GroupNorm + WS

open-mmlab / mmdetection

Last updated on
Faster R-CNN GroupNorm + WS (R-101-FPN, 1x, pytorch)
Memory (M) 8900.0
inference time (s/im) 0.11111
File Size 236.80 MB
Training Data MS COCO
Training Resources 8x NVIDIA V100 GPUs
Training Time
Architecture Softmax, RPN, Weight Standardization, Convolution, Group Normalization, FPN, RoIPool, ResNet
lr sched 1x
Memory (M) 8900.0
Backbone Layers 101
inference time (s/im) 0.11111
SHOW MORE
SHOW LESS
Faster R-CNN GroupNorm + WS (R-50-FPN, 1x, pytorch)
Memory (M) 5900.0
inference time (s/im) 0.08547
File Size 164.33 MB
Training Data MS COCO
Training Resources 8x NVIDIA V100 GPUs
Training Time
Architecture Softmax, RPN, Weight Standardization, Convolution, Group Normalization, FPN, RoIPool, ResNet
lr sched 1x
Memory (M) 5900.0
Backbone Layers 50
inference time (s/im) 0.08547
SHOW MORE
SHOW LESS
Faster R-CNN GroupNorm + WS (X-101-32x4d-FPN, 1x, pytorch)
Memory (M) 10800.0
inference time (s/im) 0.13158
File Size 235.38 MB
Training Data MS COCO
Training Resources 8x NVIDIA V100 GPUs
Training Time
Architecture Softmax, RPN, ResNeXt, Weight Standardization, Convolution, Group Normalization, FPN, RoIPool
lr sched 1x
Memory (M) 10800.0
Backbone Layers 101
inference time (s/im) 0.13158
SHOW MORE
SHOW LESS
Faster R-CNN GroupNorm + WS (X-50-32x4d-FPN, 1x, pytorch)
Memory (M) 7000.0
inference time (s/im) 0.09709
File Size 162.31 MB
Training Data MS COCO
Training Resources 8x NVIDIA V100 GPUs
Training Time
Architecture Softmax, RPN, Weight Standardization, Convolution, Group Normalization, FPN, RoIPool
lr sched 1x
Memory (M) 7000.0
Backbone Layers 50
inference time (s/im) 0.09709
SHOW MORE
SHOW LESS
Mask R-CNN GroupNorm + WS (R-101-FPN, 20-23-24e, pytorch)
Memory (M) 10300.0
Backbone Layers 101
File Size 246.89 MB
Training Data MS COCO
Training Resources 8x NVIDIA V100 GPUs
Training Time
Architecture Softmax, RPN, Weight Standardization, Convolution, Dense Connections, Group Normalization, FPN, ResNet, RoIAlign
lr sched 20-23-24e
Memory (M) 10300.0
Backbone Layers 101
SHOW MORE
SHOW LESS
Mask R-CNN GroupNorm + WS (R-101-FPN, 2x, pytorch)
Memory (M) 10300.0
inference time (s/im) 0.11628
File Size 246.89 MB
Training Data MS COCO
Training Resources 8x NVIDIA V100 GPUs
Training Time
Architecture Softmax, RPN, Weight Standardization, Convolution, Dense Connections, Group Normalization, FPN, ResNet, RoIAlign
lr sched 2x
Memory (M) 10300.0
Backbone Layers 101
inference time (s/im) 0.11628
SHOW MORE
SHOW LESS
Mask R-CNN GroupNorm + WS (R-50-FPN, 20-23-24e, pytorch)
Memory (M) 7300.0
Backbone Layers 50
File Size 174.42 MB
Training Data MS COCO
Training Resources 8x NVIDIA V100 GPUs
Training Time
Architecture Softmax, RPN, Weight Standardization, Convolution, Dense Connections, Group Normalization, FPN, ResNet, RoIAlign
lr sched 20-23-24e
Memory (M) 7300.0
Backbone Layers 50
SHOW MORE
SHOW LESS
Mask R-CNN GroupNorm + WS (R-50-FPN, 2x, pytorch)
Memory (M) 7300.0
inference time (s/im) 0.09524
File Size 174.42 MB
Training Data MS COCO
Training Resources 8x NVIDIA V100 GPUs
Training Time
Architecture Softmax, RPN, Weight Standardization, Convolution, Dense Connections, Group Normalization, FPN, ResNet, RoIAlign
lr sched 2x
Memory (M) 7300.0
Backbone Layers 50
inference time (s/im) 0.09524
SHOW MORE
SHOW LESS
Mask R-CNN GroupNorm + WS (X-101-32x4d-FPN, 20-23-24e, pytorch)
Memory (M) 12200.0
Backbone Layers 101
File Size 245.47 MB
Training Data MS COCO
Training Resources 8x NVIDIA V100 GPUs
Training Time
Architecture Softmax, RPN, ResNeXt, Weight Standardization, Convolution, Dense Connections, Group Normalization, FPN, RoIAlign
lr sched 20-23-24e
Memory (M) 12200.0
Backbone Layers 101
SHOW MORE
SHOW LESS
Mask R-CNN GroupNorm + WS (X-101-32x4d-FPN, 2x, pytorch)
Memory (M) 12200.0
inference time (s/im) 0.14085
File Size 245.47 MB
Training Data MS COCO
Training Resources 8x NVIDIA V100 GPUs
Training Time
Architecture Softmax, RPN, ResNeXt, Weight Standardization, Convolution, Dense Connections, Group Normalization, FPN, RoIAlign
lr sched 2x
Memory (M) 12200.0
Backbone Layers 101
inference time (s/im) 0.14085
SHOW MORE
SHOW LESS
Mask R-CNN GroupNorm + WS (X-50-32x4d-FPN, 20-23-24e, pytorch)
Memory (M) 8400.0
Backbone Layers 50
File Size 172.40 MB
Training Data MS COCO
Training Resources 8x NVIDIA V100 GPUs
Training Time
Architecture Softmax, RPN, Weight Standardization, Convolution, Dense Connections, Group Normalization, FPN, RoIAlign
lr sched 20-23-24e
Memory (M) 8400.0
Backbone Layers 50
SHOW MORE
SHOW LESS
Mask R-CNN GroupNorm + WS (X-50-32x4d-FPN, 2x, pytorch)
Memory (M) 8400.0
inference time (s/im) 0.10753
File Size 172.40 MB
Training Data MS COCO
Training Resources 8x NVIDIA V100 GPUs
Training Time
Architecture Softmax, RPN, Weight Standardization, Convolution, Dense Connections, Group Normalization, FPN, RoIAlign
lr sched 2x
Memory (M) 8400.0
Backbone Layers 50
inference time (s/im) 0.10753
SHOW MORE
SHOW LESS
README.md

