*This question was migrated from Github Discussions.
Original Question:
My doubt is rather conceptual. After the server collects the models from all the clients and weights them, and sends them back to the clients, do the clients all use the same central model they receive, or do they do a weighting between the central one and theirs before training further?
Answer 1:
They use the aggregated model sent by the server without any modification. But you can change this behaviour as you prefer
Question 2:
Does flower give the option to change this behaviour?
Answer 2:
Yes, with flower you can do pretty much everything. When you implement your client class, in the fit method it receives as arguments aggregated parameters together with the configuration dictonary. From there you can modify the parameters before/after starting the local training
Thank you !! 