Gradient-Aware Logit Adjustment Loss for Long-tailed Classifier

This is the source code for our paper: Gradient-Aware Logit Adjustment Loss for Long-tailed Classifier based on Pytorch.

Installation

Requirements

• Python 3.8.10
• torchvision 0.12.0
• Pytorch 1.11.0
• yacs 0.1.8
• ...

More details can be seen in requirements.txt

Install GALA

git clone https://github.com/lt-project-repository/lt-project.git
cd lt-project
pip install -r requirements.txt

Note that the torch version should be compatible with your cuda version. In the configuration file (ending with '.yaml'), norm_logits is equivalent to Prediction Re-balancing Strategy, and margin_cls means GALA Loss.

Dataset Preparation

• CIFAR-10 & CIFAR-100
• ImageNet
• iNaturalist 2018
• Places

Change the dataset path in main.py accordingly.

Get Started

CIFAR100

Cross Entropy(ce)

python main.py --cfg config/CIFAR100_LT/ce_imba100.yaml

Cross Entropy + Prediction Re-balancing Strategy

python main.py --cfg config/CIFAR100_LT/ce_pn_imba100.yaml

GALA loss

python main.py --cfg config/CIFAR100_LT/gala_imba100.yaml

GALA loss + Prediction Re-balancing Strategy

python main.py --cfg config/CIFAR100_LT/gala_pn_imba100.yaml

Results and Models

CIFAR100

Imbalance Factor	GALA loss	GALA loss + Prediction Re-balancing Strategy	Log	Model
200	48.0%	48.5%	link	link
100	52.1%	52.3%	link	link
50	56.0%	56.5%	link	link
10	63.8%	64.2%	link	link

Large-scale Datasets

Dataset	GALA loss	GALA loss + Prediction Re-balancing Strategy(τ)	Log	Model
ImageNet-LT	53.4%	54.9%(τ=1.7)	train.log fine-tune-τ.log	link
iNa'2018	71.2%	73.3%(τ=1.5)	train.log fine-tune-τ.log	link
Places-LT	41.0%	41.4%(τ=1.4)	train.log fine-tune-τ.log	link

to do list

• Support CIFAR100-LT dataset
• Support ImageNet-LT
• Support iNaturalist2018
• Support Places365-LT
• More results and models

Acknowledgment

We refer to some codes from BalancedMetaSoftmax. Many thanks to the authors.