Home

Awesome

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:

Requirements:

This code is based on:

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:

DatasetBackboneEpochsTop-1 Acc(%)
ImageNet-LTResNeXt-509058.3
ImageNet-LTResNeXt-5040058.8
iNaturalist 2018ResNet-5010073.1
iNaturalist 2018ResNet-5040074.5
CIFAR-100-LT (imb. 100)ResNet-3220053.0
CIFAR-100-LT (imb. 50)ResNet-3220057.6
CIFAR-100-LT (imb. 10)ResNet-3220065.7
CIFAR-100-LT (imb. 100)ResNet-3240054.0
CIFAR-100-LT (imb. 50)ResNet-3240058.1
CIFAR-100-LT (imb. 10)ResNet-3240067.0