Prototypical Networks for Few-shot Learning

Reimplementation of the NIPS 2017 Paper "Prototypical Networks for Few-shot Learning" by Snell, Jake and Swersky, Kevin and Zemel, Richard

Prototypical Networks

Prototypical networks learn ametric space in which classification can be performed by computing distances to prototype representations of each class.

Installation

Requiring Python 3.6, because of dependence on typing and f-string, execute from the root of the project:

$ source scripts/setup.sh

Data

Omniglot Data Set

Omniglot is a dataset of 1623 handwritten characters collected from 50 alphabets. There are 20 examples associated with each character, where each example is drawn by a different human subject.

Data Splits

Following the Vynials splitting method as in Matching Networks for One Shot Learning, using the official split files.

Training

To train the protonet on a task (e.g. Omniglot data set, according to the configuration config/omniglot.yaml):

1. cd into the root of the project folder and
2. execute:

$ python run.py omniglot

The configuration files accepts options:

• dataset: Name of the dataset to train on (i.e., 'omniglot').
• num_epochs: Number of training epochs, default to 100.
• num_episodes: Number of episodes per epoch, default to 100.
• num_ways_train: Number of random classes per episode for training, default to 60.
• num_support_train: Number of samples per class to use as support for training, default to 5.
• num_query_train: Number of samples per class to use as query for training, default to 5.
• num_ways_val: Number of random classes per episode for validation, default to 5.
• num_support_val: Number of samples per class to use as support for validation, default to 5.
• num_query_val: Number of samples per class to use as query for validation, default to 15.
• seed: Random seed, default to 0.

Evaluation

Running the configuration file config/omniglot.yaml the Prototypica Networks scores are given here.

Disclaimer

The official implementation of the code for the paper can be found here. The purpose of this repository is purely educational, and it doesn't imply that I was involved in any way to the research conducted for the original publication.

Reference

Cite the original paper, using the official .bib citation:

@article{DBLP:journals/corr/SnellSZ17,
  author    = {Jake Snell and
               Kevin Swersky and
               Richard S. Zemel},
  title     = {Prototypical Networks for Few-shot Learning},
  journal   = {CoRR},
  volume    = {abs/1703.05175},
  year      = {2017},
  url       = {http://arxiv.org/abs/1703.05175},
  archivePrefix = {arXiv},
  eprint    = {1703.05175},
  timestamp = {Wed, 07 Jun 2017 14:41:38 +0200},
  biburl    = {http://dblp.org/rec/bib/journals/corr/SnellSZ17},
  bibsource = {dblp computer science bibliography, http://dblp.org}
}

License

This project is licensed under the MIT License

Copyright (c) 2018 Angelos Filos.