search

UCSC-REAL / negative-label-smoothing

[ICML2022 Long Talk] Official Pytorch implementation of "To Smooth or Not? When Label Smoothing Meets Noisy Labels"
114 stars 10 forks source link

Question about results #2

Open MrChenFeng opened 2 years ago

MrChenFeng commented 2 years ago

Thanks for your great work! I wonder do you report the best ACC rather than the final ACC in training in Tables 1-4?

weijiaheng commented 2 years ago

Hi Chen,

That's a great question! For the final reporting of Acc in all Tables:

  1. All experiments on UCI datasets & Synthetic binary data: (Table 1, 2, 9, 10, 11) We report the accuracy of the test data in the final epoch during training;

  2. Synthetic noisy CIFAR datasets: (Table 3, 4) Since our primary purpose is to compare the potential/optimal performance of NLS/LS. To avoid reporting/comparing the optimal local performance, in an understanding view, we report the best-achieved test accuracy for Tables 3 & 4.

  3. Comparing with existing methods: (Table 5, 6) When comparing with existing methods on synthetic noisy CIFAR datasets (Table 5), Real-world noisy datasets (Table 6), we report the final accuracy of the test data after the training. Specifically, for positive label smoothing, we adopt a smooth rate 0.6; for negative label smoothing, we chose the smooth rate -6.0.