SNSCL

PyTorch Implementation of SNSCL (accepted to CVPR 2023).

1. Environment settings

Python 3.8, Pytorch 1.11, CUDA 11.1

2. Dataset

Before running the code, you should create a fold '/dataset/{}' and download datasets from following links.

   Stanford Dog: http://vision.stanford.edu/aditya86/ImageNetDogs/
   Stanford Car: https://www.kaggle.com/datasets/jessicali9530/stanford-cars-dataset
   Aircraft: https://www.robots.ox.ac.uk/~vgg/data/fgvc-aircraft/
   Cub-200-2011: https://data.caltech.edu/records/65de6-vp158

3. Training

python vanilla_w_SNSCL.py  --dataset {dog, car, aircraft, cub}
                           --loss {ce_loss, APL, Asym, GCE, Sym, label_smooth, confPenalty}
                           --batch_size 64
                           --lr 0.002
                           --noise_type sym
                           --noise_r 0.2
                           --gpu 0

Some training records can be found in Google drive.

@inproceedings{wei2023fine,
  title={Fine-grained classification with noisy labels},
  author={Wei, Qi and Feng, Lei and Sun, Haoliang and Wang, Ren and Guo, Chenhui and Yin, Yilong},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={11651--11660},
  year={2023}
}