Unofficial implementation of "Max-margin Class Imbalanced Learning with Gaussian Affinity" by TensorFlow, Keras.
Munawar Hayat, Salman Khan, Waqas Zamir, Jianbing Shen, Ling Shao. Max-margin Class Imbalanced Learning with Gaussian Affinity. 2019. https://arxiv.org/abs/1901.07711
Use "Clusterling Affinity" Layer:
from affnity_loss import *
x = ClusteringAffinity(10, 1, 10.0)(some_input) # n_classes, n_centroids, sigmaBe sure that the output dimension is one more than the number of classes. This is to pass the diversity regularizer to the loss function. Use "affinity_loss" loss function on compiling.
model.compile("adam", affinity_loss(0.75), [acc]) # lambdaMNIST, lambda=0.75, sigma=10. Evaluate on macro f1-score.
| # samples per class on test data | Softmax | Affinity m=1 | Affinity m=5 |
|---|---|---|---|
| 500 | 99.28% | 99.39% | 99.33% |
| 200 | 99.03% | 99.20% | 99.12% |
| 100 | 98.79% | 98.97% | 98.75% |
| 50 | 98.20% | 98.54% | 98.65% |
| 20 | 98.56% | 98.36% | 98.85% |
| 10 | 97.83% | 98.27% | 98.85% |
