Announcing MMFewShot: The first few shot learning toolbox for classification and detection

Few Shot learning is generally a kind of meta learning, whose goal is to train *accurate* models with very limited training data. Initially starting from image classification tasks, the scope of few shot learning, as a research hotspot, has been extended to many more different vision tasks such as detection and segmentation. However, many of these algorithms are not published with code. Interested researchers who want to incorporate them into their researches have to spend a lot of time reimplementing them. On the other hand, some other works with open-sourced implementations are scattered over the Internet. Different code styles and structures prevent researchers from systematically learning the latest works and conducting experiments across these algorithms.

And now, we are excited to announce a new member of OpenMMLab, MMFewShot!

MMFewShot is the first Pytorch-based few-shot learning toolbox that unifies extensive few-show learning algorithms into a structuralized framework. In this project, few-shot image classification and detection algorithms are all out-of-the-box with high accuracy and efficiency. Following the convention of OpenMMLab, all the models in MMFewShot are decomposed into different modules suiting our generic framework. Researchers can significantly benefit from its flexible modular design, enabling them to make fair comparisons amongst modules and replace existing modules with their creations in a plug-and-play manner.

If these are exactly what you need, don’t hesitate to check out our project on GitHub! Feel free to watch, star, and clone our projects. You are also welcome to help its growth by raising any issues and even making a PR!

https://github.com/open-mmlab/mmfewshot

Some highlights include:

• Unified Benchmark: We provide the first unified benchmark toolbox for few shot classification and detection methods.

• Modular Design: We decompose the few shot learning framework into different components and one can build a new model easily and flexibly by combining different modules.

• Strong Baselines: The toolbox provides strong baselines on few shot classification and detection methods.

Introducing MMFewShot

MMFewShot is the first few shot learning toolbox for classification and detection methods, and below is its whole framework:

MMFewShot consists of 4 parts, datasets, models, core and apis. datasets is for datasets loading and data augmentation. In this part, we support various datasets for classification and detection algorithms, and useful data augmentation transform in pipelines, and dataset wrapper for data loading. In apis, we provide high-level APIs for models training, testing, and inference, and there are evaluation tools and customized hooks for model training in core.