zerocpp / fedhkd Goto Github PK

I didn't find the code for FedHKD, but I browsed the whole code, so is FedDFKD the FedHKD code? But when I use the FedDFKD, Resnet 18 and Cifar-100 data sets to run the code, the loss3 feature will suddenly become very huge around 30 rounds, and then the entire code will explode. Is it convenient for you to provide your .sh file or further updated code?

Could you kindly explain why the global test accuracy of the FedMD consistently hovers around 10%? This seems to deviate from the results reported in the paper. Is there an issue that needs to be addressed?

The loss function of soft prediction in the paper is Euclidean norm, while the loss function of soft prediction in the code is KL divergence

The loss function of soft prediction in the paper is Euclidean norm, while the loss function of soft prediction in the code is KL divergence，I wonder why