[ICCV23] How Far Pre-trained Models Are from Neural Collapse on the Target Dataset Informs their Transferability

Paper Link

Setup

Pytorch == 1.11.0
TorchVision == 0.12.0

How-to

Step 1 (Optional if you are using the same checkpoints as specified in models/group1/xx models)

python finetune_group1.py -m resnet50 -d sun397

Step 2 Feature extraction for all candidate models on the SUN397 dataset. See get_data() function in get_dataloader.py file for all supported data sources.

python forward_feature_group1.py -d sun397

Step 3 Evaluate the transferability of models

python evaluate_metric_group1_cpu.py -me ncti -d sun397

Step 4 Calculate the ranking correlation

python tw_group1.py --me ncti -d sun397

Citation

If you find our work useful to your research, please cite

@inproceedings{wang2023far,
  title={How Far Pre-trained Models Are from Neural Collapse on the Target Dataset Informs their Transferability},
  author={Wang, Zijian and Luo, Yadan and Zheng, Liang and Huang, Zi and Baktashmotlagh, Mahsa},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages={5549--5558},
  year={2023}
}

Acknowledgement:

This code repository is developed based on SFDA.