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