Fix ZPlus gradient_mapper for negative inputs

chr5tphr / zennit

Zennit is a high-level framework in Python using PyTorch for explaining/exploring neural networks using attribution methods like LRP.

Other

190 stars 32 forks source link

Closed chr5tphr closed 2 years ago

chr5tphr commented 2 years ago

ZPlus rule was wrong for negative inputs: each output partition used its own respective sum for the denominator (i.e. like alpha-beta would handle positive and negative outputs)
for the correct ZPlus the denominator of each (2) output partitions is the sum of all positive outputs