PLReMix: Combating Noisy Labels with Pseudo-Label Relaxed Contrastive Representation Learning

Xiaoyu Liu, Beitong Zhou, Cheng Cheng

The PyTorch implementation code of the paper, PLReMix: Combating Noisy Labels with Pseudo-Label Relaxed Contrastive Representation Learning.

Abstract Recently, the usage of Contrastive Representation Learning (CRL) as a pre-training technique improves the performance of learning with noisy labels (LNL) methods. However, instead of pre-training, when trivially combining CRL loss with LNL methods as an end-to-end framework, the empirical experiments show severe degeneration of the performance. We verify through experiments that this issue is caused by optimization conflicts of losses and propose an end-to-end PLReMix framework by introducing a Pseudo-Label Relaxed (PLR) contrastive loss. This PLR loss constructs a reliable negative set of each sample by filtering out its inappropriate negative pairs, alleviating the loss conflicts by trivially combining these losses. The proposed PLR loss is pluggable and we have integrated it into other LNL methods, observing their improved performance. Furthermore, a two-dimensional Gaussian Mixture Model is adopted to distinguish clean and noisy samples by leveraging semantic information and model outputs simultaneously. Experiments on multiple benchmark datasets demonstrate the effectiveness of the proposed method. Codes will be available.

Installation

git clone https://github.com/lxysl/PLReMix.git
cd PLReMix
# Please install PyTorch using the official installation instructions (https://pytorch.org/get-started/locally/).
pip install -r requirements.txt

Training

We enable wandb logging by default. If you want to disable it, use the --no_wandb flag. You can also use the --offline flag to enable the offline logging mode.

We offer the resume training feature if you enable wandb logging. If you want to resume training from a specific run, use --resume_id <run_id>. The run_id can be found in the wandb dashboard.

To train on the CIFAR-10 dataset, run the following command:

python train.py --r 0.2 --lambda_u 0
python train.py --r 0.5 --lambda_u 25
python train.py --r 0.8 --lambda_u 25
python train.py --r 0.9 --lambda_u 50
python train.py --r 0.4 --noise_mode asym --lambda_u 0

To train on the CIFAR-100 dataset, run the following command:

python train.py --dataset cifar100 --data_path ./data/cifar100/ --r 0.2 --lambda_u 0
python train.py --dataset cifar100 --data_path ./data/cifar100/ --r 0.5 --lambda_u 150
python train.py --dataset cifar100 --data_path ./data/cifar100/ --r 0.8 --lambda_u 150
python train.py --dataset cifar100 --data_path ./data/cifar100/ --r 0.9 --lambda_u 150
python train.py --dataset cifar100 --data_path ./data/cifar100/ --r 0.4 --noise_mode asym --lambda_u 0

To train on the TinyImageNet dataset, run the following command:

cd data && bash prepare_tiny_imagenet.sh && cd ..
python train.py --dataset tiny_imagenet --data_path ./data/tiny-imagenet-200 --r 0 --lambda_u 0
python train.py --dataset tiny_imagenet --data_path ./data/tiny-imagenet-200 --r 0.2 --lambda_u 30
python train.py --dataset tiny_imagenet --data_path ./data/tiny-imagenet-200 --r 0.5 --lambda_u 200
python train.py --dataset tiny_imagenet --data_path ./data/tiny-imagenet-200 --r 0.8 --lambda_u 300
python train.py --dataset tiny_imagenet --data_path ./data/tiny-imagenet-200 --r 0.45 --noise_mode asym --lambda_u 0

To train on the Clothing1M dataset, prepare the dataset and run the following command:

Clothing1M dataset file structure (You need to download the dataset from the corresponding website.)

.
├── category_names_chn.txt
├── category_names_eng.txt
├── clean_label_kv.txt
├── clean_test_key_list.txt
├── clean_train_key_list.txt
├── clean_val_key_list.txt
├── images
│   ├── 0
│   ├── 1
│   ├── 2
│   ├── 3
│   ├── 4
│   ├── 5
│   ├── 6
│   ├── 7
│   ├── 8
│   └── 9
├── noisy_label_kv.txt
├── noisy_train_key_list.txt
├── README.md
└── venn.png

python train_clothing1m.py --dataset clothing1m --data_path PATH_TO_Clothing1M/ --pretrain

To train on the WebVision dataset, prepare the dataset and run the following command:

WebVision dataset file structure (You need to download the dataset from the corresponding website and rearrange the imagenet folder.)

.
├── flickr
├── google
├── imagenet
│   ├── ILSVRC2012_devkit_t12
│   └── val
│       ├── n01440764
│       ├── n01728920
│       ├── ...
├── info
│   ├── queries_flickr.txt
│   ├── queries_google.txt
│   ├── queries_synsets_map.txt
│   ├── synsets.txt
│   ├── test_filelist.txt
│   ├── train_filelist_flickr.txt
│   ├── train_filelist_google.txt
│   ├── train_meta_list_all.txt
│   ├── train_meta_list_flickr.txt
│   ├── train_meta_list_google.txt
│   └── val_filelist.txt
├── test_images_256
└── val_images_256

# decompress ILSVRC2012_img_val.tar and ILSVRC2012_devkit_t12.tar.gz to imagenet/val and imagenet/ILSVRC2012_devkit_t12
cp ./data/rearrange.py PATH_TO_WEBVISION/imagenet/  # copy rearrange.py to the imagenet folder
cd PATH_TO_WEBVISION/imagenet/ && python rearrange.py  # rearrange imagenet/val folder
cd PATH_TO_PLReMix
python train_webvision.py --dataset webvision --data_path PATH_TO_WebVision/ --mcrop

---

Citation

If you have any questions, do not hesitate to contact [email protected]

Also, if you find our work useful please consider citing our work:

@misc{liu2024plremix,
    title={PLReMix: Combating Noisy Labels with Pseudo-Label Relaxed Contrastive Representation Learning},
    author={Xiaoyu Liu and Beitong Zhou and Cheng Cheng},
    year={2024},
    eprint={2402.17589},
    archivePrefix={arXiv},
    primaryClass={cs.CV}
}

Acknowledgement

• DivideMix: The algorithm that our framework is based on.
• ScanMix: The codebase that we used as a reference for our implementation.
• SupContrast: The codebase that our PLR loss is based on.
• FlatCLR: The codebase that the Flat version of our PLR loss is based on.
• MoPro: Inspiration for using the prototypes.