在实现新的联邦学习算法时需要注意很多细节,或者说在进行联邦学习实验时需要注意的一些问题:
- 集群架构是server-client,还是Ring(decentralized)
- 集群中有多少个节点,每一轮通信参加的节点有多少个,这个概率是多少
- 模拟实验时,数据在各个节点是按照什么标准划分的,是否是non-IID的
- 异步延迟梯度更新怎么修改,在哪里修改
- 损失权重怎么收集并计算
In my paper, there is the staring point: more friendly to edge device. So I must have RaspiBerry experiment. But the 12 edge devices have limited RAM, CIFAR10 will not run successfully, some device will be killed when starting training. So I plan to run a real world application on them, such as Covid19 X ray dection.
FedMine can reffer to MOON and FedProxy