Self-Promoted Supervision for Few-Shot Transformer

Yet another Few-Shot ViT training framework.

Our code is mainly based on MetaBaseline, and SUN-F (means FEAT)/D (means DeepEMD) are based on the corresponding codebase. Sincerely thanks for their contribution.

Requirements

PyTorch (>= 1.9)
TorchVision
timm (latest)
einops
tqdm
numpy
scikit-learn
scipy
argparse
tensorboardx

Update

June 4th, 2022: we upload the meta-training phase and the meta-tuning phase of SUN-D.

June 4th, 2022: we upload the teacher training code in the meta-training phase of SUN.

June 3rd, 2022: we upload the meta-tuning phase of SUN-M.

Pretrained Checkpoints

Currently we provide SUN-M (Visformer) trained on miniImageNet (5-way 1-shot and 5-way 5-shot), see Google Drive for details.

More pretrained checkpoints coming soon.

Evaluate the Pretrained Checkpoints

Prepare data

For example, miniImageNet:

cd test_phase

Download miniImageNet dataset from miniImageNet (courtesy of Spyros Gidaris)

unzip the package to materials/mini-imagenet, then obtain materials/mini-imagenet with pickle files.

Prapare pretrained checkpoints

Download corresponding checkpoints from Google Drive and store the checkpoints in test_phase/ directory.

Evaluation

cd test_phase
python test_few_shot.py --config configs/test_1_shot.yaml --shot 1 --gpu 1 # for 1-shot
python test_few_shot.py --config configs/test_5_shot.yaml --shot 5 --gpu 1 # for 5-shot

For 1-shot, you can obtain: test epoch 1: acc=67.80 +- 0.45 (%)

For 5-shot, you can obtain: test epoch 1: acc=83.25 +- 0.28 (%)

Test accuracy may slightly vary with different pytorch/cuda versions or different hardwares

Citation

@inproceedings{dong2022self,
  title={Self-Promoted Supervision for Few-Shot Transformer},
  author={Dong, Bowen and Zhou, Pan and Yan, Shuicheng and Zuo, Wangmeng},
  booktitle={European Conference on Computer Vision (ECCV)},
  year={2022}
}

TODO

more checkpoints

vieozhu/few-shot-vit