[CVPR2022] This repository contains code for the paper "Nested Collaborative Learning for Long-Tailed Visual Recognition", published at CVPR 2022

Nested Collaborative Learning for Long-Tailed Visual Recognition

This repository is the official PyTorch implementation of the paper in CVPR 2022:

Nested Collaborative Learning for Long-Tailed Visual Recognition
Jun Li, Zichang Tan, Jun Wan, Zhen Lei, Guodong Guo
[PDF]  

Main requirements

torch >= 1.7.1 #This is the version I am using, other versions may be accteptable, if there is any problem, go to https://pytorch.org/get-started/previous-versions/ to get right version(espicially CUDA) for your machine.
tensorboardX >= 2.1 #Visualization of the training process.
tensorflow >= 1.14.0 #convert long-tailed cifar datasets from tfrecords to jpgs.
Python 3.6 #This is the version I am using, other versions(python 3.x) may be accteptable.

Detailed requirement

pip install -r requirements.txt

Prepare datasets

This part is mainly based on https://github.com/zhangyongshun/BagofTricks-LT

We provide three datasets in this repo: long-tailed CIFAR (CIFAR-LT), long-tailed ImageNet (ImageNet-LT), iNaturalist 2018 (iNat18) and Places_LT.

The detailed information of these datasets are shown as follows:

-"CIFAR-10-LT-100" means the long-tailed CIFAR-10 dataset with the imbalance factor beta = 100.

-"Imbalance factor" is defined as: beta = Max images / Min images.

• Data format

The annotation of a dataset is a dict consisting of two field: annotations and num_classes. The field annotations is a list of dict with image_id, fpath, im_height, im_width and category_id.

Here is an example.

{
    'annotations': [
                    {
                        'image_id': 1,
                        'fpath': '/data/iNat18/images/train_val2018/Plantae/7477/3b60c9486db1d2ee875f11a669fbde4a.jpg',
                        'im_height': 600,
                        'im_width': 800,
                        'category_id': 7477
                    },
                    ...
                   ]
    'num_classes': 8142
}

• CIFAR-LT

  Cui et al., CVPR 2019 firstly proposed the CIFAR-LT. They provided the download link of CIFAR-LT, and also the codes to generate the data, which are in TensorFlow.

  You can follow the steps below to get this version of CIFAR-LT:

  1. Download the Cui's CIFAR-LT in GoogleDrive or Baidu Netdisk (password: 5rsq). Suppose you download the data and unzip them at path /downloaded/data/.
  2. Run tools/convert_from_tfrecords, and the converted CIFAR-LT and corresponding jsons will be generated at /downloaded/converted/.

  # Convert from the original format of CIFAR-LT
  python tools/convert_from_tfrecords.py  --input_path /downloaded/data/ --output_path /downloaded/converted/

• ImageNet-LT

  You can use the following steps to convert from the original images of ImageNet-LT.

  1. Download the original ILSVRC-2012. Suppose you have downloaded and reorgnized them at path /downloaded/ImageNet/, which should contain two sub-directories: /downloaded/ImageNet/train and /downloaded/ImageNet/val.
  2. Directly replace the data root directory in the file dataset_json/ImageNet_LT_train.json, dataset_json/ImageNet_LT_val.json,You can handle this with any editor, or use the following command.

  # replace data root
  python tools/replace_path.py --json_file dataset_json/ImageNet_LT_train.json --find_root /media/ssd1/lijun/ImageNet_LT --replaces_to /downloaded/ImageNet
  
  python tools/replace_path.py --json_file dataset_json/ImageNet_LT_val.json --find_root /media/ssd1/lijun/ImageNet_LT --replaces_to /downloaded/ImageNet
  

• iNat18

  You can use the following steps to convert from the original format of iNaturalist 2018.

  1. The images and annotations should be downloaded at iNaturalist 2018 firstly. Suppose you have downloaded them at path /downloaded/iNat18/.
  2. Directly replace the data root directory in the file dataset_json/iNat18_train.json, dataset_json/iNat18_val.json,You can handle this with any editor, or use the following command.

  # replace data root
  python tools/replace_path.py --json_file dataset_json/iNat18_train.json --find_root /media/ssd1/lijun/inaturalist2018/train_val2018 --replaces_to /downloaded/iNat18
  
  python tools/replace_path.py --json_file dataset_json/iNat18_val.json --find_root /media/ssd1/lijun/inaturalist2018/train_val2018 --replaces_to /downloaded/iNat18
  

• Places_LT

  You can use the following steps to convert from the original format of Places365-Standard.

  1. The images and annotations should be downloaded at Places365-Standard firstly. Suppose you have downloaded them at path /downloaded/Places365/.
  2. Directly replace the data root directory in the file dataset_json/Places_LT_train.json, dataset_json/Places_LT_val.json,You can handle this with any editor, or use the following command.

  # replace data root
  python tools/replace_path.py --json_file dataset_json/Places_LT_train.json --find_root /media/ssd1/lijun/data/places365_standard --replaces_to /downloaded/Places365
  
  python tools/replace_path.py --json_file dataset_json/Places_LT_val.json --find_root /media/ssd1/lijun/data/places365_standard --replaces_to /downloaded/Places365
  

Usage

First, prepare the dataset and modify the relevant paths in config/CIFAR100/cifar100_im100_NCL.yaml

Parallel training with DataParallel

1, Train
# Train long-tailed CIFAR-100 with imbalanced ratio of 100. 
# `GPUs` are the GPUs you want to use, such as '0' or`0,1,2,3`.
bash data_parallel_train.sh /home/lijun/papers/NCL/config/CIFAR/CIFAR100/cifar100_im100_NCL.yaml 0

Distributed training with DistributedDataParallel

Note that if you choose to train with DistributedDataParallel, the BATCH_SIZE in .yaml indicates the number on each GPU!

Default training batch-size: CIFAR: 64; ImageNet_LT: 256; Places_LT: 256; iNat18: 512.

e.g. if you want to train NCL with batch-size=512 on 8 GPUS, you should set the BATCH_SIZE in .yaml to 64.

1, Change the NCCL_SOCKET_IFNAME in run_with_distributed_parallel.sh to [your own socket name]. 
export NCCL_SOCKET_IFNAME = [your own socket name]

2, Train
# Train inaturalist2018. 
# `GPUs` are the GPUs you want to use, such as `0,1,2,3,4,5,6,7`.
# `NUM_GPUs` are the number of GPUs you want to use. If you set `GPUs` to `0,1,2,3,4,5,6,7`, then `NUM_GPUs` should be `8`.
bash distributed_data_parallel_train.sh config/iNat18/inat18_NCL.yaml 8 0,1,2,3,4,5,6,7

Citation

If you find our work inspiring or use our codebase in your research, please consider giving a star and a citation.

@inproceedings{li2022nested,
  title={Nested Collaborative Learning for Long-Tailed Visual Recognition},
  author={Li, Jun and Tan, Zichang and Wan, Jun and Lei, Zhen and Guo, Guodong},
  booktitle={IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
  year={2022}
}

Acknowledgements

This is a project based on Bag of tricks.

The data augmentations in dataset are based on PaCo

The MOCO in constrstive learning is based on MOCO