method : str
+ Dimensionality reduction method for the output space. The available
+ method is based on principal component analysis (PCA). The Default is
+ `'no'`. There are two ways to select number of components: use
+ percentage of the explainable variance threshold (between 0 and 100)
+ (Option A) or direct prescription of components' number (Option B):
+ >>> MetaModelOpts = MetaModel()
+ >>> MetaModelOpts.dim_red_method = 'PCA'
+ >>> MetaModelOpts.var_pca_threshold = 99.999 # Option A
+ >>> MetaModelOpts.n_pca_components = 12 # Option B
+ verbose : bool
+ Prints summary of the regression results. Default is `False`.
+ """
+
+ def __init__(self, input_obj, meta_model_type='GPE', gpe_reg_method='lbfgs',
+ autoSelect=False, kernel_type='RBF', isotropy=True, noisy=False, nugget=1e-9,
+ bootstrap_method='fast', # ToDo: Remove this as a user-defined parameter, not used for GP?
+ n_bootstrap_itrs=1,
+ dim_red_method='no', verbose=False):
+
+ # Check if the surrogate outputs gaussian results: Always TRUE for GPs
+ is_gaussian = True # self.check_is_gaussian(n_bootstrap_itrs)
+
+ # Use parent init
+ super().__init__(input_obj, meta_model_type, bootstrap_method,
+ n_bootstrap_itrs, dim_red_method, is_gaussian, verbose)
+
+ # Additional inputs
+ # Parameters that are not needed from the outside are denoted with '_'
+ self.meta_model_type = meta_model_type
+ self._gpe_reg_method = gpe_reg_method
+ self.regression_dict = {}
+ self._gpe_reg_options = {}
+
+ self._autoSelect = autoSelect
+ self._kernel_isotropy = isotropy
+ self._kernel_noise = noisy
+ self._kernel_type = kernel_type
+ self._nugget = nugget
+
+ # Is currently needed inputs for 'engine', but not for the GPE class:
+ self._pce_deg = 1
+
+ # Other params
+ self._gp_poly = None
+ self._x_scaler = None
+ self._bme_score = None
+ self._kernel_name_dict = None
+ # self._LCerror = None
+
+ # Initialize the regression options as a dictionary
+
+ def check_is_gaussian(self, n_bootstrap_itrs) -> bool:
+ """
+ Check if the metamodel returns a mean and stdev.
+
+ Returns
+ -------
+ bool
+ TRUE
+
+ """
+ return True
+
+ def build_metamodel(self) -> None:
+ """
+ Builds the parts for the metamodel (,...) that are needed before fitting.
+ This is executed outside any loops related to e.g. bootstrap or
+ transformations such as pca.
+
+ Returns
+ -------
+ None
+
+ """
+ # Initialize the nested dictionaries
+ self._gp_poly = self.auto_vivification()
+ self._x_scaler = self.auto_vivification()
+ self._bme_score = self.auto_vivification()
+ self._kernel_name_dict = self.auto_vivification()
+ # self._LCerror = self.auto_vivification()
+
+ def build_kernels(self):
+ """
+ Initializes the different possible kernels, and selects the ones to train for,
+ depending on the input options.
+ ToDo: Add additional kernels
+ ToDo: Add option to include user-defined kernel
+ Returns
+ -------
+ List: with the kernels to iterate over
+ List: with the names of the kernels to iterate over
+ """
+ _ndim = self.InputSpace.ndim
+
+ # Set boundaries for length scales:
+ value = np.empty((), dtype=object)
+ value[()] = [1e-5, 1e5]
+
+ if self._kernel_isotropy:
+ # Isotropic Kernel
+ ls_bounds = list(np.full(1, value, dtype=object))
+ ls_values = 1
+ else:
+ ls_bounds = list(np.full(_ndim, value, dtype=object))
+ ls_values = list(np.full(_ndim, 1))
+
+ # Generate list of probable kernels:
+ rbf_kernel = 1*kernels.RBF(length_scale=ls_values,
+ length_scale_bounds=ls_bounds)
+ matern_kernel = 1*kernels.Matern(length_scale=ls_values,
+ length_scale_bounds=ls_bounds,
+ nu=1.5)
+ rq_kernel = 1*kernels.RationalQuadratic(length_scale=ls_values,
+ length_scale_bounds=ls_bounds,
+ alpha=1)
+ kernel_dict = {'RBF': rbf_kernel,
+ 'Matern': matern_kernel,
+ 'RQ': rq_kernel}
+
+ if self._autoSelect:
+ kernel_list = list(kernel_dict.values())
+ kernel_names = list(kernel_dict.keys())
+ else:
+ try:
+ kernel_list = [kernel_dict[self._kernel_type]]
+ kernel_names = [self._kernel_type]
+ except:
+ print(f'The kernel option {self._kernel_type} is not available. An RBF kernel was chosen instead')
+ kernel_list = [kernel_dict['RBF']]
+ kernel_names = ['RBF']
+
+ return kernel_list, kernel_names
+
+ @staticmethod
+ def transform_x(X: np.array, transform_type='norm'):
+ """
+ Scales the inputs (X) during training using either normalize ([0, 1]), or standardize (N[0, 1]).
