Introduction to LibFewShot
LibFewShot is a groundbreaking open-source library designed to streamline the process of few-shot learning, a rapidly evolving field in machine learning focused on training models using a minimal amount of sample data. Established by a team of researchers, including Wenbin Li, Ziyi Wang, Xuesong Yang, and others, and featured in TPAMI 2023, this library provides comprehensive tools that help both researchers and practitioners quickly grasp and implement various few-shot learning methods.
Supported Methods
LibFewShot supports a variety of methods categorized into three distinct approaches: Non-episodic (or fine-tuning), meta-learning, and metric-learning methods. Each category offers a range of techniques originating from various high-impact research papers:
Non-episodic Methods
These methods focus on fine-tuning model parameters:
- Baseline & Baseline++ (ICLR 2019)
- RFS (ECCV 2020)
- SKD (BMVC 2021)
- Negcos (ECCV 2020)
- S2M2 (WACV 2020)
- Meta-Baseline (ICCV 2021)
- Diffkendall (NeurIPS 2023)
Meta-learning Based Methods
These methods aim for model optimization that can quickly adapt to new data environments:
- MatchingNet (NeurIPS 2016)
- MAML (ICML 2017)
- Versa (NeurIPS 2018)
- R2D2 (ICLR 2019)
- LEO (ICLR 2019)
- MTL (CVPR 2019)
- ANIL (ICLR 2020)
- IFSL (NeurIPS 2020)
- BOIL (ICLR 2021)
- MeTAL (ICCV 2021)
Metric-learning Based Methods
These methods focus on measuring similarities or differences between data points:
- ProtoNet (NeurIPS 2017)
- RelationNet (CVPR 2018)
- ConvaMNet (AAAI 2019)
- DN4 (CVPR 2019)
- CAN (NeurIPS 2019)
- ATL-Net (IJCAI 2020)
- ADM (IJCAI 2020)
- DSN (CVPR 2020)
- FEAT (CVPR 2020)
- RENet (ICCV 2021)
- FRN (CVPR 2021)
- DeepBDC (CVPR 2022)
- MCL (CVPR 2022)
- CPEA (ICCV 2023)
Quick Installation and Reproduction
LibFewShot can be easily installed by following the guidelines provided in their comprehensive installation guide. It also offers well-documented tutorials to facilitate the reproduction of experiments, aiding users in replicating the accuracies reported by various studies. Further, it provides links to download model checkpoints and config files for different methods.
Accessible Datasets
The library includes multiple dataset links essential for few-shot learning, such as the Caltech-UCSD Birds-200-2011, Standford Cars, Standford Dogs, miniImageNet, tieredImageNet, and WebCaricature datasets. These are readily available through Google Drive and Baidu.
Contributing and Licensing
The project invites contributions and improvements from the community, adhering to the PEP 8 coding style. For those interested in contributing, detailed guidelines are provided in their contributing guide. LibFewShot is offered under the MIT License, encouraging its use in academic research.
Acknowledgement and Usage
LibFewShot stands as a testament to the effort to simplify and enhance few-shot learning research. It warmly acknowledges contributions from researchers and learners while providing an open platform for further development. The project welcomes feedback to further enhance its usability and coverage in few-shot learning. It should be noted that while LibFewShot supports academic pursuits, it is crucial to cite their work appropriately in any research utilizing this library.