search

shivashankarrs / classimb_fairness

5 stars 0 forks source link

Task 2 (class imbalance approaches) #2

Closed shivashankarrs closed 3 years ago

shivashankarrs commented 3 years ago

Implement/refactor class imbalance approaches, with easy to use api calls.

afshinrahimi commented 3 years ago

LDAM-DRW and Focal Loss https://github.com/kaidic/LDAM-DRW/blob/master/losses.py

Self-adjusted Dice Loss https://github.com/fursovia/self-adj-dice/blob/master/sadice/loss.py

Class-balanced loss based on effective number of samples (uses focal) https://github.com/vandit15/Class-balanced-loss-pytorch/blob/master/class_balanced_loss.py

List of many losses: https://github.com/JunMa11/SegLoss

afshinrahimi commented 3 years ago

There are two main approaches: (1) resampling and (2) reweighting/loss.

The resampling and reweighting approaches can be combined (as in ldam-drw).

shivashankarrs commented 3 years ago

Yeah right, we will require some implementations for resampling as well. Resampling with some dropouts is a reasonable way I think, has anyone done it?, or they just duplicate?

afshinrahimi commented 3 years ago

balanced sampling of imbalanced datasets easy to use library (will equate oversampling) https://github.com/ufoym/imbalanced-dataset-sampler

shivashankarrs commented 3 years ago

We have all objective based losses.