+ If None, then the inputs are not scaled
+ Parameters
+ ----------
+ X: 2D list or np.array of shape (#samples, #dim)
+ The parameter value combinations to train the model with.
+ transform_type: str
+ Transformation to apply to the input parameters. Default is None
+
+ Returns
+ -------
+ np.array: (#samples, #dim)
+ transformed input parameters
+ obj: Scaler object
+ transformation object, for future transformations during surrogate evaluation
+
+ """
+ if transform_type is None:
+ scaler = None
+ X_S = X
+ else:
+ if transform_type.lower() == 'norm':
+ scaler = MinMaxScaler()
+ X_S = scaler.fit_transform(X)
+ elif transform_type.lower() == 'standard':
+ scaler = StandardScaler()
+ X_S = scaler.fit_transform(X)
+ else:
+ print(f'No scaler {transform_type} found. No scaling was done')
+ scaler = None
+ X_S = X
+
+ return X_S, scaler
+
+ @_preprocessing_fit
+ @_bootstrap_fit
+ def fit(self, X: np.array, y: dict, parallel=False, verbose=False, b_i=0):
+ """
+ Fits the surrogate to the given data (samples X, outputs y).
+ Note here that the samples X should be the transformed samples provided
+ by the experimental design if the transformation is used there.
+
+ Parameters
+ ----------
+ X : 2D list or np.array of shape (#samples, #dim)
+ The parameter value combinations that the model was evaluated at.
+ y : dict of 2D lists or arrays of shape (#samples, #timesteps)
+ The respective model evaluations.
+ parallel : bool
+ Set to True to run the training in parallel for various keys.
+ The default is False.
+ verbose : bool
+ Set to True to obtain more information during runtime.
+ The default is False.
+
+ Returns
+ -------
+ None.
+
+ """
+
+ # For loop over the components/outputs
+ if self.verbose and self.n_bootstrap_itrs == 1:
+ items = tqdm(y.items(), desc="Fitting regression")
+ else:
+ items = y.items()
+
+ # Transform inputs:
+ # ToDO: Do we do this although the inputs are already transformed in the preprocessing decorator?
+ # Todo: Is this better here or in the adaptive_regression function?
+ X_S, scaler = self.transform_x(X=X) #, transform_type=None)
+ self._x_scaler[f'b_{b_i + 1}'] = scaler
+
+ for key, output in items:
+ # Parallel fit regression
+ out = None
+ if parallel: # and (not self.fast_bootstrap or b_i == 0):
+ out = Parallel(n_jobs=-1, backend='multiprocessing')(
+ delayed(self.adaptive_regression)(X_S,
+ output[:, idx],
+ idx, self.verbose)
+ for idx in range(output.shape[1]))
+ elif not parallel: # and (not self.fast_bootstrap or b_i == 0):
+ results = map(functools.partial(self.adaptive_regression, X_S, verbose=self.verbose),
+ [output[:, idx] for idx in
+ range(output.shape[1])],
+ range(output.shape[1]))
+ out = list(results)
+
+ # Create a dict to pass the variables
+ for i in range(output.shape[1]):
+ self._gp_poly[f'b_{b_i + 1}'][key][f"y_{i + 1}"] = out[i]['gp']
+ self._bme_score[f'b_{b_i + 1}'][key][f"y_{i + 1}"] = out[i]['BME']
+ self._kernel_name_dict[f'b_{b_i + 1}'][key][f"y_{i + 1}"] = out[i]['kernel_name']
+
+ # --------------------------------------------------------------------------------------------------------
+ def adaptive_regression(self, X, y, varIdx, verbose=False):
+ """
+ Adaptively fits the GPE model by comparing different Kernel options
+
+ Parameters
+ ----------
+ X : array of shape (n_samples, ndim)
+ Training set. These samples should be already transformed.
