D3

This repo is the official of implementation of "Attention-modulated frequency-aware pooling via spatial guidance".

The implementation code of D3 is located at here.

Running

Install

We implement D3 using MMPretrain V1.2.0, MMDetection V3.3.0, MMSegmentation V1.2.2 and MMCV V2.1.0.
We train and test our models under python=3.10, pytorch=2.1.1, cuda=11.8.

# First, create a virtual environment and activate it.
pip install torch==2.1.1 torchvision==0.16.1 torchaudio==2.1.1 --index-url https://download.pytorch.org/whl/cu118
pip install -U openmim
mim install -e .

# Next, install and activate the virtual environment.
pip install virtualenv
cd D3
virtualenv venv
source venv/bin/activate

Data preparation

The ImageNet dataset should be prepared as follows:

$ tree data
imagenet
├── train
│   ├── class1
│   │   ├── img1.jpeg
│   │   ├── img2.jpeg
│   │   └── ...
│   ├── class2
│   │   ├── img3.jpeg
│   │   └── ...
│   └── ...
└── val
    ├── class1
    │   ├── img4.jpeg
    │   ├── img5.jpeg
    │   └── ...
    ├── class2
    │   ├── img6.jpeg
    │   └── ...
    └── ...

Training

Our work employs 8X GPUs for training on classification and object detection.
Here is an example: train D3 + ResNet-50 with an single GPU:

python ./tools/train.py work_dirs/resnet50_8xb32_in1k_D3.py --work-dir path_to_exp --amp 

Here is an example: train D3 + ResNet-50 with an multiple GPUs:

export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
export PORT=54321
bash ./tools/dist_train.sh work_dirs/resnet50_8xb32_in1k_D3.py 8 --work-dir path_to_exp --amp 

Testing

Test D3 + ResNet-50 with an single GPU:

python ./tools/test.py work_dirs/resnet50_8xb32_in1k_D3.py path_to_checkpoint --work-dir path_to_exp

Acknowledgement

The code in this repository is developed based on the MMPretrain. Furthermore, the detection task involved in this work are implemented based on the MMDetection and Ultralytics.

Cite D3

If you find this repository useful, please use the following BibTeX entry for citation.

@article{si2025attention,
  title={Attention-modulated frequency-aware pooling via spatial guidance},
  author={Si, Yunzhong and Xu, Huiying and Zhu, Xinzhong and Liu, Rihao and Li, Hongbo},
  journal={Neurocomputing},
  pages={129507},
  year={2025},
  publisher={Elsevier}
}

Concat

If you have any questions, please feel free to contact the authors.

Yunzhong Si: [email protected]