|
pulsatrix 1.1.0
|
Configuration for LRP relevance propagation: which rule a module applies and its hyperparameters. A module that does not implement the requested rule throws (see Module::supports_lrp_rule()) – the rule is never silently substituted. More...
#include <lrp_rule_config.hpp>
Public Attributes | |
| float | epsilon = 1e-6f |
| Stabilizer added to every rule's denominator: z + epsilon * sign(z), sign(0) = +1. | |
| LRPRule | rule = LRPRule::Epsilon |
| Which rule to apply. | |
| float | gamma = 0.25f |
| Gamma rule: weight of the positive (or negative) part added to the weights. | |
| float | alpha = 1.0f |
| AlphaBeta rule: weight of the positive contributions. Needs alpha, beta >= 0, alpha - beta == 1. | |
| float | beta = 0.0f |
| AlphaBeta rule: weight of the negative contributions. | |
| float | low = 0.0f |
| ZBox rule: lowest admissible input value (scalar, applied to every input element). | |
| float | high = 1.0f |
| ZBox rule: highest admissible input value. | |
| bool | epsilon_bias_in_denominator = false |
| Epsilon rule only, honoured by LinearModule / Conv2DModule: include the bias in the denominator, z = xW + b (Zennit's Epsilon). The default (false) keeps this codebase's conservative pre-bias z = xW, under which the bias absorbs no relevance. | |
| bool | layer_norm_detach_std = false |
LayerNormModule only: AttnLRP's normalization rule as LXT applies it. The standard deviation is held constant, and the rest, gamma (x - mean) / std + beta, is linear and takes the epsilon rule, its mean subtraction included (the bias in the denominator as epsilon_bias_in_denominator says). The default (false) passes relevance through unchanged, which is the same for RMSNorm but ignores LayerNorm's mean and bias. LxtAttnLrpConfig() sets it. | |
| std::shared_ptr< const LRPRuleConfig > | conv_rule |
| The rule every Conv2DModule applies, wherever it sits, in place of this config (KS-9). Zennit maps rules by layer type through the whole network, and a residual block is one module here that passes a single config to the convolutions inside it, so a composite such as EpsilonPlus sets it ("Epsilon, and ZPlus for convolutions") instead of relying on each convolution being a top-level module. Empty: convolutions use this config. | |
| LoRARelevance | lora_relevance = LoRARelevance::Split |
| LinearModule with an unmerged, enabled LoRA adapter only: split the relevance between the base and adapter paths (the default) or treat the layer as merged. A merged adapter can't be split: its update is part of W. | |
Configuration for LRP relevance propagation: which rule a module applies and its hyperparameters. A module that does not implement the requested rule throws (see Module::supports_lrp_rule()) – the rule is never silently substituted.
epsilon stays the first member so aggregate initialization LRPRuleConfig{0.01f} keeps meaning "epsilon rule, epsilon 0.01". It is also the denominator stabilizer of Gamma / AlphaBeta / ZBox (Zennit's stabilizer, default 1e-6 there as well). | float pulsatrix::LRPRuleConfig::alpha = 1.0f |
AlphaBeta rule: weight of the positive contributions. Needs alpha, beta >= 0, alpha - beta == 1.
| float pulsatrix::LRPRuleConfig::beta = 0.0f |
AlphaBeta rule: weight of the negative contributions.
| std::shared_ptr<const LRPRuleConfig> pulsatrix::LRPRuleConfig::conv_rule |
The rule every Conv2DModule applies, wherever it sits, in place of this config (KS-9). Zennit maps rules by layer type through the whole network, and a residual block is one module here that passes a single config to the convolutions inside it, so a composite such as EpsilonPlus sets it ("Epsilon, and ZPlus for convolutions") instead of relying on each convolution being a top-level module. Empty: convolutions use this config.
| float pulsatrix::LRPRuleConfig::epsilon = 1e-6f |
Stabilizer added to every rule's denominator: z + epsilon * sign(z), sign(0) = +1.
| bool pulsatrix::LRPRuleConfig::epsilon_bias_in_denominator = false |
Epsilon rule only, honoured by LinearModule / Conv2DModule: include the bias in the denominator, z = xW + b (Zennit's Epsilon). The default (false) keeps this codebase's conservative pre-bias z = xW, under which the bias absorbs no relevance.
| float pulsatrix::LRPRuleConfig::gamma = 0.25f |
Gamma rule: weight of the positive (or negative) part added to the weights.
| float pulsatrix::LRPRuleConfig::high = 1.0f |
ZBox rule: highest admissible input value.
| bool pulsatrix::LRPRuleConfig::layer_norm_detach_std = false |
LayerNormModule only: AttnLRP's normalization rule as LXT applies it. The standard deviation is held constant, and the rest, gamma (x - mean) / std + beta, is linear and takes the epsilon rule, its mean subtraction included (the bias in the denominator as epsilon_bias_in_denominator says). The default (false) passes relevance through unchanged, which is the same for RMSNorm but ignores LayerNorm's mean and bias. LxtAttnLrpConfig() sets it.
| LoRARelevance pulsatrix::LRPRuleConfig::lora_relevance = LoRARelevance::Split |
LinearModule with an unmerged, enabled LoRA adapter only: split the relevance between the base and adapter paths (the default) or treat the layer as merged. A merged adapter can't be split: its update is part of W.
| float pulsatrix::LRPRuleConfig::low = 0.0f |
ZBox rule: lowest admissible input value (scalar, applied to every input element).
| LRPRule pulsatrix::LRPRuleConfig::rule = LRPRule::Epsilon |
Which rule to apply.