Noisy Label Detection Implementation

O2U-Net + some modifications

O2U-Net implementation source git: link

Usage Example

python main.py  --network=resnet50  --noise_rate 0.6 --remove_rate 0.8

(test)
python main.py  --network=resnet50 --noise_rate 0.6 --remove_rate 0.85 --n_epoch1 1 --n_epoch2 2 --n_epoch3 2

---

My Logs

2022/05

1. Remove Caffe2 thread-pool leak warning: link
2. 10 epoch experiment result

epoch:10 lr:0.009100 train_loss: 3.1395677614974975 test_accuarcy:55.330000 noise_accuracy:0.840500 top 0.1 noise accuracy:0.982200

3. model: ResNet50, ResNet101
4. python main.py --network=resnet101 --transforms=true (network: resnet50 / resnet101)

2022/06

1. first stage, second stage: main.py
2. third stage: curriculum.py : apply curriculum learning with masked dataset
3. mod save directory
4. Usage: python main.py --network=resnet50 --noise_rate 0.6 --remove_rate 0.8
5. Add readme flow.jpg image

2022/07

1. transform_ad.py: add more transform operations

TranslateX, TranslateY, Flip, Rotate, Posterize, AutoContrast