Multi-Label Knowledge Distillation

Penghui Yang^* · Ming-Kun Xie^* · Chen-Chen Zong · Lei Feng · Gang Niu · Masashi Sugiyama · Sheng-Jun Huang

Requirements

The repo is tested with:

mmcv==1.3.15

numpy==1.21.2

Pillow==9.2.0

randaugment==1.0.2

scikit_learn==1.1.2

timm==0.6.7

torch==1.8.1

torchvision==0.9.1

But it should be runnable with other PyTorch versions.

To install requirements:

pip install -r requirements.txt

Quick start

You can train on MS-COCO with default settings stored in ./configs/coco/resnet101_to_resnet34_l2d.py:

python main.py --cfg_file ./configs/coco/resnet101_to_resnet34_l2d.py --data_root [file path of MS-COCO stored on your own device]

You can also try your own distillers and other options by making your own configuration files under the guidance of Configuration files.

Use your own dataset

Your Pascal VOC 2007 dataset folder should be like this:

[Any name you want]
  |--VOCtrainval2007
    |--VOCdevkit
      |--VOC2007
        |--JPEGImages
          |--000005.jpg
          |--...
        |--ImageSets
          |--Main
            |--trainval.txt
  |--VOCtest2007
    |--VOCdevkit
      |--VOC2007
        |--JPEGImages
          |--000001.jpg
          |--...
        |--ImageSets
          |--Main
            |--test.txt

Your MS-COCO 2014 dataset folder should be like this:

[Any name you want]
  |--train2014
    |--COCO_train2014_000000000009.jpg
    |--...
  |--val2014
    |--COCO_val2014_000000000042.jpg
    |--...
  |--train_anno.json
  |--val_anno.json

train_anno.json and val_anno.json are in the fold ./appendix.

Your NUS-WIDE dataset folder should be like this:

[Any name you want]
  |--Flickr
    |--actor
      |--0001_2124494179.jpg
      |--0002_174174086.jpg
      |--...
    |--administrative_assistant
      |--0001_534152430.jpg
      |--0002_258761806.jpg
      |--...
    |--...
  |--ImageList
    |--Imagelist.txt
    |--TrainImagelist.txt
    |--TestImagelist.txt
  |--TrainTestLabels
    |--Labels_airport_Train.txt
    |--Labels_airport_Test.txt
    |--...
  |--Concepts81.txt

All codes of the data processing part are in the fold ./data, and you can replace them with your own code.

Configuration files

We use configuration files to pass parameters to the program. An example in the fold ./configs is shown below:

dataset = "coco"
teacher_pretrained = False
img_size = 224
batch_size = 64

model_t = "resnet101"
lr_t = 1e-4
max_epoch_t = 80
stop_epoch_t = 30

model_s = "resnet34"
lr_s = 1e-4
max_epoch_s = 80
stop_epoch_s = 80

criterion_t2s_para = dict(
    name="L2D",
    para=dict(
        lambda_ft=0.0,
        ft_dis=None,
        lambda_le=1.0,
        le_dis=dict(
            name="LED",
            para=dict(
                lambda_cd=100.0,
                lambda_id=1000.0
            )
        ),
        lambda_logits=10.0,
        logits_dis=dict(
            name="MLD",
            para=dict()
        )
    )
)

We split a distiller into three parts: feature-based part, label-wise embedding part and logits-based part. Each part has a balancing parameter lambda and corresponding parameters.

It is worth noting that you can set teacher_pretrained = True after you have already trained a teacher model and stored its weight parameters in order to avoid repetitive training and save your time.

Citation

If you find this repo useful for your research, please consider citing the paper.

@inproceedings{yang2023multi,
  title={Multi-Label Knowledge Distillation},
  author={Penghui Yang, Ming-Kun Xie, Chen-Chen Zong, Lei Feng, Gang Niu, Masashi Sugiyama, Sheng-Jun Huang},
  booktitle={International Conference on Computer Vision (ICCV)},
  year={2023},
}

Acknowledgement

This repo is partly based on the following repos, thank the authors a lot.