few-shot-GNN-RN-lrp

This repo implements two few-shot learning models: graph neural net(GNN) and RelationNet. The code is based on CrossDomainFewShot.

We improve the two models using LRP explanations. The pre-trained models are available here.

Prepare Dataset

The datasets used in this repo include cars, cub, miniImagenet, places, plantae.

These datasets can be downloaded by running

./filelists/process.py

It will process each dataset into three .json files.

Test pre-trained models

Please specify the params in test.py and test.

--name: str the path to the pre-trained model

--testset: str the testset name

--transductive: bool whether use transductive inference

--n_shot: int the number of support images per class

To get LRP explanations

Please refer the lrpmodel.py. There are two functions to generate the explanations of GNN and RN.

To train the explanation-guided models

Please specify the params.method as relationnetlrp or gnnnetlrp to train the explanation-guided models.

For details of the models, please refer to RelationNetLRP in ./methods/relationnet.py and GNNNetLRP in ./methods/gnnnet.py.

The same experiments using CAN models are here

Acknowledgement

Many thanks to the developers of innvestigate and the important input from Phillip Seegerer on the LRP Pytorch implementation.