gsnn.models.SoftmaxGroupNorm

class gsnn.models.SoftmaxGroupNorm(*args: Any, **kwargs: Any)[source]

Bases: Module

Channel-wise softmax normalized within each channel group (stable softmax via per-group max shift).

__init__(channel_groups, eps=1e-08)[source]
Parameters:

  • channel_groups – Length-C index assigning each channel to a group.

  • eps – Added to the denominator for numerical stability.

Methods

__init__(channel_groups[, eps])

param channel_groups:

Length-C index assigning each channel to a group.

forward(x)

Input (B, C) or (B, C, 1); returns group-softmax-normalized activations.
forward(x)[source]

Input (B, C) or (B, C, 1); returns group-softmax-normalized activations.