search

YyzHarry / imbalanced-regression

[ICML 2021, Long Talk] Delving into Deep Imbalanced Regression
http://dir.csail.mit.edu
MIT License
819 stars 129 forks source link

Question for Algorithm 1 #2

Closed Nate1874 closed 3 years ago

Nate1874 commented 3 years ago

Hi, I have a question for the pseudo code in Supplementary A section.

For the Algorithm 1, the LDS is only used to compute the weights for loss inverse re-weighting scheme. Why not use the smoothed labels to train the models if LDS captures the real imbalance that affects regression problems?

In addition, could you provide code for computing the effective label density distribution?

Thank you.

YyzHarry commented 3 years ago

Hi - thank you for your interests!

For LDS, we are estimating an effective label density distribution, so it is the label density (distribution) rather than label itself that is smoothed. As you noted, the effective density estimated by LDS is independent of which method you use -- a straightforward adaptation can be the re-weighting scheme; we also apply the estimated density to some classic methods and show improvements (see Sec. 4 baselines, "Synthetic samples" for example).

The code will be released very soon (in coming weeks). Please stay tuned!

YyzHarry commented 3 years ago

Hi @Nate1874 - the code has been released. Let me know if you have any questions!