#include <block_sparse_featurizer.hpp>
|
| 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 |
| |
◆ 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 |
| int64_t pulsatrix::BlockSparseOptions::k = 16 |
Blocks kept per input (Vanilla, Grassmannian).
◆ seed
| uint64_t pulsatrix::BlockSparseOptions::seed = 0 |
◆ selection
◆ 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
◆ verdict_overlap
| double pulsatrix::BlockSparseOptions::verdict_overlap = 0.5 |
The documentation for this struct was generated from the following file: