Adaptive Canonicalization

The implementation follows the prior maximization construction used in adaptive canonicalization:

• For each input x and each output channel (class) d, we search over a transformation family kappa_u(x) and compute

  s_d(x) = max_u Psi_d(kappa_u(x)).

Spectral band partition

we partition the normalized Laplacian spectrum [0, 2] into dyadic bands:

• b_0 = 0
• b_k = 2 * r^(B-k) for k = 1, ..., B-1
• b_B = 2

where:

• B = number of bands (tunable)
• r = decay rate in (0, 1) (tunable)

Usage examples

python tu/train.py --dataset MUTAG --num_bands 5 --band_decay 0.5 --candidates 32