search

JedMills / MTFL-For-Personalised-DNNs

Code for 'Multi-Task Federated Learning for Personalised Deep Neural Networks in Edge Computing', published in IEEE TPDS.
86 stars 20 forks source link

Maybe fedavg has a problem #10

Open MiaJ321 opened 2 years ago

MiaJ321 commented 2 years ago

Hello Jed, I'm sorry to bother you. Recently we did an experiment.And I found the test ccuracy of fedavg a little strange.I don't know if there are any problems with fedavg. Here is the test accuracy of some algorithms.All algorithm parameters are set in the case of control variables.
屏幕截图 2022-10-18 173718 It is obvious that in Mnist database,the test accurary of fedavg is better than the other two.However,in general, in the case of non-iid, the effect of fedavg should be lower than that of other algorithms. So we think these data are a little strange. But we are not sure where the problem occurred.So we would like to ask your opinion. Looking forward to your early reply. @JedMills

Warm regards, Mia

JedMills commented 2 years ago

Hi Mia,

Thanks for your question. I just answered a similar one from another user. Could you read that and see if it addresses your problem?

Warm regards, Jed

MiaJ321 commented 2 years ago

Hi Jed,

Thank you very much for your reply. After reading your answer, my doubts have been solved. Thanks again for your help.

Warm regards, Mia