Configuring individual clients about flower

[gioperin]

Hi @jafermarq please note that the documentation is highly misleading because the title of that section is "configuring individual clients" and the example does not achieve what the title suggests. The workaround you cited is the one I am using right now. I want to point out that this procedure is extremely inefficient because it implies copying N times all the fields inside FitIns, among which there are the parameters, ending up in poor memory usage. Don't you think it would be a good idea to use another dictionary other than FitIns to pass configuration parameters so that we don't need to copy the model parameters?

Thank you for your reply!

[danielphan2003]

This can be achieved by customizing an existing strategy or by implementing a custom strategy from scratch. Here’s a nonsensical example that customizes FedAvg by adding a custom "hello": "world" configuration key/value pair to the config dict of a single client (only the first client in the list, the other clients in this round to not receive this “special” config value):

# Add special "hello": "world" config key/value pair,
# but only to the first client in the list
_, fit_ins = client_instructions[0]  # First (ClientProxy, FitIns) pair
fit_ins.config["hello"] = "world"  # Change config for this client only

I think the code needs to be updated so that it is in fact only the first client receives {"hello": "world"}, such as:

_, fit_ins = client_instructions[0]
fit_ins = FitIns(parameters=fit_ins.parameters, fit_ins.config | {"hello": "world"})

Is this a Pythonic behavior or somewhere else in the codebase would have required such workarounds in the first place?

EDIT: I was wrong, using copy would solve the problem but would result in poor memory usage as well:

client_instructions = [(client, copy(fit_ins)) for client, fit_ins in client_instructions]
_, fit_ins = client_instructions[0]
fit_ins.config = fit_ins.config | {"hello": "world"})

[jafermarq]

Hey @gioperin, that part of the documentation indicates (as you point out) how to configure clients, but not individually. You can still achieve what you want by implementing a custom strategy and overriding the configure_fit() method and assign a different FitIns object to each client that has been sampled.