dianas1108 / proto-net

proto-net

This repository contains our implementation of the Prototypical Network for both few-shot and zero-shot learning (Snell et al., 2017), as part of the project for the MLMI 4 Advanced Machine Learning module at the University of Cambridge.

Specifically,

• few_shot_omniglot.ipynb is for few-shot learning on the Omniglot dataset
• zero_shot.ipynb is for zero-shot learning on the Caltech-UCSD Birds (CUB) 200-2011 dataset

To replicate results in the original paper, we have also performed few-shot learning on the mini-ImageNet dataset. The code is largely similar to that for Omniglot (except from the dataset) and is omitted here.

All data used is downloadable online:

• Omniglot dataset (and split): from the original implementation
• CUB dataset: from https://www.vision.caltech.edu/datasets/cub_200_2011/
• CUB - extracted image feature by GoogleNet: from https://github.com/reedscot/cvpr2016

Also, the mini-ImageNet dataset in the used split is downloaded from https://lyy.mpi-inf.mpg.de/mtl/download/Lmzjm9tX.html.