RLCD: Generalized Category Discovery via Reciprocal Learning and Class-Wise Distribution Regularization

This repository contains the official implementation of our ICML 2025 paper:

Generalized Category Discovery via Reciprocal Learning and Class-Wise Distribution Regularization
Duo Liu, Zhiquan Tan, Linglan Zhao, Zhongqiang Zhang, Xiangzhong Fang, Weiran Huang

---

📖 Overview

Generalized Category Discovery (GCD) aims to identify unlabeled samples by leveraging base knowledge from labeled ones, where the unlabeled set consists of both base and novel classes.
Our method, RLCD, introduces:

• Reciprocal Learning Framework (RLF): An auxiliary branch devoted to base classification, forming a virtuous cycle with the main branch.
• Class-Wise Distribution Regularization (CDR): Mitigates bias towards base classes and boosts novel class performance.

Together, RLCD achieves superior performance across multiple GCD benchmarks with negligible extra computation.

📂 Dataset Preparation

Download and preprocess datasets following instructions in SimGCD.

🚀 Quick Start

python train.py -d cifar100     -m 2.0 --fp16 --cross_w 0.5 --dw 0.5  --difflr
python train.py -d cub          -m 2.0 --fp16 --cross_w 0.5 --dw 0.5  --difflr
python train.py -d scars        -m 2.0 --fp16 --cross_w 0.5 --dw 1.0
python train.py -d herbarium_19 -m 2.0 --fp16 --cross_w 0.5 --dw 1.0 --fc_temp 0.07 --eval_funcs v2 v2b

⚙️ Hyperparameters and Tuning Tips

As reported in the paper, the class-wise distribution regularization weight dw and the distillation loss weight cross_w are not heavily tuned.

In practice, we recommend:

• Starting from the default configurations provided above;
• Adjusting dw and cross_w based on evaluation performance;

📜 Citation

If you find this repository or our paper helpful in your research, please consider citing:

@inproceedings{liu2025rlcd,
  title={Generalized Category Discovery via Reciprocal Learning and Class-Wise Distribution Regularization},
  author={Liu, Duo and Tan, Zhiquan and Zhao, Linglan and Zhang, Zhongqiang and Fang, Xiangzhong and Huang, Weiran},
  booktitle={Proceedings of the 42nd International Conference on Machine Learning (ICML)},
  year={2025}
}