What LRP explains: one target class (and optionally one contrast class) per row of the network's (N, num_classes) output.
More...
#include <lrp.hpp>
What LRP explains: one target class (and optionally one contrast class) per row of the network's (N, num_classes) output.
- Note
targets / contrasts hold either a single index (applied to every row) or one index per row. With contrasts, the seed is seed(target) - seed(contrast): LRP propagation is linear in the output relevance for fixed activations, so the result is exactly explanation(target) - explanation(contrast) – "why t rather than c".
◆ contrasts
| std::vector<int64_t> pulsatrix::LRPTarget::contrasts = {} |
◆ seed
◆ targets
| std::vector<int64_t> pulsatrix::LRPTarget::targets |
The documentation for this struct was generated from the following file: