Long-tail learning via logit adjustment

Aditya Krishna Menon, Sadeep Jayasumana, Ankit Singh Rawat, Himanshu Jain, Andreas Veit, Sanjiv Kumar.

---

This is the unofficial implementation of DAR-BN in the paper Long-tail learning via logit adjustment(ICLR 2021) in pytorch.

Dependency

The code is built with following libraries:

• PyTorch 1.2
• TensorboardX
• scikit-learn

Dataset

• Imbalanced CIFAR. The original data will be downloaded and converted by imbalance_cifar.py.

Training

We provide several training examples with this repo:

• To train the ERM baseline on long-tailed imbalance with ratio of 100

python cifar_train.py --gpu 0 --imb_type exp --imb_factor 0.01 --loss_type CE --train_rule None

• To train the ERM Loss along with DAR-BN on long-tailed imbalance with ratio of 100

python cifar_train.py --gpu 0 --imb_type exp --imb_factor 0.01 --loss_type CE --train_rule DAR-BN

Results

	Baseline (ERM)	Logit adjustment loss
CIFAR-10 LT	79.6	80.74
CIFAR-100 LT	36.2	41.58