BOIL: Towards Representation Change for Few-shot Learning

This repository is the official implementation of "BOIL: Towards Representation Change for Few-shot Learning" Our implementations are relied on Torchmeta.

Requirements

We run our code in the following environment using Anaconda.

• Python >= 3.5
• Pytorch == 1.4
• torchvision == 0.5

If you use Pytorch version above 1.5 (which is the latest version at this moment) and torchvision above 0.6, you may encounter problem. In that case, you are encouraged to change to the version in our environment.

To install requirements:

pip install -r requirements.txt

Training

If you want to train 4conv network in the paper, run this command:

./run_4conv.sh

If you want to train ResNet-12 in the paper, run this command:

./run_resnet.sh

If you want to see and change the arguments of training code, run this command:

python3 main.py --help

Evaluation

To evaluate the model(s) and see the results, please refer to the analysis.ipynb

Results

All results were reproduced by our group and reported as the average and standard deviation of the accuracies over 5x1000 tasks.

The values in parenthesis are the number of shots.

1. 5-Way k-shot test accuracy (%) of 4conv network on various benchmark dataset.

Domain	General Domain		Specific Domain
Dataset	miniImageNet	tieredImageNet	CUB	Cars
MAML(1)
BOIL(1)
MAML(5)
BOIL(5)

2. 5-Way k-shot test accuracy (%) of 4conv network on cross-domain adaptation.

Adaptation	General to general		General to Specific
Meta-train	tieredImageNet	miniImageNet	miniImageNet	miniImageNet
Meta-test	miniImageNet	tieredImageNet	CUB	Cars
MAML(1)
BOIL(1)
MAML(5)
BOIL(5)

---

Adaptation	Specific to general		Specific to Specific
Meta-train	CUB	CUB	Cars	CUB
Meta-test	miniImageNet	tieredImageNet	CUB	Cars
MAML(1)
BOIL(1)
MAML(5)
BOIL(5)

3. 5-Way 5-shot test accuracy (%) of ResNet-12.

Meta-train	miniImageNet			CUB
Meta-test	miniImageNet	tieredImageNet	CUB	CUB	miniImageNet	Cars
MAML w/ lsc
MAML w/o lsc
BOIL w/ lsc
BOIL w/o lsc