This repo is

(1) a PyTorch library that provides classical knowledge distillation algorithms on mainstream CV benchmarks,

(2) the official implementation of the CVPR-2022 paper: Decoupled Knowledge Distillation.

(3) the official implementation of the ICCV-2023 paper: DOT: A Distillation-Oriented Trainer.

On CIFAR-100:

On Tiny-ImageNet:

On ImageNet:

On CIFAR-100:

On ImageNet:

MDistiller supports the following distillation methods on CIFAR-100, ImageNet and MS-COCO:

Environments:

• Python 3.6
• PyTorch 1.9.0
• torchvision 0.10.0

Install the package:

sudo pip3 install -r requirements.txt
sudo python3 setup.py develop

0. Wandb as the logger

• The registeration: https://wandb.ai/home.
• If you don't want wandb as your logger, set CFG.LOG.WANDB as False at mdistiller/engine/cfg.py.

1. Evaluation

• You can evaluate the performance of our models or models trained by yourself.

• Our models are at https://github.com/megvii-research/mdistiller/releases/tag/checkpoints, please download the checkpoints to ./download_ckpts

• If test the models on ImageNet, please download the dataset at https://image-net.org/ and put them to ./data/imagenet

  # evaluate teachers
  python3 tools/eval.py -m resnet32x4 # resnet32x4 on cifar100
  python3 tools/eval.py -m ResNet34 -d imagenet # ResNet34 on imagenet
  
  # evaluate students
  python3 tools/eval.p -m resnet8x4 -c download_ckpts/dkd_resnet8x4 # dkd-resnet8x4 on cifar100
  python3 tools/eval.p -m MobileNetV1 -c download_ckpts/imgnet_dkd_mv1 -d imagenet # dkd-mv1 on imagenet
  python3 tools/eval.p -m model_name -c output/your_exp/student_best # your checkpoints

2. Training on CIFAR-100

• Download the cifar_teachers.tar at https://github.com/megvii-research/mdistiller/releases/tag/checkpoints and untar it to ./download_ckpts via tar xvf cifar_teachers.tar.

  # for instance, our DKD method.
  python3 tools/train.py --cfg configs/cifar100/dkd/res32x4_res8x4.yaml
  
  # you can also change settings at command line
  python3 tools/train.py --cfg configs/cifar100/dkd/res32x4_res8x4.yaml SOLVER.BATCH_SIZE 128 SOLVER.LR 0.1

3. Training on ImageNet

• Download the dataset at https://image-net.org/ and put them to ./data/imagenet

  # for instance, our DKD method.
  python3 tools/train.py --cfg configs/imagenet/r34_r18/dkd.yaml

4. Training on MS-COCO

• see detection.md

5. Extension: Visualizations

• Jupyter notebooks: tsne and correlation_matrices

1. create a python file at mdistiller/distillers/ and define the distiller

from ._base import Distiller

class MyDistiller(Distiller):
    def __init__(self, student, teacher, cfg):
        super(MyDistiller, self).__init__(student, teacher)
        self.hyper1 = cfg.MyDistiller.hyper1
        ...

    def forward_train(self, image, target, **kwargs):
        # return the output logits and a Dict of losses
        ...
    # rewrite the get_learnable_parameters function if there are more nn modules for distillation.
    # rewrite the get_extra_parameters if you want to obtain the extra cost.
  ...

2. regist the distiller in distiller_dict at mdistiller/distillers/__init__.py

3. regist the corresponding hyper-parameters at mdistiller/engines/cfg.py

4. create a new config file and test it.

If this repo is helpful for your research, please consider citing the paper:

@article{zhao2022dkd,
  title={Decoupled Knowledge Distillation},
  author={Zhao, Borui and Cui, Quan and Song, Renjie and Qiu, Yiyu and Liang, Jiajun},
  journal={arXiv preprint arXiv:2203.08679},
  year={2022}
}
@article{zhao2023dot,
  title={DOT: A Distillation-Oriented Trainer},
  author={Zhao, Borui and Cui, Quan and Song, Renjie and Liang, Jiajun},
  journal={arXiv preprint arXiv:2307.08436},
  year={2023}
}

MDistiller is released under the MIT license. See LICENSE for details.

• Thanks for CRD and ReviewKD. We build this library based on the CRD's codebase and the ReviewKD's codebase.

• Thanks Yiyu Qiu and Yi Shi for the code contribution during their internship in MEGVII Technology.

• Thanks Xin Jin for the discussion about DKD.