摘要
我们提出了一种名为“Meta Pseudo Labels”的半监督学习方法，在ImageNet数据集上取得了90.2%的Top-1准确率，创下新的最优纪录，较现有最优方法提升了1.6%。与Pseudo Labels方法类似，Meta Pseudo Labels同样采用教师网络在无标签数据上生成伪标签，用于指导学生网络的学习。然而，与Pseudo Labels中教师网络固定不变不同，Meta Pseudo Labels中的教师网络会根据学生网络在有标签数据集上的表现反馈进行持续自适应调整。由此，教师网络能够生成更高质量的伪标签，从而更有效地指导学生网络的学习。相关代码将公开于：https://github.com/google-research/google-research/tree/master/meta_pseudo_labels。
代码仓库
YanYan0716/MPL
tf
GitHub 中提及
kekmodel/MPL-pytorch
pytorch
GitHub 中提及
ifsheldon/MPL_Lightning
pytorch
sayakpaul/PAWS-TF
tf
GitHub 中提及
ve450su2021-group26/Algorithm
pytorch
GitHub 中提及
retoschiegg/meta-pseudo-labels
tf
GitHub 中提及
usccolumbia/tsdnn
pytorch
GitHub 中提及
基准测试
| 基准 | 方法 | 指标 |
|---|---|---|
| image-classification-on-imagenet | Meta Pseudo Labels (EfficientNet-B6-Wide) | Number of params: 390M Top 1 Accuracy: 90% |
| image-classification-on-imagenet | Meta Pseudo Labels (ResNet-50) | Top 1 Accuracy: 83.2% |
| image-classification-on-imagenet | Meta Pseudo Labels (EfficientNet-L2) | Hardware Burden: 95040G Number of params: 480M Operations per network pass: Top 1 Accuracy: 90.2% Top 5 Accuracy: 98.8 |
| image-classification-on-imagenet-real | Meta Pseudo Labels (EfficientNet-L2) | Accuracy: 91.02% |
| image-classification-on-imagenet-real | Meta Pseudo Labels (EfficientNet-B6-Wide) | Accuracy: 91.12% |
| semi-supervised-image-classification-on-2 | Meta Pseudo Labels (ResNet-50) | Top 1 Accuracy: 73.89% Top 5 Accuracy: 91.38% |
| semi-supervised-image-classification-on-cifar | Meta Pseudo Labels (WRN-28-2) | Percentage error: 3.89± 0.07 |
| semi-supervised-image-classification-on-svhn | Meta Pseudo Labels (WRN-28-2) | Accuracy: 98.01 ± 0.07 |