A Topological Filter for Learning with Label Noise (NeurIPS 2020, Paper)

Requirements

• PyTorch 0.4.1 (have not tested on other versions)
• Python 3.6 (for the purpose of compiling C++ code. Other 3.x versions should also work.)
• scipy 1.1.0 (this is due to the computation of distribution mode)
• termcolor, etc (which can be easily installed with pip)

Usage

• Compile the C++ code for computing the connected components. In folder ref, run ./compile_pers_lib.sh (by default it requires Python 3.6. If you are using other Python versions, modify the command inside compile_pers_lib.sh).
• Run train.py with the commands like below:

python train.py --every 5 --start_clean 30 --k_cc 4 --k_outlier 32 --seed 77 --type uniform --noise 0.4 --patience 65 --gpus 0 --dataset cifar10 --zeta 0.5

• For point cloud dataset, run the command with pc argument:

python train.py --gpus 2 --every 5 --start_clean 10 --k_outlier 30 --k_cc 100 --noise 0.8 --type uniform --patience 60 --seed 77 --dataset pc --net pc --milestone 35 --zeta 2

Here the major parameters are:

• every: the frequency of data collection.
• start_clean: when to start data collection.
• k_cc: the parameter for computing the KNN graph when finding the largest connected component.
• k_outlier: the parameter for computing the KNN graph when applying zeta filtering.
• seed: the random seed.
• type: the noise type. Options include uniform and asym.
• noise: the noise level.
• patience: this is a trick to save training time. If we observe no obvious improvement of validation accuracy for a consecutive number of N epochs, we stop the training.
• gpus: run on which GPU.
• dataset: which dataset to use. Options include cifar10, cifar100 and pc. For the pc dataset, it can be downloaded from https://github.com/charlesq34/pointnet
• zeta: the parameter for zeta filtering. Note that, when setting zeta to be > 1.0, we will use majority voting to remove the outliers. This sometimes achieves better performance.

Practical tips: For the extrmely noisy scenarios (noise level >= 0.8), we observe setting a larger k_cc is better.

Our code will be further improved to make it cleaner and easier to use.

Reference:

@inproceedings{wu2020topological,
  title={A Topological Filter for Learning with Label Noise},
  author={Wu, Pengxiang and Zheng, Songzhu and Goswami, Mayank and Metaxas, Dimitris and Chen, Chao},
  booktitle={Advances in Neural Information Processing Systems},
  year={2020}
}

Related Works:

• Error-Bounded Correction of Noisy Labels. In ICML, 2020. [Paper][Code]
• Learning with Feature Dependent Label Noise: A Progressive Approach. In ICLR, 2021. [Paper][Code]