def main() -> None:
"""Start server and train a number of rounds."""
args = parse_args()
# Configure logger
configure(identifier="server", host=args.log_host)
server_setting = get_setting(args.setting).server
log(INFO, "server_setting: %s", server_setting)
# Load evaluation data
(_, _), (x_test, y_test) = tf_fashion_mnist_partitioned.load_data(
iid_fraction=0.0, num_partitions=1
)
if server_setting.dry_run:
x_test = x_test[0:50]
y_test = y_test[0:50]
# Load model (for centralized evaluation)
model = orig_cnn(input_shape=(28, 28, 1), seed=SEED)
# Create client_manager
client_manager = fl.SimpleClientManager()
# Strategy
eval_fn = get_eval_fn(model=model, num_classes=10, xy_test=(x_test, y_test))
on_fit_config_fn = get_on_fit_config_fn(
lr_initial=server_setting.lr_initial,
timeout=server_setting.training_round_timeout,
partial_updates=server_setting.partial_updates,
)
if server_setting.strategy == "fedavg":
strategy = fl.strategy.FedAvg(
fraction_fit=server_setting.sample_fraction,
min_fit_clients=server_setting.min_sample_size,
min_available_clients=server_setting.min_num_clients,
eval_fn=eval_fn,
on_fit_config_fn=on_fit_config_fn,
)
if server_setting.strategy == "fast-and-slow":
if server_setting.training_round_timeout is None:
raise ValueError(
"No `training_round_timeout` set for `fast-and-slow` strategy"
)
t_fast = (
math.ceil(0.5 * server_setting.training_round_timeout)
if server_setting.training_round_timeout_short is None
else server_setting.training_round_timeout_short
)
strategy = fl.strategy.FastAndSlow(
fraction_fit=server_setting.sample_fraction,
min_fit_clients=server_setting.min_sample_size,
min_available_clients=server_setting.min_num_clients,
eval_fn=eval_fn,
on_fit_config_fn=on_fit_config_fn,
importance_sampling=server_setting.importance_sampling,
dynamic_timeout=server_setting.dynamic_timeout,
dynamic_timeout_percentile=0.8,
alternating_timeout=server_setting.alternating_timeout,
r_fast=1,
r_slow=1,
t_fast=t_fast,
t_slow=server_setting.training_round_timeout,
)
if server_setting.strategy == "fedfs-v0":
if server_setting.training_round_timeout is None:
raise ValueError("No `training_round_timeout` set for `fedfs-v0` strategy")
t_fast = (
math.ceil(0.5 * server_setting.training_round_timeout)
if server_setting.training_round_timeout_short is None
else server_setting.training_round_timeout_short
)
strategy = fl.strategy.FedFSv0(
fraction_fit=server_setting.sample_fraction,
min_fit_clients=server_setting.min_sample_size,
min_available_clients=server_setting.min_num_clients,
eval_fn=eval_fn,
on_fit_config_fn=on_fit_config_fn,
r_fast=1,
r_slow=1,
t_fast=t_fast,
t_slow=server_setting.training_round_timeout,
)
if server_setting.strategy == "fedfs-v1":
if server_setting.training_round_timeout is None:
raise ValueError("No `training_round_timeout` set for `fedfs-v1` strategy")
strategy = fl.strategy.FedFSv1(
fraction_fit=server_setting.sample_fraction,
min_fit_clients=server_setting.min_sample_size,
min_available_clients=server_setting.min_num_clients,
eval_fn=eval_fn,
on_fit_config_fn=on_fit_config_fn,
dynamic_timeout_percentile=0.8,
r_fast=1,
r_slow=1,
t_max=server_setting.training_round_timeout,
use_past_contributions=True,
)
if server_setting.strategy == "qffedavg":
strategy = fl.strategy.QffedAvg(
q_param=0.2,
qffl_learning_rate=0.1,
fraction_fit=server_setting.sample_fraction,
min_fit_clients=server_setting.min_sample_size,
min_available_clients=server_setting.min_num_clients,
eval_fn=eval_fn,
on_fit_config_fn=on_fit_config_fn,
)
# Run server
log(INFO, "Instantiating server, strategy: %s", str(strategy))
server = fl.Server(client_manager=client_manager, strategy=strategy)
fl.app.server.start_server(
DEFAULT_SERVER_ADDRESS, server, config={"num_rounds": server_setting.rounds},
)
def main() -> None:
"""Start server and train a number of rounds."""
args = parse_args()
# Configure logger
configure(identifier="server", host=args.log_host)
server_setting = get_setting(args.setting).server
log(INFO, "server_setting: %s", server_setting)
# Load evaluation data
(_,
_), (x_test,
y_test) = tf_cifar_partitioned.load_data(iid_fraction=0.0,
num_partitions=1,
cifar100=NUM_CLASSES == 100)
if server_setting.dry_run:
x_test = x_test[0:50]
y_test = y_test[0:50]
# Load model (for centralized evaluation)
model = resnet50v2(input_shape=(32, 32, 3),
num_classes=NUM_CLASSES,
seed=SEED)
# Strategy
eval_fn = get_eval_fn(model=model,
num_classes=NUM_CLASSES,
xy_test=(x_test, y_test))
fit_config_fn = get_on_fit_config_fn(
lr_initial=server_setting.lr_initial,
timeout=server_setting.training_round_timeout,
partial_updates=server_setting.partial_updates,
)
if server_setting.strategy == "fedavg":
strategy = fl.server.strategy.FedAvg(
fraction_fit=server_setting.sample_fraction,
min_fit_clients=server_setting.min_sample_size,
min_available_clients=server_setting.min_num_clients,
eval_fn=eval_fn,
on_fit_config_fn=fit_config_fn,
)
if server_setting.strategy == "fast-and-slow":
if server_setting.training_round_timeout is None:
raise ValueError(
"No `training_round_timeout` set for `fast-and-slow` strategy")
strategy = fl.server.strategy.FastAndSlow(
fraction_fit=server_setting.sample_fraction,
min_fit_clients=server_setting.min_sample_size,
min_available_clients=server_setting.min_num_clients,
eval_fn=eval_fn,
on_fit_config_fn=fit_config_fn,
importance_sampling=server_setting.importance_sampling,
dynamic_timeout=server_setting.dynamic_timeout,
dynamic_timeout_percentile=0.8,
alternating_timeout=server_setting.alternating_timeout,
r_fast=1,
r_slow=1,
t_fast=math.ceil(0.5 * server_setting.training_round_timeout),
t_slow=server_setting.training_round_timeout,
)
# Run server
fl.server.start_server(
DEFAULT_SERVER_ADDRESS,
config={"num_rounds": server_setting.rounds},
strategy=strategy,
)
