Code files for Multiclass Learning from Noisy Labels for Non-decomposable Performance Measures (AISTATS 2024).
Our implementations of NCFW and NCBS are based on implementations of FW and BS (link to code). Different variants of NCLR were implemented based on Patrini et al. (2017) (link to code). See Section 6 of our paper for details.
Command lines for real data experiments
-
To run BS with Micro F1 loss, use
python3 bisection-expt-noisy.py 0 -
To run NCBS with Micro F1 loss, use
python3 bisection-expt-noisy.py 1 -
To run FW with H-mean loss, use
python3 fwunc-expt-noisy.py 0 h -
To run NCFW with H-mean loss, use
python3 fwunc-expt-noisy.py 1 h -
To run FW with G-mean loss, use
python3 fwunc-expt-noisy.py 0 g -
To run NCFW with G-mean loss, use
python3 fwunc-expt-noisy.py 1 g -
To run FW with Q-mean loss, use
python3 fwunc-expt-noisy.py 0 q -
To run NCFW with Q-mean loss, use
python3 fwunc-expt-noisy.py 1 q -
To run different variants of NCLR, use
python3 noise-corrected-expt.py
Instructions for synthetic data experiments
calc_Bayes_simu.ipynb calculates Bayes optimal error.
run_clean_simu.py runs algorithms on clean data sets.
run_noisy_simu.py runs algorithms on noisy data sets (for five noise matrices).
show_simu.ipynb plots the results.