This repository contains the code associated to the self-service MetaDL project, based on the MetaDL competition framework. One can submit a dataset in a tfrecords format and obtains the performance of the AAAI 2021 MetaDL competition's winning solution: MetaDelta.
This is the official repository of the Meta-Learning workshop co-hosted competition for AAAI 2021.
The competition focus on few-shot learning for image classification. This is an online competition, i.e. you need to provide your submission as raw Python code that will be ran on the CodaLab platform. The code is designed to be a module and to be flexible and allows participants to any type of meta-learning algorithms.
You can find more informations on the ChaLearn website.
- [1] - E. Triantafillou Meta-Dataset: A Dataset of Datasets for Learning to Learn from Few Examples -- 2019
- [2] - J. Snell et al. Prototypical Networks for Few-shot Learning -- 2017
- [3] - C. Finn et al. Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks -- 2017
- [4] - Lake, B. M., Salakhutdinov, R., and Tenenbaum, J. B. (2015). Human-level concept learning through probabilistic program induction. Science, 350(6266), 1332-1338.
This module reuses some parts of the recent publication code E. Triantafillou et al. Meta-Dataset: GitHub repository regarging the data generation pipeline. Also the methods in the starting_kit/tutorial.ipynb such as plot_episode(), plot_batch(), iterate_dataset() have been taken from their introduction notebook.