
摘要
本文提出了一种参数化对比学习(Parametric Contrastive Learning, PaCo),以应对长尾识别问题。基于理论分析,我们发现监督对比损失倾向于偏向高频类别,从而加剧了数据不平衡学习的难度。为此,我们引入一组可学习的类别专属参数化中心,从优化角度实现类别平衡。进一步地,我们在类别平衡的设定下分析了PaCo损失函数的性质。分析表明,随着同一类样本被更紧密地拉向其对应的中心,PaCo能够自适应地增强同类样本之间的聚集强度,从而有效促进困难样本的学习。在长尾CIFAR、ImageNet、Places以及iNaturalist 2018数据集上的实验结果表明,PaCo达到了长尾识别任务的新SOTA(state-of-the-art)性能。在完整的ImageNet数据集上,使用PaCo损失训练的模型在多种ResNet骨干网络下均超越了监督对比学习方法,例如,我们的ResNet-200模型取得了81.8%的top-1准确率。相关代码已开源,地址为:https://github.com/dvlab-research/Parametric-Contrastive-Learning。
代码仓库
dvlab-research/imbalanced-learning
pytorch
GitHub 中提及
silicx/dlsa
pytorch
GitHub 中提及
dvlab-research/rescom
pytorch
GitHub 中提及
dvlab-research/parametric-contrastive-learning
官方
pytorch
GitHub 中提及
jiequancui/Parametric-Contrastive-Learning
官方
pytorch
GitHub 中提及
基准测试
| 基准 | 方法 | 指标 |
|---|---|---|
| image-classification-on-imagenet | ResNet-101 | Top 1 Accuracy: 80.9% |
| image-classification-on-imagenet | ResNet-200 | Top 1 Accuracy: 81.8% |
| image-classification-on-imagenet | ResNet-152 | Top 1 Accuracy: 81.3% |
| image-classification-on-inaturalist-2018 | PaCo(ResNet-152) | Top-1 Accuracy: 75.2% |
| long-tail-learning-on-cifar-10-lt-r-10 | PCL | Error Rate: 9.14 |
| long-tail-learning-on-cifar-100-lt-r-100 | PCL | Error Rate: 49.10 |
| long-tail-learning-on-imagenet-lt | PaCo(ResNeXt-50) | Top-1 Accuracy: 58.2 |
| long-tail-learning-on-imagenet-lt | PaCo(ResNeXt101-32x4d) | Top-1 Accuracy: 60.0 |
| long-tail-learning-on-inaturalist-2018 | PaCo(ResNet-152) | Top-1 Accuracy: 75.2% |
| long-tail-learning-on-places-lt | PaCo | Top-1 Accuracy: 41.2 |