adelie.adelie_core.state.StateMultiGlmNaive32#

class adelie.adelie_core.state.StateMultiGlmNaive32#

Core state class for multi-response GLM, naive method.

Methods

`__init__`(args, *kwargs)	Overloaded function.
`solve`(self, arg0, arg1, arg2)	Solves the state-specific problem.

Attributes

`X`	Feature matrix.
`abs_grad`	The \(\ell_2\) norms of (corrected) `grad` across each group.
`active_set`	List of indices into `screen_set` that correspond to active groups.
`active_set_size`	Number of active groups.
`active_sizes`	Active set size for every saved solution.
`adev_tol`	Percent deviance explained tolerance.
`alpha`	Elastic net parameter.
`benchmark_fit_active`	Fit time on the active set for each iteration.
`benchmark_fit_screen`	Fit time on the screen set for each iteration.
`benchmark_invariance`	Invariance time for each iteration.
`benchmark_kkt`	KKT time for each iteration.
`benchmark_screen`	Screen time for each iteration.
`beta0`	The current intercept value.
`betas`	`betas[i]` is the solution at `lmdas[i]`.
`constraint_buffer_size`	Max constraint buffer size.
`constraints`	List of constraints for each group.
`ddev_tol`	Difference in percent deviance explained tolerance.
`devs`	`devs[i]` is the (normalized) \(R^2\) at `betas[i]`.
`dual_groups`	List of starting indices to each dual group where G is the number of groups.
`duals`	`duals[i]` is the dual at `lmdas[i]`.
`early_exit`	`True` if the function should early exit based on training percent deviance explained.
`eta`	The natural parameter \(\eta = X\beta + \beta_0 \mathbf{1} + \eta^0\) where \(\beta\) and \(\beta_0\) are given by `screen_beta` and `beta0`.
`grad`	The full gradient \(-X^\top \nabla \ell(\eta)\).
`group_sizes`	List of group sizes corresponding to each element in `groups`.
`groups`	List of starting indices to each group where G is the number of groups.
`intercept`	`True` if the function should fit with intercept.
`intercepts`	`intercepts[i]` is the intercept at `lmdas[i]` for each class.
`irls_max_iters`	Maximum number of IRLS iterations.
`irls_tol`	IRLS convergence tolerance.
`lmda`	The last regularization parameter that was attempted to be solved.
`lmda_max`	The smallest \(\lambda\) such that the true solution is zero for all coefficients that have a non-vanishing group lasso penalty (\(\ell_2\)-norm).
`lmda_path`	The regularization path to solve for.
`lmda_path_size`	Number of regularizations in the path if it is to be generated.
`lmdas`	`lmdas[i]` is the regularization \(\lambda\) used for the `i` th solution.
`loss_full`	Full loss \(\ell(\eta^\star)\) where \(\eta^\star\) is the minimizer.
`loss_null`	Null loss \(\ell(\beta_0^\star \mathbf{1} + \eta^0)\) from fitting an intercept-only model (if `intercept` is `True`) and otherwise \(\ell(\eta^0)\).
`max_active_size`	Maximum number of active groups allowed.
`max_iters`	Maximum number of coordinate descents.
`max_screen_size`	Maximum number of screen groups allowed.
`min_ratio`	The ratio between the largest and smallest \(\lambda\) in the regularization sequence if it is to be generated.
`multi_intercept`	`True` if an intercept is added for each response.
`n_threads`	Number of threads.
`n_valid_solutions`	Number of valid solutions for each iteration.
`newton_max_iters`	Maximum number of iterations for the BCD update.
`newton_tol`	Convergence tolerance for the BCD update.
`offsets`	Observation offsets \(\eta^0\).
`penalty`	Penalty factor for each group in the same order as `groups`.
`pivot_slack_ratio`	If screening takes place, then `pivot_slack_ratio` number of groups with next smallest (new) active scores below the pivot point are also added to the screen set as slack.
`pivot_subset_min`	If screening takes place, then at least `pivot_subset_min` number of active scores are used to determine the pivot point.
`pivot_subset_ratio`	If screening takes place, then the `(1 + pivot_subset_ratio) * s` largest active scores are used to determine the pivot point where `s` is the current screen set size.
`resid`	Residual \(-\nabla \ell(\eta)\) where \(\eta\) is given by `eta`.
`screen_begins`	List of indices that index a corresponding list of values for each screen group.
`screen_beta`	Coefficient vector on the screen set.
`screen_hashset`	Hashmap containing the same values as `screen_set`.
`screen_is_active`	Boolean vector that indicates whether each screen group in `groups` is active or not.
`screen_rule`	Strong rule type.
`screen_set`	List of indices into `groups` that correspond to the screen groups.
`screen_sizes`	Strong set size for every saved solution.
`setup_lmda_max`	`True` if the function should setup \(\lambda_\max\).
`setup_lmda_path`	`True` if the function should setup the regularization path.
`setup_loss_null`	`True` if the function should setup `loss_null`.
`tol`	Coordinate descent convergence tolerance.

