Allows to transpose metamodels learned thanks to the tensap package (https://anthony-nouy.github.io/tensap/) to OpenTURNS functions.
Example:
fun, X = tensap.multivariate_functions_benchmark('borehole')
ORDER = X.size
DEGREE = 8
BASES = [tensap.PolynomialFunctionalBasis(X_TRAIN.orthonormal_polynomials(),
range(DEGREE+1)) for X_TRAIN in X.random_variables]
BASES = tensap.FunctionalBases(BASES)
NUM_TRAIN = 1000
X_TRAIN = X.random(NUM_TRAIN)
Y_TRAIN = fun(X_TRAIN)
NUM_TEST = 10000
X_TEST = X.random(NUM_TEST)
Y_TEST = fun(X_TEST)
SOLVER = tensap.CanonicalTensorLearning(ORDER, tensap.SquareLossFunction())
SOLVER.rank_adaptation = True
SOLVER.initialization_type = 'mean'
SOLVER.tolerance['on_error'] = 1e-6
SOLVER.alternating_minimization_parameters['stagnation'] = 1e-8
SOLVER.alternating_minimization_parameters['max_iterations'] = 100
SOLVER.linear_model_learning.regularization = False
SOLVER.linear_model_learning.basis_adaptation = True
SOLVER.bases = BASES
SOLVER.training_data = [X_TRAIN, Y_TRAIN]
SOLVER.display = True
SOLVER.alternating_minimization_parameters['display'] = False
SOLVER.test_error = True
SOLVER.test_data = [X_TEST, Y_TEST]
SOLVER.alternating_minimization_parameters['one_by_one_factor'] = False
SOLVER.alternating_minimization_parameters['inner_loops'] = 2
SOLVER.alternating_minimization_parameters['random'] = False
SOLVER.rank_adaptation_options['max_iterations'] = 4
SOLVER.model_selection = True
SOLVER.model_selection_options['type'] = 'test_error'
F, OUTPUT = SOLVER.solve()
# to the OT world
otf = ottensap.TensapFunction(F)
x = X.mean()
y = otf(x)