FedGKT client send too large model to the server (MQTT backend)

[KOUDA-AMINE]

Hi, I'm trying to test FedGKT on raspberry pi 4 using MQTT as a communication backend, and I'm confused about some params:

• whether_training_on_client;
• whether_distill_on_the_server;
• sweep. The issue is that after finishing the first comm-round the clients are unable to send the params to the server because their size is too big ( more than 4 Gb when trained on client =1 and over 1.4 Gb as a minimum). How do you explain this and how could resolve this problem? because even using compression with pickle and Zlib the problem persists.

[KOUDA-AMINE]

any suggestions ??

[chaoyanghe]

@KOUDA-AMINE Hi kouda, we were fully focusing on product in the past few months rather than checking this research paper. Could you please share the source code repo to reproduce your issue?

[fedml-dimitris]

Hello @KOUDA-AMINE can you please share a reproducible example?