Budgeted Online Continual Learning by Adaptive Layer Freezing and Frequency-based Sampling
Minhyuk Seo*, Hyunseo Koh*, Jonghyun Choi
ICLR 2025 (Spotlight)

We provide the official implementation of the proposed aL-SAR and baselines.

git clone https://github.com/snumprlab/budgeted-cl.git

conda create -n budgeted_cl python=3.10
conda activate budgeted_cl
pip install -r requirements.txt

CIFAR-10/100, CLEAR100, Bongard-HOI, and Bongard-OpenWorld can be downloaded by running the corresponding scripts in the dataset/ directory. ImageNet dataset can be downloaded from Kaggle.

First, activate the training environment budgeted_cl.

conda deactivate
conda activate budgeted_cl

Experiments for the implemented methods can be run by executing train_VLM.sh by

bash train_VLM.sh

You may change various arguments for different experiments.

• NOTE: Short description of the experiment. Experiment result and log will be saved at results/DATASET/NOTE.
  • WARNING: logs/results with the same dataset and note will be overwritten!
• MODEL_ARCH: VLM model types. Supported VLMs are: [llava, bunny_3b, bunny_8b]
• DATASET: Dataset to use in experiment. Supported datasets for multi-modal concept-IL are: [Bongard-HOI, Bongard-OpenWorld] Supported datasets for CIL are: [cifar10, cifar100, clear10, clear100, imagenet]
• MEM_SIZE: Maximum number of samples in the episodic memory.
• NUM_ITER: Number of model updates per sample.
• NUM_SET: Number of samples in each positive and negative support set for Bongard benchmarks.
• RND_SEED: Random seed for reproducibility.
• DATA_TYPE: Data source type. The default is manually annotated data. Supported types: [ma_ver3_more_text, generated, web].
• Ours: Enable our proposed adaptive layer-freezing (aL).
• SAR: Enable our proposed Similarity-aware Retrieval (SAR).

Note: All hyperparameters used for the experiments in the paper are set as default.

First, activate the evaluation environment budgeted_cl.

conda deactivate
conda activate budgeted_cl

To evaluate a model, run eval_VLM.sh by

bash train_VLM.sh

You may change various arguments for different experiments.

• RND_SEED: Random seed (must be kept the same during training).
• NOTE: Short description of the experiment (must be kept the same during training to preserve the model path directory).
• NUM_SET: Number of samples in each positive and negative support set for Bongard benchmarks.

GNU GENERAL PUBLIC LICENSE

aL-SAR

@inproceedings{
  seo2025budgeted,
  title={Budgeted Online Continual Learning by Adaptive Layer Freezing and Frequency-based Sampling},
  author={Minhyuk Seo and Hyunseo Koh and Jonghyun Choi},
  booktitle={The Thirteenth International Conference on Learning Representations},
  year={2025},
  url={https://openreview.net/forum?id=dOAkHmsjRX}
}

This work was partly supported by AI Center, Samsung Electronics, and the IITP grants (No.RS2022-II220077, No.RS-2022-II220113, No.RS-2022-II220959, No.RS-2021-II211343 (SNU AI),
No.RS-2021-II212068 (AI Innov. Hub)) funded by the Korea government (MSIT).