zeroQiaoba/ALIM

ALIM: Adjusting Label Importance Mechanism for Noisy Partial Label Learning

Correspondence to:

Mingyu Xu* (xumingyu2021@ia.ac.cn)
Zheng Lian* (lianzheng2016@ia.ac.cn)

Paper

ALIM: Adjusting Label Importance Mechanism for Noisy Partial Label Learning
Mingyu Xu *, Zheng Lian *, Lei Feng, Bin Liu, Jianhua Tao [ * Equal Contribution]

Please cite our paper if you find our work useful for your research:

@article{xu2023alim,
  title={ALIM: Adjusting Label Importance Mechanism for Noisy Partial Label Learning},
  author={Xu, Mingyu and Lian, Zheng and Feng, Lei and Liu, Bin and Tao, Jianhua},
  journal={Advances in Neural Information Processing Systems (NeurIPS)},
  year={2023}
}

Usage

Download Datasets

CIFAR-10,CIFAR-100 and CIFAR-100H can be downloaded automatically during program operation. Put the files into ./dataset/CIFAR10 and ./dataset/CIFAR100
For CUB200, PiCO provides a preprocessed copy at link and just put the files to dataset/CUB200/processed. See link for more details.

Run ALIM on CIFAR-10 and CIFAR-100:

We provide shell codes for model training. We do not tune the optimal parameters. Adjust the learning rate, epochs, and the time to start applying DALI will achieve better performance.

1.Run CIFAR10 with q=0.3 \eta=0.3:
python -u train_merge.py --seed=1 --save_root=savefinals --dataset=cifar10 --dataset_root=./dataset/CIFAR10 --partial_rate=0.3 --noise_rate=0.3 --epochs=1000 --encoder=resnet --lr=0.01 --lr_adjust=Case1 --optimizer=sgd --weight_decay=1e-3 --gpu=0 --piror_start 80 --max1

2.Run CIFAR100 with q=0.05 \eta=0.3:
python -u train_merge.py --seed=1 --save_root=savefinals --dataset=cifar100 --dataset_root=./dataset/CIFAR100 --partial_rate=0.05 --noise_rate=0.3 --epochs=1000 --encoder=resnet --lr=0.01 --lr_adjust=Case1 --optimizer=sgd --weight_decay=1e-3 --gpu=0 --piror_start 80 --max1

3.Run CIFAR100 with q=0.05 \eta=0.3 without MIXUP:
python -u train_merge.py --seed=1 --save_root=savefinals --dataset=cifar100 --dataset_root=./dataset/CIFAR100 --partial_rate=0.05 --noise_rate=0.3 --epochs=1000 --encoder=resnet --lr=0.01 --lr_adjust=Case1 --optimizer=sgd --weight_decay=1e-3 --gpu=0 --piror_start 80 --loss_weight_mixup 0 --max1

4.Run CIFAR100 with q=0.05 \eta=0.3 with SCALE NORMALIZATION:
python -u train_merge.py --seed=1 --save_root=savefinals --dataset=cifar100 --dataset_root=./dataset/CIFAR100 --partial_rate=0.05 --noise_rate=0.3 --epochs=1000 --encoder=resnet --lr=0.01 --lr_adjust=Case1 --optimizer=sgd --weight_decay=1e-3 --gpu=0 --piror_start 80 --proto_case=Case3 --max1

More examples refer to example.sh

Run ALIM on CUB-200 and CIFAR-100H:

1.Run CUB200 with q=0.05 \eta=0.2:
python -u train_merge.py --seed=1 --save_root=savefinals --dataset=cub200 --dataset_root=./dataset/CUB200 --partial_rate=0.05 --noise_rate=0.2 --epochs=200 --encoder=resnet --lr=0.01 --lr_adjust=Case1 --optimizer=sgd --weight_decay=1e-5 --gpu=0 --proto_start 30 --batch_size=256 --moco_queue=4096 --piror_start 60 --max1

2.Run CIFAR100H with q=0.5 \eta=0.2:
python -u train_merge.py --seed=1 --save_root=savefinals --dataset=cifar100H --dataset_root=./dataset/CIFAR100 --partial_rate=0.5 --noise_rate=0.2 --epochs=1000 --encoder=resnet --lr=0.01 --lr_adjust=Case1 --optimizer=sgd --weight_decay=1e-3 --gpu=0 --piror_start 80 --max1

Hyper-parameter tuning

1.Run CIFAR100 with q=0.05 \eta=0.3 with fix \lambda=0.4:
python -u train_merge.py --seed=1 --save_root=savefinals --dataset=cifar100 --dataset_root=./dataset/CIFAR100 --partial_rate=0.05 --noise_rate=0.3 --epochs=1000 --encoder=resnet --lr=0.01 --lr_adjust=Case1 --optimizer=sgd --weight_decay=1e-3 --gpu=0 --piror_start 80 --piror_auto='case3' --piror_add=0.4 --piror_max=0.4 --max1

2.Run with different e_{start}|\lambda_{mix}, please adjust hyper-parameters piror_start/loss_weight_mixup.

zeroQiaoba / ALIM