Pytorch implementation of NeurIPS 2021 paper Dynamic Distillation Network for Cross-Domain Few-Shot Recognition with Unlabeled Data.
Install the required python packages by:
pip install -r requirements.txt
Download the datasets from here, and keep them in the data/
directory.
Download the cross-entropy pretrained model (download from here) on mini-ImageNet in ckpt/ce_miniImageNet_resnet10
.
Or, train by running:
python main.py system=ce backbone=resnet10 data.dataset=miniImageNet_train model_name=ce_miniImageNet_resnet10
To train on mini-ImageNet and unlabeled target images, run the following command:
python main.py system=ce_distill_ema_sgd trainer.gpus=1 backbone=resnet10 \
data.val_dataset=EuroSAT_test data.test_dataset=null print_val=false \
trainer.log_every_n_steps=-1 \
unlabel_params.dataset=EuroSAT_train data.num_episodes=600 \
trainer.progress_bar_refresh_rate=0 print_val=false launcher.gpus=1\
model_name=dynamic_cdfsl_EuroSAT
Change EuroSAT
to other dataset to train on CropDisese
, ChestX
, ISIC
.
Model will be saved in ckpt/dynamic_cdfsl_EuroSAT/last.ckpt
You can download the pretrained checkpoints on the BSCD-FSL benchmark datasets from here.
For 5-way 5-shot evaluation on EuroSAT dataset:
python main.py system=few_shot data.test_dataset=EuroSAT_test ckpt=[pretrained-checkpoint]
If you find this repo useful for your research, please consider citing the paper:
@inproceedings{
islam2021dynamic,
title={Dynamic Distillation Network for Cross-Domain Few-Shot Recognition with Unlabeled Data},
author={Ashraful Islam and Chun-Fu Chen and Rameswar Panda and Leonid Karlinsky and Rogerio Feris and Richard Radke},
booktitle={Thirty-Fifth Conference on Neural Information Processing Systems},
year={2021},
url={https://openreview.net/forum?id=_4VxORHq-0g}
}