SCELoss-PyTorch
Official Repo: https://github.com/YisenWang/symmetric_cross_entropy_for_noisy_labels
Reproduce result for ICCV2019 paper "Symmetric Cross Entropy for Robust Learning with Noisy Labels"
Update
In the tensorflow version Official Repo, the model uses l2 weight decay of 0.01 on model.fc1, which will gives a better results. The code has been updated, now it should shows similar performance as in the paper.
How To Run
Arguments
- --loss: 'SCE', 'CE'
- --nr: 0.0 to 1.0 specify the nosie rate.
- --dataset_type: 'cifar10' or 'cifar100'
- --alpha: alpha for SCE
- --beta: beta for SCE
- --seed: random seed
- --version: For experiment notes
Example for 0.4 Symmetric noise rate with SCE loss
# CIFAR10
$ python3 -u train.py --loss SCE \
--dataset_type cifar10 \
--l2_reg 1e-2 \
--seed 123 \
--alpha 0.1 \
--beta 1.0 \
--version SCE0.4_CIFAR10 \
--nr 0.4
# CIFAR100
$ python3 -u train.py --lr 0.01 \
--loss SCE \
--dataset_type cifar100 \
--l2_reg 1e-2 \
--seed 123 \
--alpha 6.0 \
--beta 1.0 \
--version SCE0.4_CIFAR100 \
--nr 0.4
Results on CIFAR10
Result of best Epoch
| Loss | 0.0 | 0.2 | 0.4 | 0.6 | 0.8 |
|---|---|---|---|---|---|
| CE | 92.68 | 84.70 | 72.77 | 54.14 | 31.23 |
| SCE | 92.05 | 89.96 | 84.65 | 73.77 | 36.28 |
Results on CIFAR100
| Loss | 0.0 | 0.2 | 0.4 | 0.6 | 0.8 |
|---|---|---|---|---|---|
| CE | 73.84 | 61.70 | 42.88 | 20.47 | 4.88 |
| SCE | 73.57 | 62.31 | 46.50 | 24.00 | 12.51 |