A scientific and useful toolbox, which contains practical and effective long-tail related tricks with extensive experimental results

Bag of tricks for long-tailed visual recognition with deep convolutional neural networks

This repository is the official PyTorch implementation of AAAI-21 paper Bag of Tricks for Long-Tailed Visual Recognition with Deep Convolutional Neural Networks, which provides practical and effective tricks used in long-tailed image classification.

Trick gallery: trick_gallery.md

• The tricks will be constantly updated. If you have or need any long-tail related trick newly proposed, please to open an issue or pull requests. Make sure to attach the results in corresponding md files if you pull a request with a new trick.
• For any problem, such as bugs, feel free to open an issue.

Paper collection of long-tailed visual recognition

Awesome-of-Long-Tailed-Recognition

Long-Tailed-Classification-Leaderboard

Development log

• 2020-12-26 - Reorignize all the codes, according to BBN.
• 2020-12-30 - Add codes of torch.nn.parallel.DistributedDataParallel. Support apex in both torch.nn.DataParallel and torch.nn.parallel.DistributedDataParallel.
• 2021-01-02 - Add LDAMLoss, NeurIPS 2019, and a regularization method: label smooth cross-entropy, CVPR 2016.
• 2021-01-05 - Add SEQL (softmax equalization loss), CVPR 2020.
• 2021-01-10 - Add CDT (class-dependent temparature), arXiv 2020, BSCE (balanced-softmax cross-entropy), NeurIPS 2020, and support a smooth version of cost-sensitive cross-entropy (smooth CS_CE), which add a hyper-parameter $ \gamma$ to vanilla CS_CE. In smooth CS_CE, the loss weight of class i is defined as: $(\frac{N_{min}}{N_i})^\gamma$, where $\gamma \in [0, 1]$, $N_i$ is the number of images in class i. We can set $\gamma = 0.5$ to get a square-root version of CS_CE.
• 2021-01-11 - Add a mixup related method: Remix, ECCV 2020 workshop.
• 2021-02-19 - Test and add the results of two-stage training in trick_gallery.md
• 2021-01-30 - Add the results of combining mixup methods and re-balancing in trick_combination.md.
• 2021-04-22 - Add one option (TRAIN.APEX) in config.py, so you can set TRAIN.APEX to False for training without using apex.
• 2021-04-23 - Add CrossEntropyLabelAwareSmooth (label-aware smoothing, CVPR 2021) in trick_gallery.md.
• 2021-04-24 - Add classifier-balancing and corresponding experiments in Two-stage training in trick_gallery.md, including $\tau$-normalization, cRT and LWS.
• 2021-04-24 - Add the validation running command, which loads a trained model, then returns the validation acc and a corresponding confusion matrix figure. See Usage in this README for details.
• 2021-05-15 - Add CONFIGs and experimental results of our bag of tricks in trick_combination.md.
• 2021-05-19 - Add CONFIGs and experimental results of BBN-style sampling in trick_gallery.md, which consists of a uniform sampler and a reverse sampler.
• 2021-05-19 - Add results of bbn-style sampling and input mixup in trick_combination.md.

Trick gallery and combinations

Brief inroduction

We divided the long-tail realted tricks into four families: re-weighting, re-sampling, mixup training, and two-stage training. For more details of the above four trick families, see the original paper.

Detailed information :

• Trick gallery:

  Tricks, corresponding results, experimental settings, and running commands are listed in trick_gallery.md.

• Trick combinations:

  Combinations of different tricks, corresponding results, experimental settings, and running commands are listed in trick_combination.md.

• These tricks and trick combinations, which provide the corresponding results in this repo, have been reorgnized and tested. We are trying our best to deal with the rest, which will be constantly updated.

Main requirements

torch >= 1.4.0
torchvision >= 0.5.0
tensorboardX >= 2.1
tensorflow >= 1.14.0 #convert long-tailed cifar datasets from tfrecords to jpgs
Python 3
apex

• We provide the detailed requirements in requirements.txt. You can run pip install requirements.txt to create the same running environment as ours.
• The apex is recommended to be installed for saving GPU memories:

pip install -U pip
git clone https://github.com/NVIDIA/apex
cd apex
pip install -v --disable-pip-version-check --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" ./

• If the apex is not installed, the Distributed training with DistributedDataParallel in our codes cannot be used.

