File size: 904 Bytes
72fc481
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
# ELR+
This is an official PyTorch implementation of ELR+ method proposed in [Early-Learning Regularization Prevents Memorization of Noisy Labels](https://arxiv.org/abs/2007.00151). 


## Usage
Train the network on the Symmmetric Noise CIFAR-10 dataset (noise rate = 0.8):

```

python train.py -c config_cifar10.json --percent 0.8

```
Train the network on the Asymmmetric Noise CIFAR-10 dataset (noise rate = 0.4):

```

python train.py -c config_cifar10_asym.json --percent 0.4

```

Train the network on the Asymmmetric Noise CIFAR-100 dataset (noise rate = 0.4):

```

python train.py -c config_cifar100.json --percent 0.4 --asym 1

```

The config files can be modified to adjust hyperparameters and optimization settings. 


## References
- S. Liu, J. Niles-Weed, N. Razavian and C. Fernandez-Granda "Early-Learning Regularization Prevents Memorization of Noisy Labels", 2020