DenoiseRep: Denoising Model for Representation Learning

DenoiseRep is a computation-free, label-optional and model-irrelevant algorithm to incrementally improve representation learning.

Paper | Slides | Poster

DenoiseRep: Denoising Model for Representation Learning.
Zhengrui Xu&, Guan'an Wang&^, Xiaowen Huang*, Jitao Sang.
NeurIPS 2024 (Oral)

&Equal Contribution
^Project Lead
*Contact Author

Updates

• 2024.12.27: code released. thanks to Zhengrui Xu's contribution, who is the code developer.
• 2024.10.28: init project, code coming soon.

TODO

• release DenoiseRep basic code.
• implement DenoiseLinear.
• implement DenoiseConv2d.
• a tutorial of cifar10.
• implement Person-ReID experiments.
• implement Classification (ImageNet) experiments.
• implement Detection / Segmentation experiments.

Pipeline

Experimental Results

Tasks	Model	Backbone	Dataset	Metric	Baseline	+DenoiseRep
Classification	ViT	ViT patch=4	Cifar-10	acc@1	85.6%	86.2% (model)(log)
Classification	Swin-Transformer	Swinv2-T	ImageNet	acc@1	81.8%	82.1% (model)(log)
Person-ReID	TransReID-SSL	ViT-S	MSMT17	mAP	66.3%	67.3% (model)(log)

Installation RenoiseRep Lib

cd denoiserep_op
bash make.sh
pip show denoiserep

3 Steps to apply DenoiseRep to your model

load your model trained with the original pipeline, and convert to denoiserep.

train your model by adding ploss.

a training trick to obtain better performance.

please see more details by comparing train_cifar10.py and train_cifar10_denoise.py

Citing DenoiseRep

If you find denoise-rep useful in your research, please consider citing:

@inproceedings{xu2024denoiserep,
    title={DenoiseRep: Denoising Model for Representation Learning},
    author={zhengrui Xu and Guan'an Wang and Xiaowen Huang and Jitao Sang},
    booktitle={The Thirty-eighth Annual Conference on Neural Information Processing Systems},
    year={2024},
    url={https://openreview.net/forum?id=OycU0bAus6}
}

Acknowledgement

Our implementation is mainly based on the following codebases. We gratefully thank the authors for their wonderful works.

TransReID-SSL, Swin-Transformer, mmdetection, mmsegmentation.