EPSANet

This repo contains the official Pytorch implementaion code and configuration files of EPSANet：An Efficient Pyramid Split Attention Block on Convolutional Neural Network. created by Hu Zhang.

Installation

Requirements

  • Python 3.6+
  • PyTorch 1.0+

Our environments

  • OS: Ubuntu 18.04
  • CUDA: 10.0
  • Toolkit: PyTorch 1.0
  • GPU: Titan RTX

Data preparation

Download and extract ImageNet train and val images from http://image-net.org/.
The directory structure is the standard layout for the torchvision datasets.ImageFolder, and the training and validation data is expected to be in the train/ folder and val folder respectively:

/path/to/imagenet/
  train/
    class1/
      img1.jpeg
    class2/
      img2.jpeg
  val/
    class1/
      img3.jpeg
    class/2
      img4.jpeg

Usage

First, clone the repository locally:

git clone https://github.com/murufeng/EPSANet.git
cd EPSANet
  • Create a conda virtual environment and activate it:
conda create -n epsanet python=3.6 
conda activate epsanet

conda install -c pytorch pytorch torchvision

Training

To train models on ImageNet with 8 gpus run:

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python main.py -a epsanet50 --data /path/to/imagenet

Model Zoo

Models are trained with 8 GPUs on both ImageNet and MS-COCO 2017 dataset.

Image Classification on ImageNet

Model Params(M) FLOPs(G) Top-1 (%) Top-5 (%)
EPSANet-50(Small) 22.56 3.62 77.49 93.54
EPSANet-50(Large) 27.90 4.72 78.64 94.18
EPSANet-101(Small) 38.90 6.82 78.43 94.11
EPSANet-101(Large) 49.59 8.97 79.38 94.58

Object Detection on MS-COCO 2017

Faster R-CNN

model Style Lr schd Params(M) FLOPs(G) box AP AP_50 AP_75
EPSANet-50(small) pytorch 1x 38.56 197.07 39.2 60.3 42.3
EPSANet-50(large) pytorch 1x 43.85 219.64 40.9 62.1 44.6

Mask R-CNN

model Style Lr schd Params(M) FLOPs(G) box AP AP_50 AP_75
EPSANet-50(small) pytorch 1x 41.20 248.53 40.0 60.9 43.3
EPSANet-50(large) pytorch 1x 46.50 271.10 41.4 62.3 45.3

RetinaNet

model Style Lr schd Params(M) FLOPs(G) box AP AP_50 AP_75
EPSANet-50(small) pytorch 1x 34.78 229.32 38.2 58.1 40.6
EPSANet-50(large) pytorch 1x 40.07 251.89 39.6 59.4 42.3

Instance segmentation with Mask R-CNN on MS-COCO 2017

model Params(M) FLOPs(G) AP AP_50 AP_75
EPSANet-50(small) 41.20 248.53 35.9 57.7 38.1
EPSANet-50(Large) 46.50 271.10 37.1 59.0 39.5

Citing EPSANet

You can cite the paper as:

@article{hu2021epsanet,
  title={EPSANet：An Efficient Pyramid Split Attention Block on Convolutional Neural Network},
  author={Hu Zhang and Keke Zu and Jian Lu and Yuru Zou and Deyu Meng},
  journal={arXiv preprint arXiv:2105.14447},
  year={2021}
}

GitHub

https://github.com/murufeng/EPSANet