One explainer's declaration: what it answers and what it needs.
More...
#include <explainer_capabilities.hpp>
|
| std::string | name |
| | A stable identifier, the class or function name: "IntegratedGradients".
|
| |
| std::string | header |
| | The header to include: "pulsatrix/integrated_gradients.hpp".
|
| |
| std::vector< std::string > | entry_points |
| | The calls that run it: "IntegratedGradients::explain".
|
| |
| std::string | summary |
| | One sentence on what it computes.
|
| |
| std::vector< XaiQuestion > | questions |
| | The questions it answers, the main one first.
|
| |
| ExplanationScope | scope = ExplanationScope::Local |
| |
| ModelAccess | access = ModelAccess::BlackBox |
| |
| InputKind | input = InputKind::Any |
| |
| bool | attribution = false |
| | It returns an Attribution with one value per input element, so the metrics in explanation_metrics.hpp (deletion, insertion, randomization, sparseness) apply.
|
| |
| bool | complete = false |
| | The attributions add up to the output, minus a baseline's where there is one: LRP's conservation, Integrated Gradients' completeness, Shapley values' efficiency. It can be checked on every explanation (ComputeConservation()), and LRP's ε rules keep it only approximately.
|
| |
| bool | needs_base_model = false |
| | It compares a model with the one it was fine-tuned from, so it needs both.
|
| |
| bool | needs_labels = false |
| | It scores the model against true labels (the Performance metrics, conformal calibration), so it needs a labeled dataset (ModelTraits::has_labels).
|
| |
| int64_t | max_features = 0 |
| | The most input features it can take, or 0 for no limit: KernelSHAP enumerates every coalition, so it takes at most 20.
|
| |
One explainer's declaration: what it answers and what it needs.
◆ answers()
| bool pulsatrix::ExplainerCapabilities::answers |
( |
XaiQuestion |
question | ) |
const |
Whether question is among questions.
◆ access
◆ attribution
| bool pulsatrix::ExplainerCapabilities::attribution = false |
◆ complete
| bool pulsatrix::ExplainerCapabilities::complete = false |
The attributions add up to the output, minus a baseline's where there is one: LRP's conservation, Integrated Gradients' completeness, Shapley values' efficiency. It can be checked on every explanation (ComputeConservation()), and LRP's ε rules keep it only approximately.
◆ entry_points
| std::vector<std::string> pulsatrix::ExplainerCapabilities::entry_points |
The calls that run it: "IntegratedGradients::explain".
◆ header
| std::string pulsatrix::ExplainerCapabilities::header |
The header to include: "pulsatrix/integrated_gradients.hpp".
◆ input
◆ max_features
| int64_t pulsatrix::ExplainerCapabilities::max_features = 0 |
The most input features it can take, or 0 for no limit: KernelSHAP enumerates every coalition, so it takes at most 20.
◆ name
| std::string pulsatrix::ExplainerCapabilities::name |
A stable identifier, the class or function name: "IntegratedGradients".
◆ needs_base_model
| bool pulsatrix::ExplainerCapabilities::needs_base_model = false |
It compares a model with the one it was fine-tuned from, so it needs both.
◆ needs_labels
| bool pulsatrix::ExplainerCapabilities::needs_labels = false |
It scores the model against true labels (the Performance metrics, conformal calibration), so it needs a labeled dataset (ModelTraits::has_labels).
◆ questions
| std::vector<XaiQuestion> pulsatrix::ExplainerCapabilities::questions |
The questions it answers, the main one first.
◆ scope
◆ summary
| std::string pulsatrix::ExplainerCapabilities::summary |
One sentence on what it computes.
The documentation for this struct was generated from the following file: