Learning from limited exemplars(few shot learning) is a fundamental, unsolved problem that has been laboriously explored in the machine learning community. In the current study, we develop a method to learn an unsupervised few shot learner via self-training(UFSLS) which can effectively generalize to novel but related classes.

python3.6
pytorch1.2

Omniglot

1, put omniglot dataset in Dataset
2, python ./train/train_hardtriplet.py --hard_mining True

1, Compare with other unsupervised few shot learners 2, Clustering quality with T-SNE visualization at different training iterations

Please contact jizilong@mail.bnu.edu.cn