Multi-label Classification on MS COCO 2014/2017 (test)
94.29Macro AUCCentralized
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| Centralizedbeta (Dirichlet concentration parameter)=0.05, gamma (class-presence ratio)=0.75, Maximum classes per client=60 of 80, Backbone=ResNet-18, Pre-training=ImageNet-pretrained, Communication rounds=100, Batch size=322025.09 | 94.29 | 58.43 | 95.6 | 64.4 | |
| FedNCA-MLbeta (Dirichlet concentration parameter)=0.05, gamma (class-presence ratio)=0.75, Maximum classes per client=60 of 80, Backbone=ResNet-18, Pre-training=ImageNet-pretrained, Communication rounds=100, Batch size=322025.09 | 94.1 | 56.28 | 94.71 | 61.71 | |
| FedCurvbeta (Dirichlet concentration parameter)=0.05, gamma (class-presence ratio)=0.75, Maximum classes per client=60 of 80, Backbone=ResNet-18, Pre-training=ImageNet-pretrained, Communication rounds=100, Batch size=322025.09 | 94.03 | 54.34 | 95.4 | 60.82 | |
| FedProxbeta (Dirichlet concentration parameter)=0.05, gamma (class-presence ratio)=0.75, Maximum classes per client=60 of 80, Backbone=ResNet-18, Pre-training=ImageNet-pretrained, Communication rounds=100, Batch size=322025.09 | 93.74 | 53.65 | 94.87 | 60.58 | |
| FedAvgbeta (Dirichlet concentration parameter)=0.05, gamma (class-presence ratio)=0.75, Maximum classes per client=60 of 80, Backbone=ResNet-18, Pre-training=ImageNet-pretrained, Communication rounds=100, Batch size=322025.09 | 93.66 | 53.21 | 94.74 | 60.35 | |
| SCAFFOLDbeta (Dirichlet concentration parameter)=0.05, gamma (class-presence ratio)=0.75, Maximum classes per client=60 of 80, Backbone=ResNet-18, Pre-training=ImageNet-pretrained, Communication rounds=100, Batch size=322025.09 | 93.55 | 53.5 | 94.68 | 60.15 | |
| FedLGTbeta (Dirichlet concentration parameter)=0.05, gamma (class-presence ratio)=0.75, Maximum classes per client=60 of 80, Backbone=ResNet-18, Pre-training=ImageNet-pretrained, Communication rounds=100, Batch size=322025.09 | 90.9 | 55.68 | 92.37 | 62.76 | |
| SphereFedbeta (Dirichlet concentration parameter)=0.05, gamma (class-presence ratio)=0.75, Maximum classes per client=60 of 80, Backbone=ResNet-18, Pre-training=ImageNet-pretrained, Communication rounds=100, Batch size=322025.09 | 84.87 | 35.26 | 80.62 | 50.58 |