search

ongzh / ScaffoldFL

Implementation of Scaffold and Fedprox for Federated Learning using PyTorch
19 stars 10 forks source link

在cifar10上训练得到的global model的test acc始终为10% #1

Open heu-kevin opened 2 years ago

heu-kevin commented 2 years ago

使用这套代码在cifar10上进行训练，参数是20个client，local epoch为5，local batch size为128，lr为0.01， frac为1.0，优化器为sgd，同等参数下fedavg得到的模型的test acc在每次global epoch结束后都有明显提升，可是scaffold没有，想请教一下作者

ongzh commented 2 years ago

The non iid splitting is implemented differently from scaffold paper

nlokeshiisc commented 2 years ago

Even with IID setting, I could see that CIFAR acuracy using SCAFFOLD is stagnant at 30% accuracy.