ICCV'23: Concept-wise Fine-tuning Matters in Preventing Negative Transfer (Pytorch implementation).

The requiring environment is as bellow:

• Python 3.8+
• PyTorch >= 1.10
• ...

• Please download and use the datasets from their official websites under their licenses.
• We follow Transfer-Learning-Library to split the dataset.

• For supervisd pre-trained models

    # CUB {85.14% on NVIDIA RTX A6000 GPU with seed 0}
    python training.py {CUB dataset dir} -d CUB200 -a resnet50 -sr 100 --seed 0 --log {log dir} --lr 0.01 --epochs 40 --confusing 199 

    # Car
    python training.py {Car dataset dir} -d StanfordCars -a resnet50 -sr 100 --seed 0 --log {log dir} --lr 0.01 --epochs 20  --cos

• For unsupervised pre-trained: please follow the instructions in Transfer-Learning-Library to convert the format of checkpoints.

If you find this repository useful in your research, please consider citing the following paper:

@InProceedings{Yang_2023_ICCV,
    author    = {Yang, Yunqiao and Huang, Long-Kai and Wei, Ying},
    title     = {Concept-wise Fine-tuning Matters in Preventing Negative Transfer},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
    month     = {October},
    year      = {2023},
    pages     = {18753-18763}
}

Thank the Pytorch implementation of transfer learning methods in Transfer-Learning-Library.

Thank the implementation of DANet

Contact: Yunqiao Yang (hustyyq [at] gmail [dot] com)