Deterministic client sampling

Flower Help - Beginners
1

Hi,

I wanted to samples each rounds clients by id’s, meaning that if i’m using 20% of clients and i have 100 clients i want to start with 0 to 19 first clients and so on. I’m not sure if it should be on the configure_fit function on my custom strategy or it should be on the client manager… Do you have already made examples with my problem ?

Kind regards,
Mr. Sunshine.

2

Okay, i think i’ve answered my self.

The function configure_fit strategy side calls on the client_manager.sample(), this function can take a criteron that invalidate clients.

I’ve implemented my own criteron for having a deterministic layout.

class DeterministicCriteron(Criterion):
    def __init__(self):
        self.round = 0
        self.num_clients = 100
        self.ids = 0
        self.mapping = {}

    def select(self, client: ClientProxy) -> bool:
        """Decide whether a client should be eligible for sampling or not."""
        plage_min = self.round * 20
        plage_max = plage_min + 19

        if client.cid not in self.mapping:
            identifiant = self.ids
            self.mapping[client.cid] = identifiant
            self.ids += 1

        eligible = plage_min <= self.mapping[client.cid] <= plage_max

        return eligible
    
    def update(self):
        self.round = (self.round + 1) % 5

It’s only a draft right now. I should have more variable in my init to be compatible with more than my fraction fit

Hope it can be usefull to anyone reading.

3

Hi, yes, that should work, assuming that the clients have cids attributes.