The code is built on mdistiller

LumiNet: Perception-Driven Knowledge Distillation via Statistical Logit Calibration

Embarrassingly simple knowledge distillation method

Framework & Performance

CIFAR-100 Benchmark Results ( Same Architecture ):

Teacher Student	ResNet56 ResNet20	ResNet110 ResNet32	ResNet32x4 ResNet8x4	WRN-40-2 WRN-16-2	WRN-40-2 WRN-40-1	VGG13 VGG8
KD	70.66	73.08	73.33	74.92	73.54	72.98
LumiNet	72.29	74.20	77.50	76.38	75.12	74.94

CIFAR-100 Benchmark Results (Heterogeneous architecture):

Teacher Student	ResNet32x4 ShuffleNet-V1	WRN-40-2 ShuffleNet-V1	VGG13 MobileNet-V2	ResNet50 MobileNet-V2	ResNet32x4 ShuffleNet-V2
KD	74.07	74.83	67.37	67.35	74.45
LumiNet	76.66	76.95	70.50	70.97	77.55

ImageNet Benchmark Results:

On ImageNet:

Teacher Student	ResNet34 ResNet18	ResNet50 MobileNet-V1
KD	71.03	70.50
LumiNet	72.16	72.55

Installation

Supported Environments:

• Python version 3.6
• PyTorch version 1.9.0
• torchvision version 0.10.0

Install the package:

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

Getting Started

1. Wandb Integration for Logging

• Register on Wandb: Wandb Registration.
• To opt-out of Wandb logging, set CFG.LOG.WANDB to False in mdistiller/engine/cfg.py.

2. Evaluation

• Evaluate the performance of provided or custom-trained models.

• Download our models from this link and save the checkpoints in ./download_ckpts.

• For ImageNet testing, download the dataset from ImageNet and place it in ./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/luminet_resnet8x4 # luminet-resnet8x4 on cifar100
  python3 tools/eval.p -m MobileNetV1 -c download_ckpts/imgnet_luminet_mv1 -d imagenet # luminet-mv1 on imagenet
  python3 tools/eval.p -m model_name -c output/your_exp/student_best # your checkpoints

3. 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 LumiNet method.
  python3 tools/train.py --cfg configs/cifar100/luminet/res32x4_res8x4.yaml
  
  # you can also change settings at command line
  python3 tools/train.py --cfg configs/cifar100/luminet/res32x4_res8x4.yaml SOLVER.BATCH_SIZE 128 SOLVER.LR 0.1

4. Training on ImageNet

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

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

5. Training on MS-COCO ( This part will be released soon)

• see detection.md

6. Extension: Visualizations

• Jupyter notebooks: tsne and correlation_matrices

Citation

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

hossain2025luminet,
title={LumiNet: Perception-Driven Knowledge Distillation via Statistical Logit Calibration},
author={Md. Ismail Hossain and M M Lutfe Elahi and Sameera Ramasinghe and Ali Cheraghian and Fuad Rahman and Nabeel Mohammed and Shafin Rahman},
journal={Transactions on Machine Learning Research},
issn={2835-8856},
year={2025},
url={https://openreview.net/forum?id=3rU1lp9w2l},
note={}
}

Acknowledgement

• We would like to extend our sincere appreciation to the contributors of mdistiller for their dedicated efforts and significant contributions.