Popular implementations in few-shot learning

Refactored by @hli2020. This repo contains:

• Prototypical Networks for Few-shot Learning, denoted as nips17_proto. Forked repo.

• Few-shot learning with graph neural networks, denoted as iclr18_gnn. Forked repo.

• Learning to compare: relation network for few-shot learning, denoted as cvpr18_relation. Forked repo.

Overview

• Supported datasets: Omniglot, Mini-ImageNet

• PyTorch 0.4.x

• Multi-gpu if necessary

• To run, see the scripts in scripts folder. Results will be logged in output

How to run it

Check the scripts folder to have a sense. Universal arguments across different methods are stored in the basic_opt.py file. The outputs are generated in the output folder after the training is launched.

The refactored documents for each method are stored in the doc folder.

TODO list

• Support tier-ImageNet

• Merge dataset processing unified within the repo (for now, there is a gnn_specific)

• Support log visualizations in Visdom and/or TensorboardX

Dependencies

• (You might need) to install opencv:

  conda install -c defaults libprotobuf protobuf
  conda install -c conda-forge opencv