Weight Standardization

Introduction

[ALGORITHM]

@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 Mem (GB) Inf time (fps) box AP mask AP Config Download
R-50-FPN pytorch GN+WS 1x 5.9 11.7 39.7 - config model | log
R-101-FPN pytorch GN+WS 1x 8.9 9.0 41.7 - config model | log
X-50-32x4d-FPN pytorch GN+WS 1x 7.0 10.3 40.7 - config model | log
X-101-32x4d-FPN pytorch GN+WS 1x 10.8 7.6 42.1 - config model | log

Mask R-CNN

Backbone Style Normalization Lr schd Mem (GB) Inf time (fps) box AP mask AP Config Download
R-50-FPN pytorch GN+WS 2x 7.3 10.5 40.6 36.6 config model | log
R-101-FPN pytorch GN+WS 2x 10.3 8.6 42.0 37.7 config model | log
X-50-32x4d-FPN pytorch GN+WS 2x 8.4 9.3 41.1 37.0 config model | log
X-101-32x4d-FPN pytorch GN+WS 2x 12.2 7.1 42.1 37.9 config model | log
R-50-FPN pytorch GN+WS 20-23-24e 7.3 - 41.1 37.1 config model | log
R-101-FPN pytorch GN+WS 20-23-24e 10.3 - 43.1 38.6 config model | log
X-50-32x4d-FPN pytorch GN+WS 20-23-24e 8.4 - 42.1 38.0 config model | log
X-101-32x4d-FPN pytorch GN+WS 20-23-24e 12.2 - 42.7 38.5 config model | log

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.

Results

Object Detection on COCO minival

Object Detection on COCO minival
MODEL BOX AP
Mask R-CNN GroupNorm + WS (R-101-FPN, 20-23-24e, pytorch) 43.1
Mask R-CNN GroupNorm + WS (X-101-32x4d-FPN, 20-23-24e, pytorch) 42.7
Mask R-CNN GroupNorm + WS (X-101-32x4d-FPN, 2x, pytorch) 42.1
Faster R-CNN GroupNorm + WS (X-101-32x4d-FPN, 1x, pytorch) 42.1
Mask R-CNN GroupNorm + WS (X-50-32x4d-FPN, 20-23-24e, pytorch) 42.1
Mask R-CNN GroupNorm + WS (R-101-FPN, 2x, pytorch) 42.0
Faster R-CNN GroupNorm + WS (R-101-FPN, 1x, pytorch) 41.7
Mask R-CNN GroupNorm + WS (R-50-FPN, 20-23-24e, pytorch) 41.1
Mask R-CNN GroupNorm + WS (X-50-32x4d-FPN, 2x, pytorch) 41.1
Faster R-CNN GroupNorm + WS (X-50-32x4d-FPN, 1x, pytorch) 40.7
Mask R-CNN GroupNorm + WS (R-50-FPN, 2x, pytorch) 40.6
Faster R-CNN GroupNorm + WS (R-50-FPN, 1x, pytorch) 39.7
Instance Segmentation on COCO minival
MODEL MASK AP
Mask R-CNN GroupNorm + WS (R-101-FPN, 20-23-24e, pytorch) 38.6
Mask R-CNN GroupNorm + WS (X-101-32x4d-FPN, 20-23-24e, pytorch) 38.5
Mask R-CNN GroupNorm + WS (X-50-32x4d-FPN, 20-23-24e, pytorch) 38.0
Mask R-CNN GroupNorm + WS (X-101-32x4d-FPN, 2x, pytorch) 37.9
Mask R-CNN GroupNorm + WS (R-101-FPN, 2x, pytorch) 37.7
Mask R-CNN GroupNorm + WS (R-50-FPN, 20-23-24e, pytorch) 37.1
Mask R-CNN GroupNorm + WS (X-50-32x4d-FPN, 2x, pytorch) 37.0
Mask R-CNN GroupNorm + WS (R-50-FPN, 2x, pytorch) 36.6
Contact us on: hello@paperswithcode.com . Papers With Code is a free resource with all data licensed under CC-BY-SA.
Terms Data policy Cookies policy