resnest 0.0.4b20200430

ResNeSt

ResNeSt

Split-Attention Network, A New ResNet Variant. It significantly boosts the performance of downstream models such as Mask R-CNN, Cascade R-CNN and DeepLabV3.

Table of Contents

0. Pretrained Models
1. Transfer Learning Models
2. Verify Backbone Models
3. How to Train
4. Reference

Pypi / GitHub Install

0. Install this package repo, note that you only need to choose one of the options

# using github url
pip install git+https://github.com/zhanghang1989/ResNeSt

# using pypi
pip install resnest --pre

Pretrained Models

	crop size	PyTorch	Gluon
ResNeSt-50	224	81.03	81.04
ResNeSt-101	256	82.83	82.81
ResNeSt-200	320	83.84	83.88
ResNeSt-269	416	84.54	84.53

Ablation study models are available in link

PyTorch Models

• Load using Torch Hub

import torch
# get list of models
torch.hub.list('zhanghang1989/ResNeSt', force_reload=True)

# load pretrained models, using ResNeSt-50 as an example
net = torch.hub.load('zhanghang1989/ResNeSt', 'resnest50', pretrained=True)

• Load using python package

# using ResNeSt-50 as an example
from resnest.torch import resnest50
net = resnest50(pretrained=True)

Gluon Models

• Load pretrained model:

# using ResNeSt-50 as an example
from resnest.gluon import resnest50
net = resnest50(pretrained=True)

Transfer Learning Models

Detectron Models

Training code and pretrained models are released at our Detectron2 Fork.

Object Detection on MS-COCO validation set

Method	Backbone	mAP%
Faster R-CNN	ResNet-50	39.25
	ResNet-101	41.37
	ResNeSt-50 (ours)	42.33
	ResNeSt-101 (ours)	44.72
Cascade R-CNN	ResNet-50	42.52
	ResNet-101	44.03
	ResNeSt-50 (ours)	45.41
	ResNeSt-101 (ours)	47.50
	ResNeSt-200 (ours)	49.03

Instance Segmentation

Method	Backbone	bbox	mask
Mask R-CNN	ResNet-50	39.97	36.05
	ResNet-101	41.78	37.51
	ResNeSt-50 (ours)	42.81	38.14
	ResNeSt-101 (ours)	45.75	40.65
Cascade R-CNN	ResNet-50	43.06	37.19
	ResNet-101	44.79	38.52
	ResNeSt-50 (ours)	46.19	39.55
	ResNeSt-101 (ours)	48.30	41.56

Semantic Segmentation

• PyTorch models and training: Please visit PyTorch Encoding Toolkit.
• Training with Gluon: Please visit GluonCV Toolkit.

Results on ADE20K

Method	Backbone	pixAcc%	mIoU%
Deeplab-V3	ResNet-50	80.39	42.1
	ResNet-101	81.11	44.14
	ResNeSt-50 (ours)	81.17	45.12
	ResNeSt-101 (ours)	82.07	46.91
	ResNeSt-200 (ours)	82.45	48.36
	ResNeSt-269 (ours)	82.62	47.60

Verify Backbone Models:

Note: the inference speed reported in the paper are tested using Gluon implementation with RecordIO data.

Prepare ImageNet dataset:

Here we use raw image data format for simplicity, please follow GluonCV tutorial if you would like to use RecordIO format.

cd scripts/dataset/
# assuming you have downloaded the dataset in the current folder
python prepare_imagenet.py --download-dir ./

Torch Model

# use resnest50 as an example
cd scripts/torch/
python verify.py --model resnest50 --crop-size 224

Gluon Model

# use resnest50 as an example
cd scripts/gluon/
python verify.py --model resnest50 --crop-size 224

How to Train

ImageNet Models

• Training with MXNet Gluon: Please visit Gluon folder.
• Training with PyTorch: Please visit PyTorch Encoding Toolkit (slightly worse than Gluon implementation).

Detectron Models

For object detection and instance segmentation models, please visit our detectron2-ResNeSt fork.

Semantic Segmentation

• Training with PyTorch: Encoding Toolkit.
• Training with MXNet: GluonCV Toolkit.

Reference

ResNeSt: Split-Attention Networks [arXiv]

Hang Zhang, Chongruo Wu, Zhongyue Zhang, Yi Zhu, Zhi Zhang, Haibin Lin, Yue Sun, Tong He, Jonas Muller, R. Manmatha, Mu Li and Alex Smola

@article{zhang2020resnest,
title={ResNeSt: Split-Attention Networks},
author={Zhang, Hang and Wu, Chongruo and Zhang, Zhongyue and Zhu, Yi and Zhang, Zhi and Lin, Haibin and Sun, Yue and He, Tong and Muller, Jonas and Manmatha, R. and Li, Mu and Smola, Alexander},
journal={arXiv preprint arXiv:2004.08955},
year={2020}
}

Major Contributors

• ResNeSt Backbone (Hang Zhang)
• Detectron Models (Chongruo Wu, Zhongyue Zhang)
• Semantic Segmentation (Yi Zhu)
• Distributed Training (Haibin Lin)