Skip to content

README.md

Latest commit

 

History

History
75 lines (51 loc) · 2.42 KB

README.md

File metadata and controls

75 lines (51 loc) · 2.42 KB

[ICCV 2025 Highlight] VRM: Knowledge Distillation via Virtual Relation Matching

This repository contains a PyTorch implementation of Virtual Relation Matching (VRM) introduced in the paper VRM: Knowledge Distillation via Virtual Relation Matching (ICCV 2025).

Preparation

  1. Clone the repository to your local workspace:

    git clone https://github.com/VISION-SJTU/VRM.git

  2. Configure the environment:

    conda create --name vrm python=3.8
conda activate vrm
pip install torch==1.9.1+cu111 torchvision==0.10.1+cu111 torchaudio==0.9.1 -f https://download.pytorch.org/whl/torch_stable.html
pip install -r requirements.txt
python setup.py develop

    Note that other torch versions may also work.

  3. Prepare the data

    The CIFAR-100 dataset will be automatically downloaded to ./data/cifar100/. The ImageNet dataset shall be manually downloaded to ./data/imagenet/. Pretrained teachers on CIFAR-100 can be downloaded from https://github.com/megvii-research/mdistiller/releases/tag/checkpoints and should be unzipped to ./download_ckpts/cifar_teachers.

Training

To train on CIFAR-100, you may use:

  # resnet32x4-to-resnet8x4 distillation as an example
  python tools/train_vrm.py --cfg configs/cifar100/vrm/res32x4_res8x4.yaml

To train on ImageNet, you may use:

  # resnet34-to-resnet18 distillation as an example
  python tools/train_vrm.py --cfg configs/imagenet/r34_r18/vrm.yaml

To train on ImageNet using multiple GPUs, you may use:

  # resnet34-to-resnet18 distillation trained on 4 GPUs as an example
  python -m torch.distributed.launch --nproc_per_node=4 tools/train_vrm_dist.py --cfg configs/imagenet/r34_r18/vrm.yaml

Evaluation

The distilled student model will be automatically evaluated on the validation set during training. Manual evaluation is also supported by using eval.py.

Acknowledgement

This project is developed using the timm and the mdistiller library.

Reference

If you find this project useful, please consider citing it:

@inproceedings{zhang2025rsd,
  author    = {Weijia Zhang and Fei Xie and Weidong Cai and Chao Ma},
  title     = {VRM: Knowledge Distillation via Virtual Relation Matching},
  booktitle = {ICCV},
  year      = {2025}
}