coco_pipe.transforms.subject_alignment#

Experimental: subject-axis removal for frozen embedding vectors.

Attributes#

`VECTOR_TRANSFORMS`
`TOKEN_TRANSFORMS`

Classes#

`SubjectAlignment`	Base class for subject-alignment estimators.
`LeaceEraser`	Fold-local LEACE projection of the between-subject linear subspace.
`EuclideanAlign`	Per-subject mean or CORAL alignment, including unseen subjects.
`RiemannAlign`	Per-subject Riemannian recentering followed by tangent projection.

Functions#

`tokens_to_covariances`(tokens, *[, shrinkage])	Convert `(window, token, feature)` tensors to regularized SPD matrices.
`make_subject_transform`(name, **params)	Instantiate a registered subject transform by name.

Module Contents#

class coco_pipe.transforms.subject_alignment.SubjectAlignment#

Bases: sklearn.base.BaseEstimator, sklearn.base.TransformerMixin

Base class for subject-alignment estimators.

name = 'identity'#

fold_local: bool = False#

fingerprint()#

Return type:: str

class coco_pipe.transforms.subject_alignment.LeaceEraser(*, shrinkage=True, degenerate_frac=0.95)#

Bases: SubjectAlignment

Fold-local LEACE projection of the between-subject linear subspace.

Parameters:

shrinkage (bool)
degenerate_frac (float)

name = 'leace'#

fold_local = True#

shrinkage = True#

degenerate_frac = 0.95#

fit(X, y=None, groups=None)#

transform(X, groups=None)#

fit_transform(X, y=None, groups=None, **fit_params)#: Fit the LEACE projector and transform the same observations.

class coco_pipe.transforms.subject_alignment.EuclideanAlign(mode='coral', *, shrinkage=True)#

Bases: SubjectAlignment

Per-subject mean or CORAL alignment, including unseen subjects.

Parameters:

mode (str)
shrinkage (bool)

mode = 'coral'#

name = 'ea_coral'#

shrinkage = True#

fit(X, y=None, groups=None)#

transform(X, groups=None)#

fit_transform(X, y=None, groups=None, **fit_params)#: Fit subject statistics and transform the same observations.

coco_pipe.transforms.subject_alignment.tokens_to_covariances(tokens, *, shrinkage=0.1)#

Convert (window, token, feature) tensors to regularized SPD matrices.

Parameters:

tokens (numpy.ndarray)
shrinkage (float)

Return type:

numpy.ndarray

class coco_pipe.transforms.subject_alignment.RiemannAlign(*, shrinkage=0.1)#

Bases: SubjectAlignment

Per-subject Riemannian recentering followed by tangent projection.

Parameters:: shrinkage (float)

name = 'ra'#

shrinkage = 0.1#

fit(X, y=None, groups=None)#

transform(X, groups=None)#

fit_transform(X, y=None, groups=None, **fit_params)#: Fit subject references and transform the same token windows.

coco_pipe.transforms.subject_alignment.VECTOR_TRANSFORMS#

coco_pipe.transforms.subject_alignment.TOKEN_TRANSFORMS#

coco_pipe.transforms.subject_alignment.make_subject_transform(name, **params)#

Instantiate a registered subject transform by name.

Parameters:

name (str)
params (Any)

Return type:

SubjectAlignment