Welcome to the official p2pfl Google Group

[Pedro Guijas]

In this group all the news about the p2pfl library will be communicated. 

GitHub 

[Boy it]

hello, What papers should I refer to if I want to study the federated learning of ring architecture? It is best to have the source code.

[Pedro Guijas]

Hi! I recommend you to review the development memory of the first version of the library (https://github.com/pguijas/federated_learning_p2p/blob/main/other/memoria.pdf), especially the section on theoretical fundamentals, where it is explained in detail and includes references to papers used to guide the development. The document is written in Spanish but I doubt it will be a problem.

Best regards!

[Matteo Balice]

Is it possible to add new nodes and removing others during the training process without stopping it?

[Pedro Guijas]

Unfortunately at this time new node additions during learning have been disabled (the elimination of nodes is contemplated). In essence it has been disabled to avoid problems and inconsistencies in the node voting lists. However, it is planned to allow it in the future.

I encourage you to give it a spin and do a PR if you get good results. If you want more information about node inconsistencies feel free to ask.

Best regards,
Pedro.

[Matteo Balice]

I repost here the message so that anyone interested can read it.
Which type of  inconsistencies have you dealt with?

[Pedro Guijas]

In essence, what we should do would be to indicate in the handshake if a training is being executed. In this way, the node will seek to participate in the next round (remember that for an existing round it will already have defined a trainset so there is no problem).

The problem is that inconsistencies would occur if the candidates for a node (which will also be the nodes for which voting will be expected) are set and a new node is connected to the network at that small instant of time. Having different candidates, obtaining the most voted nodes will cause divergence in the trainset.

 nc_votes = {

                k: v for k, v in self.__train_set_votes.items() if k in candidates

            }

It may not be too complicated to solve, but due to lack of time it has been impossible for me to spend time to fix this little bug and I have simply decided that the inconsistencies cannot be caused.

Best regards, 

Pedro.