This code implements the paper: Long-tail Learning via Logit Adjustment : Aditya Krishna Menon, Sadeep Jayasumana, Ankit Singh Rawat, Himanshu Jain, Andreas Veit, Sanjiv Kumar. ICLR 2021.
# To produce baseline (ERM) results:
python main.py --dataset cifar10-lt
# To produce posthoc logit-adjustment results:
python main.py --dataset cifar10-lt --logit_adj_post 1
# To produce logit-adjustment loss results:
python main.py --dataset cifar10-lt --logit_adj_train 0
# To monitor the training progress using Tensorboard:
tensorboard --logdir logs
Replace cifar10-lt above with cifar100-lt to obtain results for the CIFAR-100 long-tail dataset.
| Baseline | Post-hoc logit adjustment | Logit-adjusted loss | |
|---|---|---|---|
| CIFAR10LT | 0.7127 | 0.7816 | 0.7857 |
| CIFAR100LT | 0.3985 | 0.4404 | 0.4402 |