+ y : array of shape (n_samples,)
+ Target values, i.e. simulation results for the Experimental design.
+ varIdx : int
+ Index of the output.
+ verbose : bool, optional
+ Print out summary. The default is False.
+
+ Returns
+ -------
+ returnVars : Dict
+ Fitted estimator, BME score
+
+ """
+
+ # Initialization
+ gp_list = {}
+ BME = []
+
+ # Get kernels:
+ kernel_list, kernel_names = self.build_kernels()
+
+ for i, K in enumerate(kernel_list):
+ if self._kernel_noise:
+ K = K + kernels.WhiteKernel(noise_level=np.std(y)/math.sqrt(2))
+
+ # ToDo: Check if we want to increase number of iterations/tolerance. Set n_restars as variable?
+ # gp_list[i] = MySklGPE(kernel=K,
+ # n_restarts_optimizer=20,
+ # alpha=self._nugget,
+ # normalize_y=True)
+ #
+ gp_list[i] = GaussianProcessRegressor(kernel=K,
+ n_restarts_optimizer=10,
+ alpha=self._nugget,
+ normalize_y=True)
+
+ # Fit to data using Maximum Likelihood Estimation
+ gp_list[i].fit(X, y)
+
+ # Store the MLE as BME score
+ BME.append(gp_list[i].log_marginal_likelihood())
+
+ # Select the GP with the highest BME
+ idx_max = np.argmax(BME)
+ gp = gp_list[idx_max]
+
+ if varIdx is not None and verbose:
+ gp_score = gp.score(X, y)
+ print('=' * 50)
+ print(' ' * 10 + ' Summary of results ')
+ print('=' * 50)
+
+ print(f'Output variable {varIdx}:')
+ print('The estimation of GPE coefficients converged,')
+ print(f'with the R^2 score: {gp_score:.3f} ')
+ print(f'using a {kernel_names[idx_max]} Kernel')
+ print('=' * 50)
+
+ # Create a dict to pass the outputs
+ # ToDo: Save the Kernel hyperparameters?
+ returnVars = {}
+ returnVars['gp'] = gp
+ returnVars['BME'] = np.argmax(BME)
+ returnVars['kernel_name'] = kernel_names[idx_max]
+
+ return returnVars
+
+ # -------------------------------------------------------------------------
+ @staticmethod
+ def scale_x(X: np.array, transform_obj: object):
+ """
+ Transforms the inputs based on the scaling done during training
+ Parameters
+ ----------
+ X: 2D list or np.array of shape (#samples, #dim)
+ The parameter value combinations to evaluate the model with.
+ transform_obj: Scikit-Learn object
+ Class instance to transform inputs
+
+ Returns
+ -------
+ np.array (#samples, #dim)
+ Transformed input sets
+ """
+ if transform_obj is None:
+ x_t = X
+ else:
+ x_t = transform_obj.transform(X)
+ return x_t
+
+ @_preprocessing_eval
+ @_bootstrap_eval
+ def eval_metamodel(self, samples, b_i=0):
+ """
+ Evaluates GP metamodel at the requested samples.
+
+ Parameters
+ ----------
+ samples : array of shape (n_samples, ndim), optional
+ Samples to evaluate metamodel at. The default is None.
+
+ Returns
+ -------
+ mean_pred : dict
+ Mean of the predictions.
+ std_pred : dict
+ Standard deviation of the predictions.
+ """
+ # Transform the input parameters
+ samples_sc = self.scale_x(X=samples, transform_obj=self._x_scaler[f'b_{b_i + 1}'])
+
+ # Extract model dictionary
+ model_dict = self._gp_poly[f'b_{b_i + 1}']
+
+ # Loop over outputs
+ mean_pred = {}
+ std_pred = {}
+
+ for output, values in model_dict.items():
+ mean = np.empty((len(samples), len(values)))
+ std = np.empty((len(samples), len(values)))
+ idx = 0
+ for in_key, _ in values.items():
+
+ gp = self._gp_poly[f'b_{b_i + 1}'][output][in_key]
+
+ # Prediction
+ y_mean, y_std = gp.predict(samples_sc, return_std=True)
+
+ mean[:, idx] = y_mean
+ std[:, idx] = y_std
+ idx += 1
+
+ mean_pred[output] = mean
+ std_pred[output] = std
+
+ return mean_pred, std_pred
+
--
GitLab