Paddle FL MPC running on Multy Machines
Created by: SaviorD7
Hello everyone!
I am trying to run Paddle FL MPC on multy machines. (I used uci_housing_demo example, but remade it for my data) So , I have two linux (ubuntu 18.04) machines running in Google Cloud.
- I need to prepare data on each machine according to point 1 in the guide:
Data owner encrypts data. Concrete operations are consistent with “Prepare Data” in “Running on Single Machine”.
What does it mean? For example, I have a half part of data on first machine and next half part of data on second machine. I need to training these parts of data on each machine and send results in the one place.
So in my mind it seems like that: First machine (will be main machine): has its part of data , training NN and has results after Secoon machine: has its part of data, training NN and send results to the first machine So now, first machine have all the results recieve from two training.
Or maybe I dont understand it right ? Cause in point 2: According to the suffix of file name, distribute encrypted data files to /tmp directories of all 3 computation parties. For example, send house_feature.part0 and house_label.part0 to /tmp of party 0 with scp command.
says that there are 3 computation (okay, maybe it just use 3 machine) and there are divide data and send to party 0 (which is server or main machine I think so)
But , however, it seems that machine with part 0 will recieve all data from 1 and 2 parties and just training it by itself :/ So it is not going to be Federate Learning.
Can you please understand it more understandable? :~
Thanks