__init__(*args, **kwargs)#

Overloaded function.

__init__(self: adelie.adelie_core.state.StateMultiGlmNaive32, n_classes: int, multi_intercept: bool, X: adelie.adelie_core.matrix.MatrixNaiveBase32, eta: numpy.ndarray[numpy.float32[1, n]], resid: numpy.ndarray[numpy.float32[1, n]], constraints: adelie.adelie_core.constraint.VectorConstraintBase32, groups: numpy.ndarray[numpy.int64[1, n]], group_sizes: numpy.ndarray[numpy.int64[1, n]], dual_groups: numpy.ndarray[numpy.int64[1, n]], alpha: float, penalty: numpy.ndarray[numpy.float32[1, n]], offsets: numpy.ndarray[numpy.float32[1, n]], lmda_path: numpy.ndarray[numpy.float32[1, n]], loss_null: float, loss_full: float, lmda_max: float, min_ratio: float, lmda_path_size: int, max_screen_size: int, max_active_size: int, pivot_subset_ratio: float, pivot_subset_min: int, pivot_slack_ratio: float, screen_rule: str, irls_max_iters: int, irls_tol: float, max_iters: int, tol: float, adev_tol: float, ddev_tol: float, newton_tol: float, newton_max_iters: int, early_exit: bool, setup_loss_null: bool, setup_lmda_max: bool, setup_lmda_path: bool, intercept: bool, n_threads: int, screen_set: numpy.ndarray[numpy.int64[1, n]], screen_beta: numpy.ndarray[numpy.float32[1, n]], screen_is_active: numpy.ndarray[bool[1, n]], active_set_size: int, active_set: numpy.ndarray[numpy.int64[1, n]], beta0: float, lmda: float, grad: numpy.ndarray[numpy.float32[1, n]]) -> None
__init__(self: adelie.adelie_core.state.StateMultiGlmNaive32, arg0: adelie.adelie_core.state.StateMultiGlmNaive32) -> None

solve(self: adelie.adelie_core.state.StateMultiGlmNaive32, arg0: adelie.adelie_core.glm.GlmMultiBase32, arg1: bool, arg2: Callable[[adelie.adelie_core.state.StateMultiGlmNaive32], bool]) → dict#: Solves the state-specific problem.

X#: Feature matrix.

abs_grad#: The \(\ell_2\) norms of (corrected) grad across each group. abs_grad[i] is given by np.linalg.norm(grad[g:g+gs] - lmda * penalty[i] * (1-alpha) * beta[g:g+gs] - correction) where g = groups[i], gs = group_sizes[i], beta is the full solution vector represented by screen_beta, and correction is the output from calling constraints[i].gradient().

active_set#: List of indices into screen_set that correspond to active groups. screen_set[active_set[i]] is the i th active group. An active group is one with non-zero coefficient block, that is, for every i th active group, screen_beta[b:b+p] == 0 where j = active_set[i], k = screen_set[j], b = screen_begins[j], and p = group_sizes[k].

active_set_size#: Number of active groups. active_set[i] is only well-defined for i in the range [0, active_set_size).

active_sizes#: Active set size for every saved solution.

adev_tol#: Percent deviance explained tolerance.

alpha#: Elastic net parameter.

benchmark_fit_active#: Fit time on the active set for each iteration.

benchmark_fit_screen#: Fit time on the screen set for each iteration.

benchmark_invariance#: Invariance time for each iteration.

benchmark_kkt#: KKT time for each iteration.

benchmark_screen#: Screen time for each iteration.

beta0#: The current intercept value.

betas#: betas[i] is the solution at lmdas[i].

