RemoteTrimmer: Adaptive Structural Pruning for Remote Sensing Image Classification

Guanwenjie Zou (邹广文杰)* ,     Liang Yao (姚亮)* ,     Fan Liu (刘凡) ✉ ,     Chuanyi Zhang (张传一) ,    

Xin Li (李鑫) ,     Ning Chen (陈宁) ,     Shengxiang Xu (徐圣翔) ,     Jun Zhou (周峻)

News

• 2024/12/21: The paper is accepted by ICASSP 2025!
• 2024/07/29: We propose an efficient structural pruning method for remote sensing image classification. Codes and models will be open-sourced at this repository.

Contact

Please Contact yaoliang@hhu.edu.cn

Installation

Our pruning method utiles the torch-pruning framework, which is compatible with both PyTorch 1.x and 2.x versions.

pip install torch-pruning

How It Works

We leverage intermediate outputs from SENet and scaling factors from the BN layer to map channel importance into the attention space. During the post-pruning fine-tuning phase, we design a lateral inhibition loss function to emphasize difficult samples. Our method effectively addresses two key challenges in remote sensing model pruning: the lack of distinct channel importance and the prevalence of difficult samples.

A Example of Our Method

loading resnet18 model with SENet

import torch
from resnet18_SE.py import resnet18_SE

model = resnet18_SE(class_num) # class_num is the number of types of datasets

training model on EuroSAT datasets

python train.py

extracting the intermediate outputs of SENet

python get_attention.py

deleting SENet upon resnet18 and updating BN layer parameters

python delete_SE_resnet18.py

pruning model

python pruning.py

fine-tuning pruned model with Adaptive Mining Loss function

from Adaptive_Mining_Loss import MyLoss

loss_fn = MyLoss(r1=1, r2=1)