This repository contains the code for the ICLR 2024 paper:
ConR: Contrastive Regularizer for Deep Imbalanced Regression
Mahsa Keramati, Lili Meng, R David Evans
ConR key insights. a) Without ConR, it
is common to have minority examples mixed with
majority examples. b) ConR adds additional loss
weight for minority, and mis-labelled examples,
resulting in better feature representations and c)
better prediction error.
ConR is complementary to conventional imbalanced learning techniques. The following code snippent shows the implementation of ConR for the task of Age estimation
def ConR(features, targets, preds, w=1, weights=1, t=0.07, e=0.01):
q = torch.nn.functional.normalize(features, dim=1)
k = torch.nn.functional.normalize(features, dim=1)
l_k = targets.flatten()[None, :]
l_q = targets
p_k = preds.flatten()[None, :]
p_q = preds
l_dist = torch.abs(l_q - l_k)
p_dist = torch.abs(p_q - p_k)
pos_i = l_dist.le(w)
neg_i = ((~ (l_dist.le(w))) * (p_dist.le(w)))
for i in range(pos_i.shape[0]):
pos_i[i][i] = 0
prod = torch.einsum("nc,kc->nk", [q, k]) / t
pos = prod * pos_i
neg = prod * neg_i
pushing_w = weights * torch.exp(l_dist * e)
neg_exp_dot = (pushing_w * (torch.exp(neg)) * neg_i).sum(1)
# For each query sample, if there is no negative pair, zero-out the loss.
no_neg_flag = (neg_i).sum(1).bool()
# Loss = sum over all samples in the batch (sum over (positive dot product/(negative dot product+positive dot product)))
denom = pos_i.sum(1)
loss = ((-torch.log(
torch.div(torch.exp(pos), (torch.exp(pos).sum(1) + neg_exp_dot).unsqueeze(-1))) * (
pos_i)).sum(1) / denom)
loss = (weights * (loss * no_neg_flag).unsqueeze(-1)).mean()
return lossPlease go into the sub-folder to run experiments for different datasets.
The code is based on Yang et al., Delving into Deep Imbalanced Regression, ICML 2021 and Ren et al.,Balanced MSE for Imbalanced Visual Regression, CVPR 2022.