Long-Tailed Recognition by Mutual Information Maximization between Latent Features and Ground-Truth Labels

Python code for ICML 2023 paper entitled "Long-Tailed Recognition by Mutual Information Maximization between Latent Features and Ground-Truth Labels"

Test environment:

• Ubuntu 18.04
• CUDA 11.3, cuDNN 8.3.2

Requirements:

• Python 3.8
• Pytorch 1.12.1
• tqdm
• scikit-learn
• tensorboard
• matplotlib

This code is based on:

• https://github.com/facebookresearch/moco
• https://github.com/dvlab-research/Parametric-Contrastive-Learning/tree/main/PaCo/LT

Experiment procedure:

1. (Set dataset directory) Change --data argument of sh/ImageNetLT_train_teacher.sh and sh/ImageNetLT_train_student.sh
2. (Train teacher) Run bash ./sh/ImageNetLT_train_teacher.sh
3. (Train student) Run bash ./sh/ImageNetLT_train_student.sh

Long-tailed recognition accuracy:

Dataset	Backbone	Epochs	Top-1 Acc(%)
ImageNet-LT	ResNeXt-50	90	58.3
ImageNet-LT	ResNeXt-50	400	58.8
iNaturalist 2018	ResNet-50	100	73.1
iNaturalist 2018	ResNet-50	400	74.5
CIFAR-100-LT (imb. 100)	ResNet-32	200	53.0
CIFAR-100-LT (imb. 50)	ResNet-32	200	57.6
CIFAR-100-LT (imb. 10)	ResNet-32	200	65.7
CIFAR-100-LT (imb. 100)	ResNet-32	400	54.0
CIFAR-100-LT (imb. 50)	ResNet-32	400	58.1
CIFAR-100-LT (imb. 10)	ResNet-32	400	67.0