Can Less be More? When Increasing-to-Balancing Label Noise Rates Considered Beneficial

This is the source code for the NeurIPS 2021 paper Can Less be More? When Increasing-to-Balancing Label Noise Rates Considered Beneficial.

Data

The pre-processed data including image features and clean labels is available at https://www.dropbox.com/s/ztkj5o3xenf2nx5/data.tar.gz?dl=0.

To reproduce the experiments, simply run

python3 CanLessBeMore.py --min_noise 0.2 --max_noise 0.4

You can set any noise rates as you want by changing the parameters in min_noise and max_noise