IM-DCL

The code of paper "Enhancing Information Maximization with Distance-Aware Contrastive Learning for Source-Free Cross-Domain Few-Shot Learning"

1. Setup

conda creat --name im-dcl python=3.9

conda activate im-dcl

conda install pytorch torchvision -c pytorch

conda install pandas

pip install numpy

pip install argparse

pip install math

pip install os

pip install sklearn

pip install scipy

pip install PIL

pip install abc

2. Code clone

git clone https://github.com/xuhuali-mxj/IM-DCL.git

cd IM-DCL

3. Dataset

For the 4 datasets CropDiseases, EuroSAT, ISIC, and ChestX, we refer to the BS-CDFSL repo.

4. Run IM-DCL

Based on ResNet

Our method aims at improving the performance of pretrained source model on the target FSL task. We introduce the information maximization, and propose a distance-aware contrastive learning, helping the pretrained source model to learn the decision boundary.

Please set your data address in configs.py.

We also provide the pretrained source model in mini_models/checkpoints/ResNet10_ce_aug/, We use ResNet10_ce_1200.tar to evaluate our IM-DCL.

We start from run.sh. Taking 5-way 1-shot as an example, the code runing process can be done as,

python ./adapt_da.py --model ResNet10 --train_aug --use_saved --dtarget CropDisease --n_shot 1

Based on ViT

Coming soon.

5. Acknowledge

Our code is built upon the implementation of FTEM_BSR_CDFSL and SHOT. Thanks for their work.