Source code for

.. module:: PerformanceEvaluationXVal
   :synopsis: Best parameters estimation with Cross-Validation

.. moduleauthor:: Marco Melis <>
.. moduleauthor:: Ambra Demontis <>

from import CPerfEvaluator
from secml.array import CArray
from secml.core.type_utils import is_scalar

[docs]class CPerfEvaluatorXVal(CPerfEvaluator): """Evaluate the best estimator parameters using Cross-Validation. Parameters ---------- splitter : CXVal or str XVal object to be used for splitting the dataset into train and validation. metric : CMetric or str Name of the metric that we want maximize / minimize. Attributes ---------- class_type : 'xval' """ __class_type = 'xval'
[docs] def compute_performance(self, estimator, dataset): """Split data in folds and return the mean estimator performance. Parameters ---------- estimator : CClassifier The Classifier that we want evaluate dataset : CDataset Dataset that we want use for evaluate the classifier Returns ------- score : float Mean performance score of estimator computed on the K-Folds. """ # Placeholder for folds' score fold_number = len(self.splitter.tr_idx) splits_score = CArray.zeros(fold_number) # estimate the performance of the estimator on each fold for split_idx in range(fold_number): train_dataset = dataset[self.splitter.tr_idx[split_idx], :] test_dataset = dataset[self.splitter.ts_idx[split_idx], :] # Train the estimator, train_dataset.Y) pred_label, pred_score = estimator.predict( test_dataset.X, return_decision_function=True) if dataset.num_classes > 2: pred_score = None # Score cannot be used in multiclass case else: # Extracting score of the positive class pred_score = pred_score[:, 1].ravel() this_test_score = self.metric.performance_score( test_dataset.Y, y_pred=pred_label, score=pred_score) splits_score[split_idx] = this_test_score return splits_score.mean()
def _get_best_params(self, res_vect, params, params_matrix, pick='first'): """Returns the best parameters given input performance scores. The best parameters have the closest associated performance score to the metric's best value. Parameters ---------- res_vect : CArray Array with the performance results associated to each parameters combination. params : dict Dictionary with the parameters to be evaluated. params_matrix : CArray Indices of each combination of parameters to evaluate. pick : {'first', 'last', 'random'}, optional Defines which of the best parameters set pick. Usually, 'first' (default) correspond to the smallest parameters while 'last' correspond to the biggest. The order is consistent to the parameters dict passed as input. Returns ------- best_params_dict : dict Dictionary with the parameters that have obtained the best performance score. best_value : any Performance value associated with the best parameters. """ if not is_scalar(self.metric.best_value): raise TypeError( "XVal only works with metric with the best value as scalar") # Get the index of the results closest to the best value diff = abs(res_vect - self.metric.best_value) condidates_idx = diff.find(diff == diff.nanmin()) if len(condidates_idx) < 1: raise ValueError("all metric outputs are equal to Nan!") # Get the value of the result closest to the best value best_score = res_vect[condidates_idx[0]] # Get the index of the corresponding parameters best_params_idx = params_matrix[condidates_idx, :] # Build the list of candidate parameters best_params_list = [] for c_idx in range(best_params_idx.shape[0]): # For each candidate get corresponding parameters best_params_dict = dict() for j, par in enumerate(params): value_idx = best_params_idx[c_idx, j].item() best_params_dict[par] = params[par][value_idx] best_params_list.append(best_params_dict) # Chose which candidate parameters assign to classifier if pick == 'first': # Usually the smallest best_params_dict = best_params_list[0] elif pick == 'last': # Usually the biggest best_params_dict = best_params_list[-1] elif pick == 'random': import random best_params_dict = random.choice(best_params_list) else: raise ValueError("pick strategy '{:}' not known".format(pick)) return best_params_dict, best_score