README.md

Weight Standardization

Introduction

@article{weightstandardization,
  author    = {Siyuan Qiao and Huiyu Wang and Chenxi Liu and Wei Shen and Alan Yuille},
  title     = {Weight Standardization},
  journal   = {arXiv preprint arXiv:1903.10520},
  year      = {2019},
}

Results and Models

Faster R-CNN

Backbone	Style	Normalization	Lr schd	box AP	mask AP	Download
R-50-FPN	pytorch	GN	1x	37.8	-	-
R-50-FPN	pytorch	GN+WS	1x	38.9	-	model
R-101-FPN	pytorch	GN	1x	39.8	-	-
R-101-FPN	pytorch	GN+WS	1x	41.4	-	model
X-50-32x4d-FPN	pytorch	GN	1x	36.5	-	-
X-50-32x4d-FPN	pytorch	GN+WS	1x	39.9	-	model
X-101-32x4d-FPN	pytorch	GN	1x	33.2	-	-
X-101-32x4d-FPN	pytorch	GN+WS	1x	41.8	-	model

Mask R-CNN

Backbone	Style	Normalization	Lr schd	box AP	mask AP	Download
R-50-FPN	pytorch	GN	2x	39.9	36.0	-
R-50-FPN	pytorch	GN+WS	2x	40.3	36.2	model
R-101-FPN	pytorch	GN	2x	41.6	37.3	-
R-101-FPN	pytorch	GN+WS	2x	42.0	37.3	model
X-50-32x4d-FPN	pytorch	GN	2x	39.2	35.5	-
X-50-32x4d-FPN	pytorch	GN+WS	2x	40.7	36.7	model
X-101-32x4d-FPN	pytorch	GN	2x	36.4	33.1	-
X-101-32x4d-FPN	pytorch	GN+WS	2x	42.1	37.7	model
R-50-FPN	pytorch	GN	20-23-24e	40.6	36.6	-
R-50-FPN	pytorch	GN+WS	20-23-24e	41.1	37.0	model
R-101-FPN	pytorch	GN	20-23-24e	42.3	38.1	-
R-101-FPN	pytorch	GN+WS	20-23-24e	43.0	38.4	model
X-50-32x4d-FPN	pytorch	GN	20-23-24e	39.6	35.9	-
X-50-32x4d-FPN	pytorch	GN+WS	20-23-24e	41.9	37.7	model
X-101-32x4d-FPN	pytorch	GN	20-23-24e	36.6	33.4	-
X-101-32x4d-FPN	pytorch	GN+WS	20-23-24e	43.4	38.7	model

Note:

• GN+WS requires about 5% more memory than GN, and it is only 5% slower than GN.
• In the paper, a 20-23-24e lr schedule is used instead of 2x.
• The X-50-GN and X-101-GN pretrained models are also shared by the authors.