Meta-learning has recently yielded state-of-the-art results in few-shot learning. However, current algorithm implementations are deeply tied to the datasets they were developed on.
As a consequence researchers often reimplement algorithms from scratch to compare their results.
To improve this situation Tobias Würfl teamed up with Tristan Deleu, Mandana Samiei, Joseph Paul Cohen and Yoshua Bengio from the Montreal Institute for Learning Algorithms (MILA) and created Torchmeta which provides extensions for PyTorch to simplify the development of meta-learning algorithms.
This library was now featured at the PyTorch Developer Conference in San Francisco as winner of the best in show award of the Global PyTorch Summer Hackathon.

Torchmeta is available as open source project from https://github.com/tristandeleu/pytorch-meta.