boundlab.prop.input_smear

boundlab.prop.input_smear(e, target, radius, reduce='max')

ReluVal-style smear per target dimension.

\[\mathrm{smear}_j = \mathrm{reduce}_i |J_{ij}| \cdot \mathrm{radius}_j\]

where J = input_attribution(e, target). reduce is "max" (ReluVal: \(\max_i |J_{ij}|\)), "sum" (total width contribution), or "none" (return |J| with no output reduction, just scaled by radius). radius broadcasts against target.shape.