pulsatrix 1.1.0
Loading...
Searching...
No Matches
null_model_baseline.hpp
Go to the documentation of this file.
1
10#pragma once
11
12#include <cstdint>
13#include <string>
14#include <type_traits>
15#include <utility>
16#include <vector>
17
19#include "pulsatrix/module.hpp"
20
21namespace pulsatrix {
22
25public:
26 explicit ParameterSnapshot(Module& model) {
27 for (ParamRef p : model.parameters()) {
28 saved_.emplace_back(p.value, *p.value);
29 }
30 }
34
36 void restore() {
37 for (auto& [target, value] : saved_) {
38 *target = value; // keeps the target's device and requires_grad flag
39 }
40 }
41
42private:
43 std::vector<std::pair<Tensor*, Tensor>> saved_;
44};
45
54void ReinitializeParameters(Module& model, uint64_t seed, const std::string& prefix = "");
55
57template <typename Result>
59 Result trained;
60 Result null_model;
61};
62
69template <typename Analysis>
70[[nodiscard]] auto NullModelBaseline(Module& model, Analysis&& analysis, uint64_t seed)
72 using Result = std::decay_t<std::invoke_result_t<Analysis&>>;
73 Result trained = analysis();
74 ParameterSnapshot saved(model);
75 ReinitializeParameters(model, seed);
76 Result null_model = analysis();
77 return {std::move(trained), std::move(null_model)};
78}
79
88
90[[nodiscard]] AttributionNullReport NullModelBaseline(Module& model, const ExplainFn& explain, const Tensor& input,
91 uint64_t seed);
92
93} // namespace pulsatrix
Base class for every layer type (LinearModule, Conv2DModule, activations, ...).
Definition module.hpp:78
virtual std::vector< ParamRef > parameters()
This module's trainable parameters and their gradients, for an optimizer to update uniformly across m...
Definition module.hpp:217
Saves every parameter of a module and restores them exactly when it goes out of scope.
Definition null_model_baseline.hpp:24
void restore()
Puts the saved values back now (the destructor does it again, harmlessly).
Definition null_model_baseline.hpp:36
ParameterSnapshot(Module &model)
Definition null_model_baseline.hpp:26
ParameterSnapshot & operator=(const ParameterSnapshot &)=delete
~ParameterSnapshot()
Definition null_model_baseline.hpp:31
ParameterSnapshot(const ParameterSnapshot &)=delete
N-dimensional tensor. Owns its data buffer exclusively; a DeviceBackend* is injected (not owned) – th...
Definition tensor.hpp:29
Explanation-quality metrics in the Quantus families (roadmap XAI-5): faithfulness (deletion and inser...
Abstract base every layer subclasses – NVI forward(), pure-virtual LRP contract.
Definition acquisition_functions.hpp:16
void ReinitializeParameters(Module &model, uint64_t seed, const std::string &prefix="")
Replaces each parameter tensor with Gaussian noise of mean 0 and that tensor's own standard deviation...
auto NullModelBaseline(Module &model, Analysis &&analysis, uint64_t seed) -> NullModelComparison< std::decay_t< std::invoke_result_t< Analysis & > > >
Runs analysis() on model, then again after ReinitializeParameters(model, seed), then restores the mod...
Definition null_model_baseline.hpp:70
std::function< Attribution(const Tensor &input)> ExplainFn
Explains input using the model in its current state.
Definition explanation_metrics.hpp:27
NullModelBaseline() for an explanation of one input.
Definition null_model_baseline.hpp:81
Attribution null_model
Definition null_model_baseline.hpp:83
float rank_similarity
Spearman rank correlation of the two attributions' magnitudes. Close to 1 means the explanation barel...
Definition null_model_baseline.hpp:86
Attribution trained
Definition null_model_baseline.hpp:82
An explanation result: the raw attribution values, the method that produced them, and any relevant me...
Definition attribution.hpp:23
NullModelBaseline()'s result: the analysis on the trained model and on the null model.
Definition null_model_baseline.hpp:58
Result null_model
Definition null_model_baseline.hpp:60
Result trained
Definition null_model_baseline.hpp:59
A trainable parameter and its accumulated gradient, as owned by some Module.
Definition module.hpp:24