Awesome

MeTAL - Meta-Learning with Task-Adaptive Loss Function for Few-Shot Learning (ICCV2021 Oral)

Sungyong Baik, Janghoon Choi, Heewon Kim, Dohee Cho, Jaesik Min, Kyoung Mu Lee

Official PyTorch implementation of Meta-Learning with Task-Adaptive Loss Function for Few-Shot Learning (ICCV2021 Oral)

The code is based off the public code of MAML++, where their reimplementation of MAML is used as the baseline. The code also includes the implementation of ALFA.

[Paper-arXiv] [Video]

Requirements

Ubuntu 18.04

• Anaconda3
• Python==3.7.10
• PyTorch==1.4
• numpy==1.19.2

To install requirements, first download Anaconda3 and then run the following:

conda create -n metal python=3.7.10
conda activate metal
bash install.sh

Datasets

For miniIamgenet, the dataset can be downloaded from the link provided from MAML++ public code. make a directory named 'datasets' and place the downloaded miniImagnet under the 'datasets' directory.

Training

To train a model, run the following command in experiments_scripts directory

bash MeTAL.sh $GPU_ID

Evaluation

After training is finished, evaluation is performed automatically. To run an evaluation manually, run the same command

bash MeTAL.sh $GPU_ID

Results

Reference

@InProceedings{baik2021meta,
 title={Meta-Learning with Task-Adaptive Loss Function for Few-Shot Learning},
 author={Sungyong Baik, Janghoon Choi, Heewon Kim, Dohee Cho, Jaesik Min, Kyoung Mu Lee}
 booktitle = {International Conference on Computer Vision (ICCV)}, 
 year={2021}
}