Category Traversal Module for Few-shot Learning

A PyTorch implementation of our paper "Finding Task-Relevant Features for Few-Shot Learning by Category Traversal", published at CVPR 2019, an ORAL presentation.

By Hongyang Li, David Eigen, Samuel Dodge, Matthew Zeiler, and Xiaogang Wang.

[arXiv Paper] [Poster]

[End-of-internship Presentation Slides] (40 mins)

[Short Slides] (5 mins at CVPR)

(a) describes the conventional metric-based methods and (b) depicts the proposed CTM where features are traversed across categories for acquiring better representations.

The following figure shows a detailed configuration of our proposed CTM module.

Overview

• PyTorch 0.4 or above, tested in Linux/cluster/multi/single-gpu(s).
• Datasets: tieredImagenet and miniImagenet
• A metric-based few-shot learning algorithm
• The proposed Category Traversal Module (CTM) serves as a plug-and-play unit to most existing methods, with ~2% improvement in accuracy.

Install

There are some dependencies; be sure to install the newer version to be compatible with the latest pytorch. For example:

conda install pytorch torchvision cudatoolkit=10.0 -c pytorch
conda install -c anaconda pillow pyyaml opencv requests
conda install -c conda-forge visdom

Prepare the dataset (miniImageNet for example):

sh dataset/get_tier_and_mini.sh

How to run

python main.py --yaml_file configs/demo/mini/20way_1shot.yaml

Datasets

We conduct all the experiments on tieredImagenet and miniImagenet benchmarks; to download them, please refer to DATASET.md.

Adapting CTM module to your own task

Please refer to forward_CTM method in the core/model.py file for details.

The current version contains some legacy variable names in early trial experiments; we would remove them later and make the repo cleaner.

Citation

Please cite in the following manner if you find it useful in your research:

@inproceedings{li2019ctm,
  title = {{Finding Task-Relevant Features for Few-Shot Learning by Category Traversal}},
  author = {Hongyang Li and David Eigen and Samuel Dodge and Matthew Zeiler and Xiaogang Wang},
  booktitle = {CVPR},
  year = {2019}
}