pulsatrix 1.1.0
Loading...
Searching...
No Matches
pulsatrix::BlockSparseOptions Struct Reference

#include <block_sparse_featurizer.hpp>

Public Attributes

BsfVariant variant = BsfVariant::Vanilla
 
int64_t block_size = 3
 b, the directions per block (the paper's main runs use 3).
 
int64_t k = 16
 Blocks kept per input (Vanilla, Grassmannian).
 
BlockSelection selection = BlockSelection::TopK
 
double duel_overlap = 0.1
 Tournament: a candidate block duels an accepted one when their subspace overlap |Q_g Q_hᵀ|²_F / b exceeds duel_overlap; it loses and is skipped, and the accepted block gains a win. Above verdict_overlap the verdict is permanent: the block with more wins keeps the feature and the other is never selected again.
 
double verdict_overlap = 0.5
 
int64_t tournament_warmup = 0
 Training steps of plain top-k before duels begin.
 
double target_l0 = 16
 GroupLasso: the mean number of active blocks per input to hold, by dual ascent on log λ with step dual_lr (the multiplier starts at initial_lambda).
 
float initial_lambda = 1e-2f
 
float dual_lr = 3e-2f
 
float gain = 10.0f
 GroupLasso: θ_g = softplus(gain · raw_g); the straight-through Gaussian kernel's bandwidth is this multiple of the block norms' standard deviation (a moving average).
 
float bandwidth = 0.25f
 
uint64_t seed = 0
 

Member Data Documentation

◆ bandwidth

float pulsatrix::BlockSparseOptions::bandwidth = 0.25f

◆ block_size

int64_t pulsatrix::BlockSparseOptions::block_size = 3

b, the directions per block (the paper's main runs use 3).

◆ dual_lr

float pulsatrix::BlockSparseOptions::dual_lr = 3e-2f

◆ duel_overlap

double pulsatrix::BlockSparseOptions::duel_overlap = 0.1

Tournament: a candidate block duels an accepted one when their subspace overlap |Q_g Q_hᵀ|²_F / b exceeds duel_overlap; it loses and is skipped, and the accepted block gains a win. Above verdict_overlap the verdict is permanent: the block with more wins keeps the feature and the other is never selected again.

◆ gain

float pulsatrix::BlockSparseOptions::gain = 10.0f

GroupLasso: θ_g = softplus(gain · raw_g); the straight-through Gaussian kernel's bandwidth is this multiple of the block norms' standard deviation (a moving average).

◆ initial_lambda

float pulsatrix::BlockSparseOptions::initial_lambda = 1e-2f

◆ k

int64_t pulsatrix::BlockSparseOptions::k = 16

Blocks kept per input (Vanilla, Grassmannian).

◆ seed

uint64_t pulsatrix::BlockSparseOptions::seed = 0

◆ selection

BlockSelection pulsatrix::BlockSparseOptions::selection = BlockSelection::TopK

◆ target_l0

double pulsatrix::BlockSparseOptions::target_l0 = 16

GroupLasso: the mean number of active blocks per input to hold, by dual ascent on log λ with step dual_lr (the multiplier starts at initial_lambda).

◆ tournament_warmup

int64_t pulsatrix::BlockSparseOptions::tournament_warmup = 0

Training steps of plain top-k before duels begin.

◆ variant

BsfVariant pulsatrix::BlockSparseOptions::variant = BsfVariant::Vanilla

◆ verdict_overlap

double pulsatrix::BlockSparseOptions::verdict_overlap = 0.5

The documentation for this struct was generated from the following file: