I am working on a mini-library on top of fastai for meta learning. Few shot learning algorithms often require complicated gradient updates which require a lot of code. I am trying to reduce the hassle in using these algorithms.
Key features:
- Supports MAML, Meta-SGD and Reptile algorithms currently.
- Native support for all torchvision resnet models. Also allows use of pretrained imagenet weights.
- Functional versions of nn.Conv2d, nn.Batchnorm2d, nn.Linear and nn.Sequential to allow users to easily create their own models for few shot learning.
- Support for fastai’s callback system.
It is still a work in progress and any feedback is appreciated, especially from people with experience in this domain.
Link: https://github.com/Atom-101/Meta.AI
Note that it isn’t pip installable yet. In order to use it, sys.path has to be modified.