unique-chan / Complement-Cross-Entropy

Imbalanced Image Classification with Complement Cross Entropy (Pytorch)

Imbalanced Image Classification with Complement Cross Entropy (Pytorch)

Yechan Kim, Younkwan Lee, and Moongu Jeon

Cite this paper

News:

• (06/2022) Now, you can easily try our loss function with Holocron. Holocron includes implementations of recent Deep Learning tricks in computer vision, easily paired up with your favorite framework and model zoo.
• (08/2021) Our paper is accepted to Pattern Recognition Letters 🎉.

This repository contains:

• Complement Cross Entropy (code)
• For simplicity, classification code is provided separately in this GitHub repo 🖱️: you can easily use Complement Cross Entropy by passing --loss_function='CCE' for executing train.py. For details, please visit the above repository.

Prerequisites

• See requirements.txt

torch
torchvision

Code

class CCE(nn.Module):
    def __init__(self, device, balancing_factor=1):
        super(CCE, self).__init__()
        self.nll_loss = nn.NLLLoss()
        self.device = device # {'cpu', 'cuda:0', 'cuda:1', ...}
        self.balancing_factor = balancing_factor

    def forward(self, yHat, y):
        # Note: yHat.shape[1] <=> number of classes
        batch_size = len(y)
        # cross entropy
        cross_entropy = self.nll_loss(F.log_softmax(yHat, dim=1), y)
        # complement entropy
        yHat = F.softmax(yHat, dim=1)
        Yg = yHat.gather(dim=1, index=torch.unsqueeze(y, 1))
        Px = yHat / (1 - Yg) + 1e-7
        Px_log = torch.log(Px + 1e-10)
        y_zerohot = torch.ones(batch_size, yHat.shape[1]).scatter_(
            1, y.view(batch_size, 1).data.cpu(), 0)
        output = Px * Px_log * y_zerohot.to(device=self.device)
        complement_entropy = torch.sum(output) / (float(batch_size) * float(yHat.shape[1]))

        return cross_entropy - self.balancing_factor * complement_entropy

Citation

If you use this code for your research, please cite the following paper:

@article{kim2021imbalanced,
  title={Imbalanced image classification with complement cross entropy},
  author={Kim, Yechan and Lee, Younkwan and Jeon, Moongu},
  journal={Pattern Recognition Letters},
  volume={151},
  pages={33--40},
  year={2021},
  publisher={Elsevier}
}

Contribution

If you find any bugs or have opinions for further improvements, please feel free to create a pull request or contact me ([e-mail]). All contributions are welcome.

Reference

1. Hao-Yun Chen, Pei-Hsin Wang, Chun-Hao Liu, Shih-Chieh Chang, Jia-Yu Pan, Yu-Ting Chen, Wei Wei, and Da-Cheng Juan. Complement objective training. arXiv preprint arXiv:1903.01182, 2019.
2. Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Doll ́ar.Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980–2988, 2017.
3. Tong He, Zhi Zhang, Hang Zhang, Zhongyue Zhang, Junyuan Xie, andMu Li. Bag of tricks for image classification with convolutional neuralnetworks. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 558–567, 2019.
4. https://github.com/calmisential/Basic_CNNs_TensorFlow2
5. https://github.com/Hsuxu/Loss_ToolBox-PyTorch
6. https://github.com/weiaicunzai/pytorch-cifar100