karhoutam/per-fedavg

PyTorch Implementation of Personalized federated learning with theoretical guarantees: A model-agnostic meta-learning approach

Hi there 👋

• 🤖 I'm a postgraduate student of Shenzhen University, China (Y3).
• 🌱 In school, I major in federated learning, especially personalized federated learning (pFL).
• 👀 Feel free to contact me for asking questions or seeking collaboration.
• 🧐 I'm currently learning CUDA C programming and LLMs.
• 🎊 The Next Stage: Huawei Cloud, Shenzhen.

per-fedavg's Issues

About gradient descent on the client side

Hi, Jiahao Tan.
Thanks for your work.

I have some confusion about the code on lines 98 of "per-fedavg /perfedavg.py".
param.data.sub_(self.beta * grad1 - self.beta * self.alpha * grad2)
According to the formula in the article, I think "self.beta * self.alpha * grad2" seems to miss "grad1".

question about SGD one step

In the section 5, the author wrote "Note that the model obtained by any of these three methods is later updated using one step of stochastic gradient descent at the test time".
May I ask you why the testset data is used here instead of the validation set data ? This problem confuses me a lot.