|
|
--- |
|
|
language: |
|
|
- en |
|
|
license: mit |
|
|
tags: |
|
|
- continual-learning |
|
|
- general-continual-learning |
|
|
- online-learning |
|
|
- vision |
|
|
- image-classification |
|
|
- vit |
|
|
- prompt-tuning |
|
|
library_name: pytorch |
|
|
pipeline_tag: image-classification |
|
|
inference: false |
|
|
--- |
|
|
|
|
|
# FlyGCL Checkpoints (FlyPrompt & ViT Baselines) |
|
|
|
|
|
This repository provides **research checkpoints** for **FlyGCL**, a lightweight framework for **General Continual Learning (GCL) / online class-incremental learning** in the **Si-Blurry** setting. |
|
|
|
|
|
It is designed to be used together with the FlyGCL codebase: |
|
|
|
|
|
- Code: `https://github.com/AnAppleCore/FlyGCL` |
|
|
- Paper (arXiv): `https://www.arxiv.org/abs/2602.01976` |
|
|
- OpenReview: `https://openreview.net/forum?id=8pi1rP71qv` |
|
|
|
|
|
## What is included |
|
|
|
|
|
This model repo may contain: |
|
|
|
|
|
- **Backbone checkpoints** (ViT-B/16 variants) referenced by FlyGCL via `--backbone`. |
|
|
- **Prompt checkpoints** (optional) for DualPrompt/MISA-style prompts: |
|
|
- `g_prompt.pt` |
|
|
- `e_prompt.pt` |
|
|
|
|
|
For the exact filename mapping and where to place these files in FlyGCL, see the code repository README: |
|
|
|
|
|
- `https://github.com/AnAppleCore/FlyGCL/blob/main/README.md` |
|
|
|
|
|
## Model details |
|
|
|
|
|
- **Architecture family**: Vision Transformer (ViT-B/16) backbones + prompt-based continual learning heads. |
|
|
- **Framework**: PyTorch. |
|
|
- **Training setting**: online GCL / Si-Blurry (see paper and code for details). |
|
|
|
|
|
## Intended use |
|
|
|
|
|
These checkpoints are released for: |
|
|
|
|
|
- **Research / reproducibility** of the FlyGCL paper and baselines |
|
|
- **Benchmarking** continual learning methods in comparable settings |
|
|
|
|
|
Not intended for: |
|
|
|
|
|
- Safety-critical or medical/diagnostic use |
|
|
- Deployment without careful evaluation in your target environment |
|
|
|
|
|
## Limitations and biases |
|
|
|
|
|
- Continual learning performance depends on data ordering, hyperparameters, and backbone initialization. |
|
|
- Backbones pretrained on large-scale datasets may encode biases from their pretraining data. |
|
|
- Prompt checkpoints may not transfer to datasets/settings different from those used during training. |
|
|
|
|
|
## License |
|
|
|
|
|
- Code license: MIT (see FlyGCL `LICENSE`). |
|
|
- **Checkpoint licensing** may depend on upstream sources (e.g., DINO/iBOT/MoCo pretrained backbones). If you redistribute upstream-derived weights here, ensure the redistribution terms are compatible and include required notices. |
|
|
|
|
|
## Citation |
|
|
|
|
|
If you use FlyGCL or these checkpoints in your research, please cite: |
|
|
|
|
|
```bibtex |
|
|
@inproceedings{flyprompt2026, |
|
|
title={FlyPrompt: Brain-Inspired Random-Expanded Routing with Temporal-Ensemble Experts for General Continual Learning}, |
|
|
author={Yan, Hongwei and Sun, Guanglong and Zhou, Kanglei and Li, Qian and Wang, Liyuan and Zhong, Yi}, |
|
|
booktitle={ICLR}, |
|
|
year={2026} |
|
|
} |
|
|
``` |
|
|
|
|
|
## Contact |
|
|
|
|
|
- Maintainer: `Hongwei Yan` (`yanhw22@mails.tsinghua.edu.cn`) |
|
|
|
