This repository contains the code for the paper:
RankSim: Ranking Similarity Regularization for Deep Imbalanced Regression
Yu Gong, Greg Mori, Fred Tung
ICML 2022
[arXiv] [Paper] [Poster]
RankSim (ranking similarity) regularizer encodes an inductive bias that samples that are closer in label space should also be closer in feature space.
RankSim regularizer is straightforward to implement, can be computed in closed form, and complementary to conventional imbalanced learning techniques. It has two hyperparameters --- interpolation strength lambda and balancing weight gamma.
def batchwise_ranking_regularizer(features, targets, interp_strength_lambda):
loss = 0
# Reduce ties and boost relative representation of infrequent labels by computing the
# regularizer over a subset of the batch in which each label appears at most once
batch_unique_targets = torch.unique(targets)
if len(batch_unique_targets) < len(targets):
sampled_indices = []
for target in batch_unique_targets:
sampled_indices.append(random.choice((targets == target).nonzero()[:,0]).item())
x = features[sampled_indices]
y = targets[sampled_indices]
else:
x = features
y = targets
# Compute feature similarities for ranking
xxt = torch.matmul(F.normalize(x.view(x.size(0),-1)), F.normalize(x.view(x.size(0),-1)).permute(1,0))
# Compute the loss of ranking similarity
for i in range(len(y)):
label_ranks = rank_normalised(-torch.abs(y[i] - y).transpose(0,1))
feature_ranks = TrueRanker.apply(xxt[i].unsqueeze(dim=0), interp_strength_lambda) # differentiable ranking operation, defined in ranking.py
loss += F.mse_loss(feature_ranks, label_ranks)
return lossPlease go into the sub-folder to run experiments for different datasets.
If you find RankSim helpful, please cite our work:
@InProceedings{gong2022ranksim,
title={{R}ank{S}im: Ranking Similarity Regularization for Deep Imbalanced Regression},
author={Gong, Yu and Mori, Greg and Tung, Frederick},
booktitle={International Conference on Machine Learning (ICML)},
year={2022}
The code is based on Yang et al., Delving into Deep Imbalanced Regression, ICML 2021.