This is the official repository for the Meta-Album benchmark for few-shot learning and meta-learning which has also been used for the NeurIPS MetaDL Competition 2021. The benchmark consists of 15 image datasets coming from 5 different domains (ecology, bio-medicine, manufacturing, remote sensing, and optical character recognition). Each domain has 3 datasets (2 publicly available, and 1 hidden on the Codabench platform for external testing). In this repository, you can find information about the 10 public datasetsis, utilities for creating your own datasets (data format, fact sheets, and data generators), and the code and instructions for performing few-shot learning experiments.
Here, you can find information to contribute to the Meta-Album Benchmark.
Here, you can find information about the 10 public datasets used in Meta-Album.
3. Data Format
This page covers the uniform data format for all datasets
4. Factsheets
This page contains the code used to generate fact sheets (PDFs containing information about a given dataset)
This page contains utility code for generating a dataset from tabular data.
On this page, you can find the code that we used for all few-shot learning experiments in the paper.
The following people have enabled the creation of the Meta-Album benchmark: