Skip to content

icey-zhang/GHOST

Repository files navigation

Guided Hybrid Quantization for Object Detection in Multimodal Remote Sensing Imagery via One-to-one Self-teaching

⭐ The code matches our paper article!!!⭐

If our code is helpful to you, please cite:


@article{zhang2023guided,
  title={Guided Hybrid Quantization for Object Detection in Remote Sensing Imagery via One-to-one Self-teaching},
  author={Zhang, Jiaqing and Lei, Jie and Xie, Weiying and Li, Yunsong and Yang, Geng and Jia, Xiuping},
  journal={IEEE Transactions on Geoscience and Remote Sensing},
  year={2023},
  publisher={IEEE}
}

Requirements

pip install -r requirements.txt

Use GHOST

1. Prepare training data

  • 1.1 The publicly available dataset VEDAI designed for multimodal remote sensing image object detection is adopted in our experiments. In addition to validation on the multimodal object detection dataset, three single modal datasets (DOTA, NWPU and DIOR ) are utilized in experiments to verify the generation of our proposed algorithm.
  • 1.2 Download VEDAI data for our experiment from baiduyun (code: hvi4) or google drive. And the path of dataset is like that
  • 1.3 The process of the dataset can be seen in here transform_xxx.py

2. begin to train

Here we take the NWPU dataset as an example.
  • 2.1 Train a small full-precision model as the pretrained teacher model (SuperYOLO)

    python train.py --data data/NWPU.yaml --cfg models/SRyolo_noFocus.yaml --ch 3 --input_mode RGB --batch-size 8 --epochs 150 --train_img_size 1024 --test_img_size 512 --device 0

    result: You will get a weight_SuperYOLO, which serves as a network of teachers to guide the optimization process of quantifying the network.

  • 2.2 Get a small mixed-bit model (GHOST) by completing n-bit quantization under the guidance of pretrained teacher via distillation

      <!-- note: --weights teacher==weight_superYOLO --weight==weight_SuperYOLO -->
    
    python quantization_conv_automix_autodis.py --distillation 6 --inter_threshold 0.1 --device 0 --kd_weight 400 --epochs 150 --data data/NWPU.yaml --weights_teacher runs/train/use/exp/weights/best.pt --weights runs/train/use/exp/weights/best.pt --cfg models/SRyolo_noFocus.yaml --ch 3 --input_mode RGB --batch-size 8 --hyp data/hyp.scratch.yaml --train_img_size 1024 --test_img_size 512

    result:You will get a weight_GHOST.

3. test

  • 3.1 test Parameters and mAP of the SuperYOLO

    <!-- note: --weights == weight_GHOST --full weights==weight_SuperYOLO  -->
    <!-- note:if you want to get n-bit SuperYOLO you can set -- bit_width==n  -->
    
    python test_flops.py --weights runs/train/use/exp/weights/best.pt --full_weights runs/train/use/exp/weights/best.pt --bit_width 32 --input_mode RGB 
    python test.py --data data/NWPU.yaml --weights runs/train/exp/weights/best.pt --batch-size 4 --device 0 --iou-thres 0.6
  • 3.2 test Parameters and mAP of the GHOST

    python test_flops.py --weights runs/train/use/exp1/weights/best.pt --full_weights runs/train/use/exp/weights/best.pt --input_mode RGB --inter-threshold 0.1
    python test.py --data data/NWPU.yaml --weights runs/train/exp1/weights/best.pt --batch-size 4 --device 0 --iou-thres 0.6

4. If you want to use other datasets, you can try :

The DOTA, DIOR, and VEDAI

Time

2023.2.14 open the code

Acknowledgements

This code is built on YOLOv5 (PyTorch). We thank the authors for sharing the codes.

Thanks for the code of Dota dataset processing DOTA_devkit_YOLO.

Contact

If you have any questions, please contact me by email (jqzhang_2@stu.xidian.edu.cn).