Scikit Modules

Scikit Modules#

class brainmaze_eeg.scikit_modules.FeatureAugmentorModule#

Feature augmentation using an ‘augment_features’ function from the ‘PiesUtils’ package. See the code for additional details.

fit(X=None, Y=None)#

Fit method (no-op for this transformer).

fit_transform(X, Y=None)#

Fit and transform in one step (same as transform for this module).

transform(X)#

Transform features by applying mutual and standalone operations.

Parameters:

X (np.ndarray) – Input feature matrix of shape [n_samples, n_features].

Returns:

Augmented feature matrix.

Return type:

np.ndarray

class brainmaze_eeg.scikit_modules.Log10Module#

Base-10 logarithm transformation module compatible with scikit-learn pipelines.

fit(X, Y=None)#

Fit method (no-op for this transformer).

fit_transform(X, Y=None)#

Fit and transform in one step.

transform(X, Y=None)#

Apply base-10 logarithm transformation to features.

class brainmaze_eeg.scikit_modules.LogModule#

Natural logarithm transformation module compatible with scikit-learn pipelines.

fit(X, Y=None)#

Fit method (no-op for this transformer).

fit_transform(X, Y=None)#

Fit and transform in one step.

transform(X, Y=None)#

Apply natural logarithm transformation to features.

class brainmaze_eeg.scikit_modules.PCAModule(var_threshold=0.98)#

PCA module using scikit-learn’s PCA with automatic component selection. Extends sklearn.decomposition.PCA.

fit(X, y=None)#

Fit the model with X.

Parameters:

  • X ({array-like, sparse matrix} of shape (n_samples, n_features)) – Training data, where n_samples is the number of samples and n_features is the number of features.

  • y (Ignored) – Ignored.

Returns:

self – Returns the instance itself.

Return type:

object

fit_transform(X, y=None)#

Fit and transform data in one step.

Parameters:

  • X (np.ndarray) – Input data matrix.

  • y (array-like, optional) – Target values (ignored, for scikit-learn compatibility).

Returns:

Transformed data in PCA space.

Return type:

np.ndarray

class brainmaze_eeg.scikit_modules.PCAModuleSVD(var_threshold=0.98)#

PCA via eigendecomposition of the covariance matrix, with automatic component selection by an explained-variance threshold. Compatible with scikit-learn pipelines.

Note

Despite the name, this does not use an SVD: it eigendecomposes the sample covariance C = X.T @ X / (n - 1). It does not centre X; it assumes the input is already zero-mean per feature (e.g. the output of a z-score step). On non-centred data the “covariance” is the second-moment matrix and the components are not principal components. Use PCAModule (scikit-learn, centres the data) if that assumption does not hold.

eigen_vals#

Set by fit. Eigenvalues of C (variance along each component), real and sorted in descending order. Tiny negative values from round-off are clipped to 0.

Type:

np.ndarray, shape (n_features,), float

eigen_vecs#

Matching unit eigenvectors in columns (eigen_vecs[:, i] <-> eigen_vals[i]). Sign convention (like scikit-learn’s svd_flip): the largest-magnitude loading of each column is positive (the first one on ties), so the projections do not depend on the LAPACK build.

Type:

np.ndarray, shape (n_features, n_features), float

explained_variance_ratio#

eigen_vals / eigen_vals.sum().

Type:

np.ndarray, shape (n_features,)

n#

Number of retained components: the smallest n such that the cumulative explained-variance ratio of the first n components is >= var_threshold (clipped to n_features).

Type:

int

fit(X, Y=None)#

Fit the components on X.

Parameters:

  • X (np.ndarray, shape (n_samples, n_features)) – Zero-mean data (see the class note; X is not centred here).

  • Y (ignored)

Return type:

self

Notes

Up to v1.0.0 this used np.linalg.eig (general, non-symmetric solver) and took the first n eigenpairs as if they were in descending order. eig does not guarantee any order (e.g. it returned [.., 0.896, 0.292, 0.589, 0.449] for an 8-feature covariance) and returns complex dtype, so the selected components and n could be wrong. It now uses np.linalg.eigh (symmetric solver, real output) and sorts explicitly in descending order (issue #39).

fit_transform(X, Y=None)#

Fit and transform in one step.

transform(X, Y=None)#

Transform data using fitted PCA components.

Parameters:

  • X (np.ndarray) – Input data matrix.

  • Y (array-like, optional) – Ignored parameter for scikit-learn compatibility.

Returns:

Transformed data in PCA space.

Return type:

np.ndarray

class brainmaze_eeg.scikit_modules.ZScoreModule(trainable=False, continuous_learning=False, multi_class=False)#

Z-score normalization compatible with scikit.pipeline.Pipeline Enables continuous learning - enabling continuous adaptation.

Modes

  • Zscore normalization

  • Zscore normalization with fixed mean and std values based on the initial training dataset

    • Possible category-wise normalization with mean and std values estimated from the training dataset - number of features is multiplied by number of categories

  • Zscore normalization with an initial mean and std values trained on the training dataset - adaptation during inference

    https://stats.stackexchange.com/questions/211837/variance-of-subsample

continuous_learning#

If true - An instance updates mean and variance values during each prediction step. Initial outlier filtering is recommended

Type:

bool

trainable#

If false - An instance normalizes inference data based on their current mean value and std If true - An instance remembers mean and variance values of training data

Type:

bool

multi_class#

If true - An instance performs normalization for each training class separately Number of output features is multiplied by a number of training categories

Type:

bool

mean#

Trained mean values for each feature. In case multi_class == True -> list of numpy ndarrays for each category

Type:

numpy ndarray / list

std#
Type:

numpy ndarray

N#
Type:

int

fit(X=None, Y=None)#
Parameters:

  • X (numpy ndarray) – shape[n_samples, n_features]

  • Y (list or numpy array, optional) – category reference for each sample - required only for option with multi_class normalization

Return type:

None

fit_transform(X=None, Y=None)#
Parameters:

  • X (numpy ndarray) – shape[n_samples, n_features]

  • Y (list or numpy array, optional) – category reference for each sample - required only for option with multi_class normalization

Returns:

transformed_data – shape[n_samples, n_features]

Return type:

numpy ndarray

transform(X=None)#
Parameters:

X (numpy ndarray) – shape[n_samples, n_features]

Returns:

transformed_data – shape[n_samples, n_features]

Return type:

numpy ndarray