How a removed feature is filled in when perturbing an input.
More...
#include <explanation_metrics.hpp>
|
| static Imputation | Constant (float value=0.0f) |
| | Every removed element becomes value.
|
| |
| static Imputation | NoisyLinear (float noise_std=0.01f, uint64_t seed=0) |
| | ROAD (Rong et al., "A Consistent and Efficient Evaluation Strategy for Attribution
Methods", ICML 2022): each removed pixel becomes the weighted average of its neighbors in its (H, W) plane – direct neighbors 1/6, diagonal ones 1/12, solved jointly for all removed pixels – plus Gaussian noise. A constant fill leaks which pixels were removed through the shape of the hole; this doesn't.
|
| |
How a removed feature is filled in when perturbing an input.
◆ Kind
| Enumerator |
|---|
| Constant | |
| NoisyLinear | |
◆ Constant()
| static Imputation pulsatrix::Imputation::Constant |
( |
float |
value = 0.0f | ) |
|
|
inlinestatic |
Every removed element becomes value.
◆ NoisyLinear()
| static Imputation pulsatrix::Imputation::NoisyLinear |
( |
float |
noise_std = 0.01f, |
|
|
uint64_t |
seed = 0 |
|
) |
| |
|
inlinestatic |
ROAD (Rong et al., "A Consistent and Efficient Evaluation Strategy for Attribution
Methods", ICML 2022): each removed pixel becomes the weighted average of its neighbors in its (H, W) plane – direct neighbors 1/6, diagonal ones 1/12, solved jointly for all removed pixels – plus Gaussian noise. A constant fill leaks which pixels were removed through the shape of the hole; this doesn't.
- Note
- Meant for images: the last two dimensions are treated as (H, W).
◆ kind
◆ noise_std
| float pulsatrix::Imputation::noise_std = 0.01f |
◆ seed
| uint64_t pulsatrix::Imputation::seed = 0 |
◆ value
| float pulsatrix::Imputation::value = 0.0f |
The documentation for this struct was generated from the following file: