GroupNorm

Group Normalization

Yuxin Wu and Kaiming He · ECCV 2018

Read paper · arXiv:1803.08494

Normalize contiguous channel groups within each sample and explore batch composition and group counts.

What you will implement

Implement differentiable grouped moments and per-channel affine; scalar oracles, gradient checks and a training-free synthetic experiment are supplied.

• 92. GroupNorm: Grouped Statistics and Batch Independence

  Implement per-sample group normalization with per-channel affine parameters and differentiable statistics.

  Complete one function; independent scalar checks, gradient checks and a batch-composition experiment are supplied, without model training.