Happy: A Debiased Learning Framework for Continual Generalized Category Discovery

Official implementation of our NeurIPS 2024 paper: Happy: A Debiased Learning Framework for Continual Generalized Category Discovery [arXiv]

We study the under-explored setting of continual generalized category discovery (C-GCD) as follows:

🔖 The core difference between C-GCD and class-incremental learning (CIL) is that C-GCD is unsupervised continual learning, while CIL is purely supervised. At each continual stage of C-GCD, unlabeled training data could contain both old and new classes.

We introduce our method: Happy, which is characterized by Hardness-aware prototype sampling and soft entropy regularization, as follows:

Running 🏃

1. Datasets

We conduct experiments on 6 datasets:

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

2. Config

Set paths to datasets in config.py

3. Stage-0: Labeled Training

CIFAR100

CUDA_VISIBLE_DEVICES=0 python train_happy.py --dataset_name 'cifar100' --batch_size 128 --transform 'imagenet' --lr 0.1 --memax_weight 1 --eval_funcs 'v2' --num_old_classes 50 --prop_train_labels 0.8 --train_session offline --epochs_offline 100 --continual_session_num 5 --online_novel_unseen_num 400 --online_old_seen_num 25 --online_novel_seen_num 25

CUB

CUDA_VISIBLE_DEVICES=0 python train_happy.py --dataset_name 'cub' --batch_size 128 --transform 'imagenet' --lr 0.1 --memax_weight 2 --eval_funcs 'v2' --num_old_classes 100 --prop_train_labels 0.8 --train_session offline --epochs_offline 100 --continual_session_num 5 --online_novel_unseen_num 25 --online_old_seen_num 5 --online_novel_seen_num 5

⚠️ Please specify the args --train_session as offline

4. Stage-1 $\sim$ T: Continual Training

CIFAR100

CUDA_VISIBLE_DEVICES=0 python train_happy.py --dataset_name 'cifar100' --batch_size 128 --transform 'imagenet' --warmup_teacher_temp 0.05 --teacher_temp 0.05 --warmup_teacher_temp_epochs 10 --lr 0.01 --memax_old_new_weight 1 --memax_old_in_weight 1 --memax_new_in_weight 1 --proto_aug_weight 1 --feat_distill_weight 1 --radius_scale 1.0 --hardness_temp 0.1 --eval_funcs 'v2' --num_old_classes 50 --prop_train_labels 0.8 --train_session online --epochs_online_per_session 30 --continual_session_num 5 --online_novel_unseen_num 400 --online_old_seen_num 25 --online_novel_seen_num 25 --init_new_head --load_offline_id Old50_Ratio0.8_20240418-002807 --shuffle_classes --seed 0

CUB

CUDA_VISIBLE_DEVICES=0 python train_happy.py --dataset_name 'cub' --batch_size 128 --transform 'imagenet' --warmup_teacher_temp 0.05 --teacher_temp 0.05 --warmup_teacher_temp_epochs 10 --lr 0.01 --memax_old_new_weight 1 --memax_old_in_weight 1 --memax_new_in_weight 1 --proto_aug_weight 1 --feat_distill_weight 1 --radius_scale 1.0 --eval_funcs 'v2' --num_old_classes 100 --prop_train_labels 0.8 --train_session online --epochs_online_per_session 20 --continual_session_num 5 --online_novel_unseen_num 25 --online_old_seen_num 5 --online_novel_seen_num 5 --init_new_head --load_offline_id Old100_Ratio0.8_20240506-165445 --shuffle_classes --seed 0

⚠️ Please specify the args --train_session as online

⚠️ Please change the args --load_offline_id according to offline training save path

⚠️ Please keep the four args --continual_session_num, --online_novel_unseen_num, --online_old_seen_num and --online_novel_seen_num the same as offline training stage.

Citing this work 📋

@article{ma2024happy,
  title={Happy: A Debiased Learning Framework for Continual Generalized Category Discovery},
  author={Ma, Shijie and Zhu, Fei and Zhong, Zhun and Liu, Wenzhuo and Zhang, Xu-Yao and Liu, Cheng-Lin},
  journal={arXiv preprint arXiv:2410.06535},
  year={2024}
}

Acknowledgements 🎁

In building this repository, 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 ([email protected])