UNICON: Combating Label Noise Through Uniform Selection and Contrastive Learning

If you like our project, please give us a star ⭐ on GitHub for the latest update.

Paper

Training Pipeline

UNICON Framework

Installation Guide

1. Create a conda environment

   conda create -n unicon 
   conda activate unicon

2. After creating a virtual environment, install the required packages

   pip install -r requirements.txt

Download the Datasets

• For adding Synthetic Noise, download these datasets

  1. CIFAR10
  2. CIFAR100
  3. Tiny-ImageNet

• For Datasets with Real-World Label Noise

  1. Clothing1M (Please contact tong.xiao.work[at]gmail[dot]com to get the download link)
  2. WebVision

UNICON Training

• Example run (CIFAR10 with 50% symmetric noise)

   python Train_cifar.py --dataset cifar10 --num_class 10 --data_path ./data/cifar10 --noise_mode 'sym' --r 0.5 

• Example run (CIFAR100 with 90% symmetric noise)

   python Train_cifar.py --dataset cifar100 --num_class 100 --data_path ./data/cifar100 --noise_mode 'sym' --r 0.9 

This will throw an error as downloaded files will not be in the proper folder. That is why they must be manually moved to the "data_path".

• Example Run (TinyImageNet with 50% symmetric noise)

   python Train_TinyImageNet.py --ratio 0.5

• Example run (Clothing1M)

  python Train_clothing1M.py --batch_size 32 --num_epochs 200   

• Example run (Webvision)

   python Train_webvision.py 

Reference

If you have any questions, do not hesitate to contact nazmul.karim170@gmail.com

Also, if you find our work useful please consider citing our work:

@InProceedings{Karim_2022_CVPR,
    author    = {Karim, Nazmul and Rizve, Mamshad Nayeem and Rahnavard, Nazanin and Mian, Ajmal and Shah, Mubarak},
    title     = {UniCon: Combating Label Noise Through Uniform Selection and Contrastive Learning},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2022},
    pages     = {9676-9686}
}