README.md

Examples

Noisy from Human

Noise Type	aggre	rand1	rand2	rand3	worst	noisy100
Noise Rate(%)	9.03	17.23	18.12	17.64	40.21	40.2

Type noisy100 belongs to cifar100, and other types belong to cifar10.

python main.py --dataset cifar10 --human --noise_type aggre

python main.py --dataset cifar100 --human --noise_type noisy100

Noisy from Synthesis

symmetric noise:

• sym to generate symmetric noise, otherwise asymmetric noise.
• noise_rate is the noise rate of the dataset.

python main.py --dataset cifar10 --sym --noise_rate 0.5

python main.py --dataset cifar100 --sym --noise_rate 0.5

asymmetric noise:

• cifar10: the noisy samples are flipped in some classes i.e. automobile < - truck, bird -> airplane, cat <-> dog, deer -> horse
• cifar100: the noisy samples are flipped in same superclass.

python main.py --dataset cifar10 --noise_rate 0.5

python main.py --dataset cifar100 --noise_rate 0.5

Others

• Under the same resnet structure, the resnet provided by torchvision will reduce performance compared to the custom resnet. (Why?)
• Issue is welcomed.

Reference

1. cifar-10-100n
2. SCELoss-Reproduce