Layers

Implement some simple re-parameterization.

Classes

Identity	Identity layer does nothing to the parameters. It is used to be base class for other layers.
NonNegative	NonNegative layer ensures that all parameters are non-negative.
LinearConstraint	LinearConstraint layer ensures that the parameters satisfy the linear constraint: <coef, params> = 1.
SimplexConstraint	SimplexConstraint layer ensures that the parameters satisfy the linear constraint: <coef, params> = 1 and all parameters are non-negative.
BoxConstraint	BoxConstraint layer ensures that the parameters are in the box: lower <= params <= upper.
OffsetSparse	OffsetSparse layer ensures that the sparse constraint of sparse solvers changes from ||params||_0 = s to ||params - offset||_0 = s. In other words, the layer ensures that the parameters corresponding to the non-selected features are equal to offset rather than zero.

class skscope.layer.BoxConstraint(dimensionality, lower, upper)

BoxConstraint layer ensures that the parameters are in the box: lower <= params <= upper.

Parameters:

• dimensionality (int) – Dimensionality of the parameters.

• lower (float or array with shape (dimensionality,)) – Lower bound of the box, if lower is a float, then lower * ones(dimensionality) is used. lower must be non-positive.

• upper (float or array with shape (dimensionality,)) – Upper bound of the box, if upper is a float, then upper * ones(dimensionality) is used. upper must be non-negative.

transform_params(params)

tree_flatten()

classmethod tree_unflatten(aux_data, children)

class skscope.layer.Identity(dimensionality)

Identity layer does nothing to the parameters. It is used to be base class for other layers.

Parameters:

dimensionality (int) – Dimensionality of the parameters.

random_initilization = False

transform_group(group)

transform_params(params)

transform_preselect(preselect)

transform_sparsity(sparsity)

tree_flatten()

classmethod tree_unflatten(aux_data, children)

class skscope.layer.LinearConstraint(dimensionality, coef=None)

LinearConstraint layer ensures that the parameters satisfy the linear constraint: <coef, params> = 1.

Parameters:

• dimensionality (int) – Dimensionality of the parameters.

• coef (float or array with shape (dimensionality,)) – Coefficients of the linear constraint <coef, params> = 1. If coef is a float, then coef * ones(dimensionality) is used.

random_initilization = True

transform_params(params)

tree_flatten()

classmethod tree_unflatten(aux_data, children)

class skscope.layer.NonNegative(dimensionality)

NonNegative layer ensures that all parameters are non-negative.

Parameters:

dimensionality (int) – Dimensionality of the parameters.

transform_params(params)

class skscope.layer.OffsetSparse(dimensionality, offset)

OffsetSparse layer ensures that the sparse constraint of sparse solvers changes from ||params||_0 = s to ||params - offset||_0 = s. In other words, the layer ensures that the parameters corresponding to the non-selected features are equal to offset rather than zero.

Parameters:

• dimensionality (int) – Dimensionality of the parameters.

• offset (float or array with shape (dimensionality,)) – Offset of the sparse constraint.

transform_params(params)

tree_flatten()

classmethod tree_unflatten(aux_data, children)

class skscope.layer.SimplexConstraint(dimensionality, coef=None)

SimplexConstraint layer ensures that the parameters satisfy the linear constraint: <coef, params> = 1 and all parameters are non-negative.

Parameters:

• dimensionality (int) – Dimensionality of the parameters.

• coef (float or array with shape (dimensionality,)) – Coefficients of the linear constraint <coef, params> = 1. If coef is a float, then coef * ones(dimensionality) is used.

random_initilization = True

transform_params(params)

tree_flatten()

classmethod tree_unflatten(aux_data, children)