constraint_buffer_size#: Max constraint buffer size. Equivalent to np.max([0 if c is None else c.buffer_size() for c in constraints]).

constraints#: List of constraints for each group. constraints[i] is the constraint object corresponding to group i.

ddev_tol#: Difference in percent deviance explained tolerance.

devs#: devs[i] is the (normalized) \(R^2\) at betas[i].

dual_groups#: List of starting indices to each dual group where G is the number of groups. dual_groups[i] is the starting index of the i th dual group.

duals#: duals[i] is the dual at lmdas[i].

early_exit#: True if the function should early exit based on training percent deviance explained.

eta#: The natural parameter \(\eta = X\beta + \beta_0 \mathbf{1} + \eta^0\) where \(\beta\) and \(\beta_0\) are given by screen_beta and beta0.

grad#: The full gradient \(-X^\top \nabla \ell(\eta)\).

group_sizes#: List of group sizes corresponding to each element in groups. group_sizes[i] is the group size of the i th group.

groups#: List of starting indices to each group where G is the number of groups. groups[i] is the starting index of the i th group.

intercept#: True if the function should fit with intercept.

intercepts#: intercepts[i] is the intercept at lmdas[i] for each class.

irls_max_iters#: Maximum number of IRLS iterations.

irls_tol#: IRLS convergence tolerance.

lmda#: The last regularization parameter that was attempted to be solved.

lmda_max#: The smallest \(\lambda\) such that the true solution is zero for all coefficients that have a non-vanishing group lasso penalty (\(\ell_2\)-norm).

lmda_path#: The regularization path to solve for.

lmda_path_size#: Number of regularizations in the path if it is to be generated.

lmdas#: lmdas[i] is the regularization \(\lambda\) used for the i th solution.

loss_full#: Full loss \(\ell(\eta^\star)\) where \(\eta^\star\) is the minimizer.

loss_null#: Null loss \(\ell(\beta_0^\star \mathbf{1} + \eta^0)\) from fitting an intercept-only model (if intercept is True) and otherwise \(\ell(\eta^0)\).

max_active_size#: Maximum number of active groups allowed.

max_iters#: Maximum number of coordinate descents.

max_screen_size#: Maximum number of screen groups allowed.

min_ratio#: The ratio between the largest and smallest \(\lambda\) in the regularization sequence if it is to be generated.

multi_intercept#: True if an intercept is added for each response.

n_threads#: Number of threads.

n_valid_solutions#: Number of valid solutions for each iteration.

newton_max_iters#: Maximum number of iterations for the BCD update.

newton_tol#: Convergence tolerance for the BCD update.

offsets#: Observation offsets \(\eta^0\).

penalty#: Penalty factor for each group in the same order as groups.

pivot_slack_ratio#: If screening takes place, then pivot_slack_ratio number of groups with next smallest (new) active scores below the pivot point are also added to the screen set as slack.

pivot_subset_min#: If screening takes place, then at least pivot_subset_min number of active scores are used to determine the pivot point.

pivot_subset_ratio#: If screening takes place, then the (1 + pivot_subset_ratio) * s largest active scores are used to determine the pivot point where s is the current screen set size.

resid#: Residual \(-\nabla \ell(\eta)\) where \(\eta\) is given by eta.

screen_begins#: List of indices that index a corresponding list of values for each screen group. screen_begins[i] is the starting index corresponding to the i th screen group. From this index, reading group_sizes[screen_set[i]] number of elements will grab values corresponding to the full i th screen group block.

screen_beta#: Coefficient vector on the screen set. screen_beta[b:b+p] is the coefficient for the i th screen group where k = screen_set[i], b = screen_begins[i], and p = group_sizes[k].

screen_hashset#: Hashmap containing the same values as screen_set.

screen_is_active#: Boolean vector that indicates whether each screen group in groups is active or not. screen_is_active[i] is True if and only if screen_set[i] is active.

screen_rule#: Strong rule type.

screen_set#: List of indices into groups that correspond to the screen groups. screen_set[i] is i th screen group.

screen_sizes#: Strong set size for every saved solution.

setup_lmda_max#: True if the function should setup \(\lambda_\max\).

setup_lmda_path#: True if the function should setup the regularization path.

setup_loss_null#: True if the function should setup loss_null.

tol#: Coordinate descent convergence tolerance.