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

#include <parameter_decomposition.hpp>

Public Attributes

DecompositionMethod method = DecompositionMethod::SPD
 
OutputDivergence divergence = OutputDivergence::MeanSquaredError
 
float faithfulness_coefficient = 1.0f
 Σ_l |W - V U|² / (number of weights): SPD's faithfulness, VPD's Δ loss (whose paper weights it 1e7).
 
float stochastic_coefficient = 1.0f
 Reconstruction with every layer masked (SPD), or k of them (VPD; 0.5 there).
 
float layerwise_coefficient = 1.0f
 SPD: reconstruction with one layer masked at a time, averaged over layers.
 
float adversarial_coefficient = 0.5f
 VPD: reconstruction under adversarial masks.
 
float importance_coefficient = 3e-3f
 Importance minimality's weight and exponent p (set_p() anneals it).
 
float p = 1.0f
 
float frequency_beta = 0.5f
 VPD: the frequency term's β in Σ_c [mean_c + β mean_c log2(1 + sum_c)].
 
float leak = 0.01f
 The leaky slope of the causal importance's clamps.
 
int64_t pgd_steps = 2
 VPD's adversary: ascent steps per training step (before the one that uses the main loss's gradient), and Adam's learning rate and betas for the sources.
 
float pgd_lr = 0.01f
 
float pgd_beta1 = 0.5f
 
float pgd_beta2 = 0.99f
 
uint64_t seed = 0
 

Member Data Documentation

◆ adversarial_coefficient

float pulsatrix::DecompositionOptions::adversarial_coefficient = 0.5f

VPD: reconstruction under adversarial masks.

◆ divergence

OutputDivergence pulsatrix::DecompositionOptions::divergence = OutputDivergence::MeanSquaredError

◆ faithfulness_coefficient

float pulsatrix::DecompositionOptions::faithfulness_coefficient = 1.0f

Σ_l |W - V U|² / (number of weights): SPD's faithfulness, VPD's Δ loss (whose paper weights it 1e7).

◆ frequency_beta

float pulsatrix::DecompositionOptions::frequency_beta = 0.5f

VPD: the frequency term's β in Σ_c [mean_c + β mean_c log2(1 + sum_c)].

◆ importance_coefficient

float pulsatrix::DecompositionOptions::importance_coefficient = 3e-3f

Importance minimality's weight and exponent p (set_p() anneals it).

◆ layerwise_coefficient

float pulsatrix::DecompositionOptions::layerwise_coefficient = 1.0f

SPD: reconstruction with one layer masked at a time, averaged over layers.

◆ leak

float pulsatrix::DecompositionOptions::leak = 0.01f

The leaky slope of the causal importance's clamps.

◆ method

DecompositionMethod pulsatrix::DecompositionOptions::method = DecompositionMethod::SPD

◆ p

float pulsatrix::DecompositionOptions::p = 1.0f

◆ pgd_beta1

float pulsatrix::DecompositionOptions::pgd_beta1 = 0.5f

◆ pgd_beta2

float pulsatrix::DecompositionOptions::pgd_beta2 = 0.99f

◆ pgd_lr

float pulsatrix::DecompositionOptions::pgd_lr = 0.01f

◆ pgd_steps

int64_t pulsatrix::DecompositionOptions::pgd_steps = 2

VPD's adversary: ascent steps per training step (before the one that uses the main loss's gradient), and Adam's learning rate and betas for the sources.

◆ seed

uint64_t pulsatrix::DecompositionOptions::seed = 0

◆ stochastic_coefficient

float pulsatrix::DecompositionOptions::stochastic_coefficient = 1.0f

Reconstruction with every layer masked (SPD), or k of them (VPD; 0.5 there).


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