Embedding Adaptation is Still Needed for Few-Shot Learning

Code Release for "Embedding Adaptation is Still Needed for Few-Shot Learning"

This code provides:

Re-implementation of the ATG algorithm in examples/atg.py.
Loaders for the dataset splits introduced in the paper.
Demonstration code for training the algorithms, borrowed from learn2learn.

Resources

Citation

Please cite this work as follows:

"Embedding Adaptation is Still Needed for Few-Shot Learning", Sébastien M. R. Arnold and Fei Sha

or with the following BibTex entry:

@article{arnold2021embedding,
    title={Embedding Adaptation is Still Needed for Few-Shot Learning},
    author={Sebastien M. R. Arnold, Fei Sha},
    year={2021},
    archivePrefix={arXiv},
    primaryClass={cs.LG}
}

Usage

Dependencies include the following Python packages:

PyTorch>=1.3.0
torchvision>=0.5.0
scikit-learn>=0.19.2
tqdm>=4.48.2
learn2learn on the master branch

Running ATG

A standalone re-implementation of ATG is provided in examples/atg.py. To run it on a synthetic dataset:

python examples/atg.py

Training on ATG Partitions

python examples/train.py --algorithm='protonet' --dataset='mini-imagenet' --taskset='original'

where

taskset takes values easy, medium-easy, medium-hard, hard or randomX where X is the seed to reproduce random splits.
dataset takes values mini-imagenet, tiered-imagenet, emnist, lfw10, cifar100.
algorithm takes values protonet, maml, anil.