Model-Agnostic Boundary-Adversarial Sampling for Test-Time Generalization in Few-Shot Learning

In ECCV 2020 (Oral)

This is the code for our paper,

• Jaekyeom Kim, Hyoungseok Kim, and Gunhee Kim. Model-Agnostic Boundary-Adversarial Sampling for Test-Time Generalization in Few-Shot Learning. In ECCV, 2020. [paper] [appx] [talk] [slides]

This is the application of our method to Dynamic Few-Shot Visual Learning without Forgetting.

Citing the paper

If you find our work or this code useful in your research, please cite

@inproceedings{kim2020_mabas,
    title={Model-Agnostic Boundary-Adversarial Sampling For Test-Time Generalization in Few-Shot Learning},
    author={Kim, Jaekyeom and Kim, Hyoungseok and Kim, Gunhee},
    booktitle={Computer Vision--ECCV 2020: 16th European Conference, Glasgow, UK, August 23--28, 2020, Proceedings, Part I 16},
    pages={599--617},
    year={2020},
    organization={Springer}
}

Environment setup

You can create a conda environment by

conda env create -f environment.yml

and activate it with

conda activate mabas-fswf

Downloading datasets

• Create a subdirectory named datasets.
• Download and decompress miniImageNet (from FSwF and MetaOptNet) into datasets/MiniImagenet/.
• Download and decompress CIFAR-FS (from MetaOptNet) into datasets/CIFAR_FS/.
• Download and decompress FC100 (from MetaOptNet) into datasets/FC100/.

Meta-training

Command	Dataset	Type
python train.py --config miniImageNet_ResNetLikeCosineClassifier	miniImageNet	Base
python train.py --config miniImageNet_ResNetLikeCosineClassifierGenWeightAttN5 --parent_exp "miniImageNet_ResNetLikeCosineClassifier"	miniImageNet	5-shot
python train.py --config miniImageNet_ResNetLikeCosineClassifierGenWeightAttN1 --parent_exp "miniImageNet_ResNetLikeCosineClassifier"	miniImageNet	1-shot
python train.py --config CIFARFS_ResNetLikeCosineClassifier	CIFAR-FS	Base
python train.py --config CIFARFS_ResNetLikeCosineClassifierGenWeightAttN5 --parent_exp "CIFARFS_ResNetLikeCosineClassifier"	CIFAR-FS	5-shot
python train.py --config CIFARFS_ResNetLikeCosineClassifierGenWeightAttN1 --parent_exp "CIFARFS_ResNetLikeCosineClassifier"	CIFAR-FS	1-shot
python train.py --config FC100_ResNetLikeCosineClassifier	FC100	Base
python train.py --config FC100_ResNetLikeCosineClassifierGenWeightAttN5 --parent_exp "FC100_ResNetLikeCosineClassifier"	FC100	5-shot
python train.py --config FC100_ResNetLikeCosineClassifierGenWeightAttN1 --parent_exp "FC100_ResNetLikeCosineClassifier"	FC100	1-shot

Test-time fine-tuning with MABAS

Command	Dataset	Type
python evaluate_finetune.py --config miniImageNet_ResNetLikeCosineClassifierGenWeightAttN5 --parent_exp "miniImageNet_ResNetLikeCosineClassifier/miniImageNet_ResNetLikeCosineClassifierGenWeightAttN5"	miniImageNet	5-shot
python evaluate_finetune.py --config miniImageNet_ResNetLikeCosineClassifierGenWeightAttN1 --parent_exp "miniImageNet_ResNetLikeCosineClassifier/miniImageNet_ResNetLikeCosineClassifierGenWeightAttN1"	miniImageNet	1-shot
python evaluate_finetune.py --config CIFARFS_ResNetLikeCosineClassifierGenWeightAttN5 --parent_exp "CIFARFS_ResNetLikeCosineClassifier/CIFARFS_ResNetLikeCosineClassifierGenWeightAttN5"	CIFAR-FS	5-shot
python evaluate_finetune.py --config CIFARFS_ResNetLikeCosineClassifierGenWeightAttN1 --parent_exp "CIFARFS_ResNetLikeCosineClassifier/CIFARFS_ResNetLikeCosineClassifierGenWeightAttN1"	CIFAR-FS	1-shot
python evaluate_finetune.py --config FC100_ResNetLikeCosineClassifierGenWeightAttN5 --parent_exp "FC100_ResNetLikeCosineClassifier/FC100_ResNetLikeCosineClassifierGenWeightAttN5"	FC100	5-shot
python evaluate_finetune.py --config FC100_ResNetLikeCosineClassifierGenWeightAttN1 --parent_exp "FC100_ResNetLikeCosineClassifier/FC100_ResNetLikeCosineClassifierGenWeightAttN1"	FC100	1-shot

Acknowledgments

This source code is based on the implementations for Dynamic Few-Shot Visual Learning without Forgetting and Meta-learning with differentiable convex optimization.