Enhanced Meta Label Correction (EMLC)

Description ℹ️

This is an official PyTorch implementation of the "Enhanced Meta Label Correction for Coping with Label Corruption" paper.

• Description
• Results
• Running the Code
  • Training
• Dependencies
• Credits

Results 📈

CIFAR-10 with 1k clean samples:

Method/Noise	20%	50%	80%	90%	Asym. 40%
MLC	92.6	88.1	77.4	67.9	–
Previous SOTA	93.4	90.07	86.78	79.52	91.6
EMLC	93.53	92.63	90.95	90.71	91.82

CIFAR-100 with 1k clean samples:

Method/Noise	20%	50%	80%	90%
MLC	66.5	52.4	18.9	14.2
Previous SOTA	72.5	65.4	55.22	16.7
EMLC	73.05	68.61	60.51	54.04

Clothing1M:

Method	Accuracy
MLC	75.78
Previous SOTA	77.83
EMLC	79.35

Running The Code 🏃

Before training the models please:

1. Put the datasets in the data sub-directory.
2. For the CIFAR experiments, put the self-supervised pretrained models in the pretrained sub-directory. You may use SupContrast implementation for training the SSL models.
3. For the Clothing experiment, run the preprocessing procedure:

python CLOTHING1M/load_cloth1m_data.py --root=[path_to_clothing_dataset]

4. Go over the dependencies.

Training 🏋️

To start training on the Clothing1M dataset run:

python main_clothing.py --gpuid=a --n_gpus=b

To start training on the CIFAR datasets run:

python main_cifar.py --gpuid=a --n_gpus=b --dataset=c --corruption_type=d --corruption_level=e --ssl_path=f

The parameters and their meaning:

Parameter	Description	Type	Possible values
gpuid	The id of the primary GPU to train on	int	[0,#GPUS)
n_gpus	Number of GPUs to train on	int	[1,#GPUS]
dataset	CIFAR dataset to train on	string	Either cifar10 or cifar100
corruption_type	Type of artifical noise	string	unif for symmetric noise injection and flip for asymmetric noise injection
corruption_level	Injected Noise level	float	(0,1)
ssl_path	Path to the relevant SSL pretrained model	string	pretrained/*.pth

Dependencies 💾

Before trying to run anything please ensure that the packages below are up to date.

Library	Minimal Version
NumPy	1.23.5
PyTorch	1.13.1
torchvision	0.14.1
tensorboard	2.11.2
argparse	1.1
tqdm	4.64.0

Credits ©️

This repository is heavily based on MLC.