This repository contains the code for the paper "Meta Evidential Transformer for Few-Shot Open-Set Recognition".

Dataset Information:

All datasets are publicly available. Please put the dataset as follow:

1. data: This contains 4 folders imagenet, tieredimagenet, cifar100, caltech. Please download data and put in corresponding folder
2. class_info: Our approach has an unique way to split the data. Similar to data, it contains four folders each for dataset. It contains class index of openset and closeset classes
3. saves: Contains initialization folder with four folders within each for each dataset. Those folders hold pretrained weights.

Code Running Information:

To train the model run main.py with following parameters:

--open_loss -> boolen telling whether to use explicit open set loss during training.
--open_loss_coeff -> weights given for openset loss.
--loss_type-> edl_loss or ce_loss used to train the model.
--dataset -> dataset type use to train the model.
--shot -> number of support set samples per task [1, 5].
--query -> Number of samples per class in query set [15].

For other commands directly look at main.py

After training, the trained model will be stored under checkpoints. Also, all training losses, and validation accuracies are sotred in the same folder.