CNLL: A Semi-Supervised Approach for Continual Learning with Noisy Labels

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

Paper

Code for Training

System Dependencies

• Python >= 3.6.1
• CUDA >= 9.0 supported GPU

Installation

Using virtual env is recommended.

$ conda create --name CNLL python=3.6

Install pytorch==1.7.0 and torchvision==0.8.1. Then, install the rest of the requirements.

$ pip install -r requirements.txt

First, generate the different tasks out of a single dataset

Users can perform task/class incremental learning in this manner. We create class-wise tasks where each task has M number of classes to deal with. Specify parameters in config yaml, episodes yaml files. Here config contains dataset description and episodes contain task information.

python main.py --log-dir [log directory path] --c [config file path] --e [episode file path] --override "|" --random_seed [seed]

Run CIFAR10 asymmetric noise rate of 40% experiment-

python main.py --log-dir ./data --c configs/cifar10_spr.yaml --e episodes/cifar10-split_epc1_asym_a.yaml --override "asymmetric_noise=True|corruption_percent=0.4";

Run CIFAR100 superclass symmetric noise rate of 40% experiment. Noise labels can be generated within 20 superclasses or randomly.

python main.py --log-dir ./data --c configs/cifar100_spr.yaml --e episodes/cifar100sup-split_epc1_a.yaml --override "superclass_noise=True|corruption_percent=0.4";

Run CNLL Algorithm for Continual Noisy Label Learning on These Tasks

Make sure the ".npy" files for different tasks are in the same data folder. Check "data_path" argument in "Train_cifar_CNLL.py". Also, please make sure noise mode and noise ratio are consistent with the task specification.

For the CIFAR10 asymmetric noise rate of 40% experiment-

python Train_cifar_CNLL.py --dataset cifar10 --noise_mode asym --r 0.4

For CIFAR100 symmetric and superclass noise rate of 40% experiment-

python Train_cifar_CNLL.py --dataset cifar100 --noise_mode sup --r 0.4	

For the CIFAR100 symmetric and random noise rate of 40% experiment-

python Train_cifar_CNLL.py --dataset cifar100 --noise_mode rand --r 0.4

Thanks! If you have any queries please send an email [email protected]. If you find the implementation useful, please cite our paper!

@InProceedings{Karim_2022_CVPR,
    author    = {Karim, Nazmul and Khalid, Umar and Esmaeili, Ashkan and Rahnavard, Nazanin},
    title     = {CNLL: A Semi-Supervised Approach for Continual Noisy Label Learning},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
    month     = {June},
    year      = {2022},
    pages     = {3878-3888}
}