IndexingTerm

class liesel_gam.IndexingTerm(basis, penalty, scale, name='', inference=None, coef_name=None, _update_on_init=True, validate_scalar_scale=True)

Bases: StrctTerm

Term object for memory-efficient representation of sparse bases.

Derived from StrctTerm. If the basis matrix of a term is a dummy matrix, where each column consists only of binary (0/1) entries, and each row has only one non-zero entry, then it is not necessary to store the full matrix in memory and evaluate the term as a dot product basis @ coef.

Instead, we can simply store a 1-D array of indices, identifying the nonzero column for each row of the basis matrix, and use this index to access the corresponding coefficient. This scenario is common for independent random intercepts.

This class implements such a sparse representation.

In case you do need to materialize the full, sparse basis of such a term, you can use IndexingTerm.init_full_basis().

Examples

>>> basis = Basis(jnp.array([0, 1, 0, 1]), xname="group", penalty=None)
>>> term = IndexingTerm(basis, penalty=jnp.eye(2), scale=1.0)
>>> term.value.shape, term.nclusters
((4,), 2)

Methods

init_full_basis

Materializes a Basis object that holds the full basis matrix corresponding to this term.

Attributes

nbases	Number of coefficients represented by the indexed basis.
nclusters	Number of unique clusters in this term (equals the number of coefficients).