AS-MLP architecture for Image Classification

This repo is the official implementation of our ICLR2022 paper "AS-MLP: An Axial Shifted MLP Architecture for Vision" (arXiv).

Model Zoo

Image Classification on ImageNet-1K

Network	Resolution	Top-1 (%)	Params	FLOPs	Throughput (image/s)	model
AS-MLP-T	224x224	81.3	28M	4.4G	1047	onedrive
AS-MLP-S	224x224	83.1	50M	8.5G	619	onedrive
AS-MLP-B	224x224	83.3	88M	15.2G	455	onedrive

Usage

Install

• Clone this repo:

git clone https://github.com/svip-lab/AS-MLP
cd AS-MLP

• Create a conda virtual environment and activate it:

conda create -n asmlp python=3.7 -y
conda activate asmlp

• Install CUDA==10.1 with cudnn7 following the official installation instructions
• Install PyTorch==1.7.1 and torchvision==0.8.2 with CUDA==10.1:

conda install pytorch==1.7.1 torchvision==0.8.2 cudatoolkit=10.1 -c pytorch

• Install timm==0.3.2:

pip install timm==0.3.2

• Install cupy-cuda101:

pip install cupy-cuda101

• Install Apex:

git clone https://github.com/NVIDIA/apex
cd apex
pip install -v --disable-pip-version-check --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" ./

• Install other requirements:

pip install opencv-python==4.4.0.46 termcolor==1.1.0 yacs==0.1.8

Evaluation

To evaluate a pre-trained AS-MLP on ImageNet val, run:

bash train_scripts/test.sh

Training from scratch

To train a AS-MLP on ImageNet from scratch, run:

bash train_scripts/train.sh

You can easily reproduce our results. Enjoy!

Throughput

To measure the throughput, run:

bash train_scripts/get_throughput.sh

Citation

If this project is helpful for you, you can cite our paper:

@InProceedings{Lian_2021_ASMLP,
	title={AS-MLP: An Axial Shifted MLP Architecture for Vision},
	author={Lian, Dongze and Yu, Zehao and Sun, Xing and Gao, Shenghua},
	booktitle={International Conference on Learning Representations (ICLR)},
	year={2022}
}

Other Links

Object Detection and Instance Segmentation: See AS-MLP for Object Detection.

Semantic Segmentation: See AS-MLP for Semantic Segmentation.

Acknowledgement

The code is built upon Swin-Transformer, the cuda kernel is modified from Involution.