File size: 2,784 Bytes
0d554b3 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 |
---
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`)
|