This is the official pytorch implementation of CVPR 2024 paper ["Simple Semantic-Aided Few-Shot Learning"] [paper] [ArXiv]. If you use any content of this repo for your work, please cite the following bib entry:

@InProceedings{Zhang_2024_CVPR,
    author    = {Zhang, Hai and Xu, Junzhe and Jiang, Shanlin and He, Zhenan},
    title     = {Simple Semantic-Aided Few-Shot Learning},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2024},
    pages     = {28588-28597}
}

We propose a semantic evolution approach for automatic collection of rich semantics, and with high-quality semantics the model achieves excellent results even with the simplest network structure.

Semantic Evolution

Overall framework

The following prerequisites are required to run the scripts:

• PyTorch and torchvision

• Dataset: Please download the dataset from Google Cloud and modify the corresponding dataset path in the training file.

• Pre-trained weights: Please download checkpoints from Google Cloud, then put them into ./checkpoints directory.

• Install packages

pip instal -r requirements.txt

Please excute following commands in order.

python compute_center.py
    --dataset MiniImageNet/TiredImageNet/CIFAR-FS/FC100
    --backbone resnet/swin

python method/train.py 
    --max-epoch 50 
    --mode clip 
    --semantic-size 512 
    --text_type gpt 
    --shot 1  
    --step-size 40 
    --test-batch 600 
    --batch-size 128 
    --num-workers 8 
    --dataset MiniImageNet/TiredImageNet/CIFAR-FS/FC100
    --lr 1e-4 
    --backbone resnet/swin

python method/test.py 
    --dataset MiniImageNet/TiredImageNet/CIFAR-FS/FC100 
    --shot 1 
    --backbone resnet/swin

We thank the following repos providing helpful components/functions in our work.

• ProtoNet

• FEAT

• FewTURE

• CLIP

• BERT

• ChatGPT