PENCIL

Third-party implement of paper 'Probabilistic End-to-end Noise Correction for Learning with Noisy Labels'

Attention! I find that using the hyper parameters in the paper can not get good result, so I try some other parameters and the result is not bad.

Requirements

pytorch >= 1.0

hyper parameters

experiment result:

Run

for example:

1. python backbone_train.py --lr 0.03
2. python pencil_train.py --alpha 0.01 --beta 0.1 --lamda 500 --lr 0.03 --percent 0.3
3. python fine_tune.py --lr 0.1 --percent 0.3