EPSANet：An Efficient Pyramid Split Attention Block on Convolutional Neural Network

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

• Install CUDA==10.0 with cudnn7 following the official installation instructions
• Install PyTorch==1.0.1 and torchvision==0.2.0 with CUDA==10.0:

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

Object Detection on MS-COCO 2017

Faster R-CNN

Mask R-CNN

RetinaNet

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

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}
}