OpenLT: An open-source project for long-tail classification

Supported Methods for Long-tailed Recognition:

Reproduce Results

Here we simply show part of results to prove that our implementation is reasonable.


Method Backbone Reported Result Our Implementation
CE ResNet-10 34.8 35.3
Decouple-cRT ResNet-10 41.8 41.8
Decouple-LWS ResNet-10 41.4 41.6
BalanceSoftmax ResNet-10 41.8 41.4
CE ResNet-50 41.6 43.2
LDAM-DRW* ResNet-50 48.8 51.2
Decouple-cRT ResNet-50 47.3 48.7
Decouple-LWS ResNet-50 47.7 49.3

CIFAR100-LT (Imbalance Ratio 100)

${\dagger}$ means the reported results are copied from LADE

Method Datatset Reported Result Our Implementation
CE CIFAR100-LT 39.1 40.3
LDAM-DRW CIFAR100-LT 42.04 42.9
LogitAdjust CIFAR100-LT 43.89 45.3
BalanceSoftmax$^{\dagger}$ CIFAR100-LT 45.1 46.47



  • Python >= 3.7, < 3.9
  • PyTorch >= 1.6
  • tqdm (Used in
  • tensorboard >= 1.14 (for visualization)
  • pandas
  • numpy

Dataset Preparation

CIFAR code will download data automatically with the dataloader. We use data the same way as classifier-balancing. For ImageNet-LT and iNaturalist, please prepare data in the data directory. ImageNet-LT can be found at this link. iNaturalist data should be the 2018 version from this repo (Note that it requires you to pay to download now). The annotation can be found at here. Please put them in the same location as below:

├── cifar-100-python
│   ├── file.txt~
│   ├── meta
│   ├── test
│   └── train
├── cifar-100-python.tar.gz
├── ImageNet_LT
│   ├── ImageNet_LT_open.txt
│   ├── ImageNet_LT_test.txt
│   ├── ImageNet_LT_train.txt
│   ├── ImageNet_LT_val.txt
│   ├── Tiny_ImageNet_LT_train.txt (Optional)
│   ├── Tiny_ImageNet_LT_val.txt (Optional)
│   ├── Tiny_ImageNet_LT_test.txt (Optional)
│   ├── test
│   ├── train
│   └── val
└── iNaturalist18
    ├── iNaturalist18_train.txt
    ├── iNaturalist18_val.txt
    └── train_val2018

Training and Evaluation Instructions

Single Stage Training

python -c path_to_config_file

For example, to train a model with LDAM Loss on CIFAR-100-LT:

python -c configs/CIFAR-100/LDAMLoss.json

Decouple Training (Stage-2)

python -c path_to_config_file -crt path_to_stage_one_checkpoints

For example, to train a model with LWS classifier on ImageNet-LT:

python -c configs/ImageNet-LT/R50_LWS.json -lws path_to_stage_one_checkpoints


To test a checkpoint, please put it with the corresponding config file.

python -r path_to_checkpoint


python -c path_to_config_file -r path_to_resume_checkpoint

Please see the pytorch template that we use for additional more general usages of this project

FP16 Training

If you set fp16 in utils/, it will enable fp16 training. However, this is susceptible to change (and may not work on all settings or models) and please double check if you are using it since we don’t plan to focus on this part if you request help. Only some models work (see autograd in the code). We do not plan to provide support on this because it is not within our focus (just for faster training and less memory requirement).
In our experiments, the use of FP16 training does not reduce the accuracy of the model, regardless of whether it is a small dataset (CIFAR-LT) or a large dataset(ImageNet_LT, iNaturalist).


We use tensorboard as a visualization tool, and provide the accuracy changes of each class and different groups during the training process:

tensorboard --logdir path_to_dir

We also provide the simple code to visualize feature distribution using t-SNE and calibration using the reliability diagrams, please check the parameters in and, and then run:




Pytorch template

This is a project based on this pytorch template. The readme of the template explains its functionality, although we try to list most frequently used ones in this readme.


This project is licensed under the MIT License. See LICENSE for more details. The parts described below follow their original license.


This project is mainly based on RIDE‘s code base. In the process of reproducing and organizing the code, it also refers to some other excellent code repositories, such as decouple and LDAM.


View Github