ProtoGCD: Unified and Unbiased Prototype Learning for Generalized Category Discovery

Official implementation of our TPAMI 2025 paper "ProtoGCD: Unified and Unbiased Prototype Learning for Generalized Category Discovery".

🏃 ​Running

Dependencies

loguru
numpy
pandas
scikit_learn
scipy
torch==1.10.0
torchvision==0.11.1
tqdm

Datasets

We conduct experiments on 7 datasets:

• Generic datasets: CIFAR-10, CIFAR-100, ImageNet-100
• Fine-grained datasets: CUB, Stanford Cars, FGVC-Aircraft, Herbarium19

Config

Set paths to datasets in config.py

Training ProtoGCD

CIFAR100:

CUDA_VISIBLE_DEVICES=0 python train_fix.py --dataset_name 'cifar100' --batch_size 128 --epochs 200 --num_workers 4 --use_ssb_splits --weight_decay 5e-5 --lr 0.1 --eval_funcs 'v2' --weight_sup 0.35 --weight_entropy_reg 2 --weight_proto_sep 0.1 --temp_logits 0.1 --temp_teacher_logits 0.05 --wait_ratio_epochs 0 --ramp_ratio_teacher_epochs 100 --init_ratio 0.0 --final_ratio 1.0 --exp_name cifar100_protogcd

CUB:

CUDA_VISIBLE_DEVICES=0 python train_fix.py --dataset_name 'cub' --batch_size 128 --epochs 200 --num_workers 2 --use_ssb_splits --weight_decay 5e-5 --lr 0.1 --eval_funcs 'v2' --weight_sup 0.35 --weight_entropy_reg 2 --weight_proto_sep 0.05 --temp_logits 0.1 --temp_teacher_logits 0.05 --wait_ratio_epochs 0 --ramp_ratio_teacher_epochs 100 --init_ratio 0.0 --final_ratio 1.0 --exp_name cub_protogcd

Evaluate OOD detection

CIFAR:

CUDA_VISIBLE_DEVICES=0 python test_ood_cifar.py --dataset_name 'cifar100' --batch_size 128 --num_workers 4 --use_ssb_splits --num_to_avg 10 --score msp --ckpts_date YOUR_CKPTS_NAME --temp_logits 0.1

ImageNet:

CUDA_VISIBLE_DEVICES=0 python test_ood_imagenet.py --dataset_name 'imagenet_100' --batch_size 128 --num_workers 4 --use_ssb_splits --num_to_avg 10 --score msp --ckpts_date YOUR_CKPTS_NAME --temp_logits 0.1

📋 ​Citing this work

@article{ma2025protogcd,
  title={ProtoGCD: Unified and Unbiased Prototype Learning for Generalized Category Discovery},
  author={Ma, Shijie and Zhu, Fei and Zhang, Xu-Yao and Liu, Cheng-Lin},
  journal={IEEE Transactions on Pattern Analysis and Machine Intelligence},
  volume={47},
  number={07},
  pages={6022--6038},
  year={2025},
  publisher={IEEE Computer Society}
}

🎁 ​Acknowledgements

In building the ProtoGCD codebase, we reference SimGCD.

✅ ​License

This project is licensed under the MIT License - see the LICENSE file for details.

📧 ​Contact

If you have further questions or discussions, feel free to contact me:

Shijie Ma (mashijie2021@ia.ac.cn)