This multi-prototype method is based off PANet https://github.com/kaixin96/PANet
After git clone, create a folder named pretrained_model and put https://download.pytorch.org/models/vgg16-397923af.pth under it
Download the dataset from https://drive.google.com/file/d/1pY4uFUxXVUModA0AQqMEwc2Qmj-DZcAa/view?usp=sharing and place it in the folder
Follow few-shot.ipynb to setup your training in google colab.
Snapshot for Set 2 1way1shot with vgg backbone + multi-prototype + k-means adjustment + distance loss => 0.504 https://drive.google.com/file/d/18rnxDxccjyaJ0exNaQmLBsD-CZ_NDdJI/view?usp=sharing
Snapshot for Set2 1way1shot with vgg backbone + multi-prototype + k-means adjustment => 0.503 https://drive.google.com/file/d/16IZtq3a_Lk_jref9G3hnQWnar22iqLlH/view?usp=sharing
Original paper's snapshot I was only able to achieve 0.477 on colab after training
Other things that have tried which didn't help
- Using saliency classifier to assist
- Training saliency on a dominant dataset then assist
- multi step k-means
- Few-shot semantic segmentation paper https://github.com/xiaomengyc/Few-Shot-Semantic-Segmentation-Papers
- Meta-learning tutorial https://lilianweng.github.io/lil-log/2018/11/30/meta-learning.html