Preparing the datasets

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

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 $\beta = \frac{\text{Max images}}{\text{Min images}}$.

• 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

  There are two versions of CIFAR-LT.

  1. 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/ --out_path /downloaded/converted/

  2. Cao et al., NeurIPS 2019 followed Cui et al., CVPR 2019's method to generate the CIFAR-LT randomly. They modify the CIFAR datasets provided by PyTorch as this file shows.

• 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. Download the train/test splitting files (ImageNet_LT_train.txt and ImageNet_LT_test.txt) in GoogleDrive or Baidu Netdisk (password: cj0g). Suppose you have downloaded them at path /downloaded/ImageNet-LT/.
  3. Run tools/convert_from_ImageNet.py, and you will get two jsons: ImageNet_LT_train.json and ImageNet_LT_val.json.

  # Convert from the original format of ImageNet-LT
  python tools/convert_from_ImageNet.py --input_path /downloaded/ImageNet-LT/ --image_path /downloaed/ImageNet/ --output_path ./

• 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. Run tools/convert_from_iNat.py, and use the generated iNat18_train.json and iNat18_val.json to train.

  # Convert from the original format of iNaturalist
  # See tools/convert_from_iNat.py for more details of args 
  python tools/convert_from_iNat.py --input_json_file /downloaded/iNat18/train2018.json --image_path /downloaded/iNat18/images --output_json_file ./iNat18_train.json
  
  python tools/convert_from_iNat.py --input_json_file /downloaded/iNat18/val2018.json --image_path /downloaded/iNat18/images --output_json_file ./iNat18_val.json 

Usage

In this repo:

• The results of CIFAR-LT (ResNet-32) and ImageNet-LT (ResNet-10), which need only one GPU to train, are gotten by DataParallel training with apex.

• The results of iNat18 (ResNet-50), which need more than one GPU to train, are gotten by DistributedDataParallel training with apex.

• If more than one GPU is used, DistributedDataParallel training is efficient than DataParallel training, especially when the CPU calculation forces are limited.

Training

Parallel training with DataParallel

1, To train
# To train long-tailed CIFAR-10 with imbalanced ratio of 50. 
# `GPUs` are the GPUs you want to use, such as `0,4`.
bash data_parallel_train.sh configs/test/data_parallel.yaml GPUs

Distributed training with DistributedDataParallel

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, To train
# To train long-tailed CIFAR-10 with imbalanced ratio of 50. 
# `GPUs` are the GPUs you want to use, such as `0,1,4`.
# `NUM_GPUs` are the number of GPUs you want to use. If you set `GPUs` to `0,1,4`, then `NUM_GPUs` should be `3`.
bash distributed_data_parallel_train.sh configs/test/distributed_data_parallel.yaml NUM_GPUs GPUs

Validation

You can get the validation accuracy and the corresponding confusion matrix after running the following commands.

See main/valid.py for more details.

1, Change the TEST.MODEL_FILE in the yaml to your own path of the trained model firstly.
2, To do validation
# `GPUs` are the GPUs you want to use, such as `0,1,4`.
python main/valid.py --cfg [Your yaml] --gpus GPUS

The comparison between the baseline results using our codes and the references [Cui, Kang]

• We use Top-1 error rates as our evaluation metric.

• From the results of two CIFAR-LT, we can see that the CIFAR-LT provided by Cao has much lower Top-1 error rates on CIFAR-10-LT, compared with the baseline results reported in his paper. So, in our experiments, we use the CIFAR-LT of Cui for fairness.
• For the ImageNet-LT, we find that the color_jitter augmentation was not included in our experiments, which, however, is adopted by other methods. So, in this repo, we add the color_jitter augmentation on ImageNet-LT. The old baseline without color_jitter is 64.89, which is +1.15 points higher than the new baseline.
• You can click the Baseline in the table below to see the experimental settings and corresponding running commands.

Citation

@inproceedings{zhang2020tricks,
  author    = {Yongshun Zhang and Xiu{-}Shen Wei and Boyan Zhou and Jianxin Wu},
  title     = {Bag of Tricks for Long-Tailed Visual Recognition with Deep Convolutional Neural Networks},
  booktitle = {AAAI},
  year      = {2021},
}

Contacts

If you have any question about our work, please do not hesitate to contact us by emails provided in the paper.