Recognition-by-Components for Interpretable Few-Shot Learning

Official PyTorch implementation of the paper Recognition-by-Components for Interpretable Few-Shot Learning.

Requirements

Python 3.8, Pytorch 1.7.0, timm 0.3.2

Datasets

Please follow mini-imagenet-tools to obtain the miniImageNet dataset and put it in ./datasets/mini/.

Please follow tiered-imagenet-tools to obtain the tieredImageNet dataset and put it in ./datasets/tiered/.

Please follow download_cifar_fs.sh to obtain the CIFAR-FS dataset and put it in ./datasets/cifarfs/.

Please follow download_fc100.sh to obtain the FC100 dataset and put it in ./datasets/fc100/.

Pretraining

Please follow https://github.com/mrkshllr/FewTURE/tree/main to pretrain the backbone ViT-small and put it in ./initialization/miniimagenet.

Training and inference

Please see ./run.sh.

Quick start

• Please refer to https://github.com/mrkshllr/FewTURE/tree/main to download the miniImageNet dataset and the checkpoint of the corresponding pretrained ViT-small model.

• Put them in the corresponding folders.

• Run ./run.sh in bash shell.