timedecoder_utils module¶

timedecoder_utils.decode_time_step(X_i, y, y_dict, clf, scaler, n_splits=5, sampler=[], do_pca=False, n_components=0.95, permute_train_labels=False, compute_auc=False)¶

Run a classification of data X_i with labels y using a Stratified K-Folds cross-validation where n_splits is the number of folds. Classification pipeline, for each one of the Stratified K-Folds :

1. If do_pca is True, first do a standardization scaling and then apply PCA and keep only the first n_components
2. Apply the scaler. Fit the transform on the training set and apply it on the test set
3. Resample the train set so that the train set is balanced between classes, using the sampler argument
4. Train the classifier on the train set (If permute_train_labels is True, permute the training labels first)
5. Predict labels on the test set
6. Evaluate the performances with several scores

See also timedecoder.TimeDecoder.decode()

Parameters:

X_i : 2D array [n_features, n_trials] Feature array y : array [n_trials] Label of each trial in X_i y_dict : dict Label dictionnary clf : sklearn classifier Scikit-Learn classifier scaler : sklearn scaler Scikit-Learn scaler n_splits : int Number of folds for the Stratified K-Folds cross-validation sampler : sklearn sampler Scikit learn sampler - see timedecoder.get_sampler() do_pca : bool (default: False) If True, apply a PCA decomposition and keep only a certain amount of components given n_components n_components : int | float Number of components to keep when running the PCA If 0 < n_components < 1 and svd_solver == 'full', select the number of components such that the amount of variance that needs to be explained is greater than the percentage specified by n_components. permute_train_labels : bool (default: False) If True, permute the training labels. Used for computing the chance level. compute_auc : bool (default: False) If True, will try to compute the Area Under the Curve score

Returns:

accuracy_mean : float Accuracy, mean over the K-folds f1_mean : float F1-score, mean over the K-folds precision_mean : float Precision, mean over the K-folds recall_mean : float Recall, mean over the K-folds class_acc_mean : array [n_classes] Class-accuracy for each class, mean over the K-folds auc_mean : float Area under the curve, mean over the K-folds

timedecoder_utils.temporal_generalization_step(X, i_time, y, y_dict, clf, scaler, n_splits=5, sampler=[], do_pca=False, n_components=0.95, permute_train_labels=False)¶

Temporal generalization step. Train at time i_time and test for all time points. Use a stratified K-folds cross validation.

Pipeline, for each stratified K-folds:

1. If do_pca is True, first do a standardization scaling and then apply PCA and keep only the first n_components on the train set
2. Apply the scaler. Fit the transform on the training set
3. Resample the train set so that the train set is balanced between classes, using the sampler argument
4. Train the classifier on the train set (If permute_train_labels is True, permute the training labels first)
5. For each time point :

• If do_pca is True, do a standardization scaling and then apply PCA on the test set
• Apply the scaler on the test set
• Predict labels on the test set
• Evaluate the performances with several scores

See also timedecoder.TimeDecoder.temporal_generalization()

Parameters:

X : 3D array [n_features, n_pnts, n_trials] Temporal feature array i_time : int Training time point y : array [n_trials] Label array y_dict : dict Label dictionnary clf : sklearn classifier Scikit-Learn classifier scaler : sklearn scaler Scikit-Learn scaler n_splits : int Number of folds for the Stratified K-Folds cross-validation sampler : sklearn sampler Scikit learn sampler - see timedecoder.get_sampler() do_pca : bool (default: False) If True, apply a PCA decomposition and keep only a certain amount of components given n_components n_components : int | float Number of components to keep when running the PCA If 0 < n_components < 1 and svd_solver == 'full', select the number of components such that the amount of variance that needs to be explained is greater than the percentage specified by n_components. permute_train_labels : bool (default: False) If True, permute the training labels. Used for computing the chance level.

Returns:

accuracy_mean : float Accuracy, mean over the K-folds class_acc_mean : array [n_classes] Class-accuracy for each class, mean over the K-folds