Temporal-Adjusted Loss (TAL)

This repository contains the official implementation of the CVPR 2026 paper Temporal Imbalance of Positive and Negative Supervision in Class-Incremental Learning.

TAL studies temporal imbalance in class-incremental learning and dynamically reweights negative supervision in cross-entropy loss. The code framework is based on PyCIL, with TAL added as the core loss implementation and integrated into selected class-incremental learning baselines. This repository currently focuses on the reproduction paths used in our released experiments.

Slides: TAL slides

Status

The code is still being completed and cleaned up. Documentation and packaging will be improved over time.

Usage

Install the required Python dependencies, including PyTorch, torchvision, numpy, scipy, tqdm, scikit-learn, and POT.

From the repository root, run an experiment with:

python main.py --config=./exps/[CONFIG].json

The reproduction scripts in scripts/ automatically switch to the repository root. From the repository root, use:

# iCaRL on CIFAR-100
bash scripts/icarl.sh

# iCaRL + TAL on CIFAR-100
bash scripts/icarl_tal.sh

# DER on ImageNet-100
bash scripts/der.sh

# DER + TAL on ImageNet-100
bash scripts/der_tal.sh

For ImageNet-100, update the dataset paths or file lists in utils/data.py and data/imagenet_subset/ for your local environment.

Citation

If you find this repository useful, please cite:

@inproceedings{ma2026temporal,
  title={Temporal Imbalance of Positive and Negative Supervision in Class-Incremental Learning},
  author={Ma, Jinge and Zhu, Fengqing},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={32299--32308},
  year={2026}
}

Acknowledgments

This implementation is based on PyCIL. We thank the PyCIL authors for their open-source codebase.