Large Separable Kernel Attention (LSKA)

This repository implements the model proposed in the paper:

Kin Wai Lau, Lai-Man Po, Yasar Abbas Ur Rehman, Large Separable Kernel Attention: Rethinking the Large Kernel Attention Design in CNN

[arXiv paper]

The implementation code is based on the Visual Attention Network (VAN), Computational Visual Media, 2023. For more information, please refer to the link.

Citing

When using this code, kindly reference:

@article{lau2024large,
  title={Large Separable Kernel Attention: Rethinking the Large Kernel Attention Design in CNN},
  author={Lau, Kin Wai and Po, Lai-Man and Rehman, Yasar Abbas Ur},
  journal={Expert Systems with Applications},
  volume={236},
  pages={121352},
  year={2023},
  publisher={Elsevier}
}

Pretrained models

You can download our pretrained models on ImageNet-1K:

Pretrained ImageNet-1K LSKA (kernel size: 23): link.

Preparation

• Requirements:

1. Pytorch >= 1.7
2. timm == 0.4.12

Train

We use 4 GPUs for training by default. Run command (It has been writen in train_lska.sh):

MODEL=van_tiny # van_{tiny, small, base}
DROP_PATH=0.1 # drop path rates [0.1, 0.1, 0.1] for [tiny, small, base]
# Kernel size should be [7, 11, 23, 35, 53]
CUDA_VISIBLE_DEVICES=0,1,2,3 bash distributed_train.sh 4 /path/to/imagenet \
	  --model $MODEL -b 128 --lr 1e-3 --drop-path $DROP_PATH --k_size kernel_size \
	  --log_name log_file_name

Validate

Run command (It has been writen in eval.sh) as:

MODEL=van_tiny # van_{tiny, small, base}
CUDA_VISIBLE_DEVICES=0 python3 validate.py /path/to/imagenet --model $MODEL --k_size kernel_size \
  --checkpoint /path/to/model -b 128

Object Detection and Segmentation

We also include the configuration file for the object detection and semantic segmentation in this repository. For details, please check folder mmdetection and mmsegmentation.