def run_fedavg_round(aggregator: 'BaseAggregatorParticipant',
participants: List['BaseTrainingParticipant'],
training_args: TrainArgs,
client_fraction=1.0):
"""
Routine to run a training round with the given clients based on the server model and then aggregate the results
:param client_fraction: client fraction to train with
:param aggregator: aggregator participant that will aggregate the resulting training models
:param participants: training participants in this round
:param training_args: training arguments for this round
:return:
"""
logger.debug('distribute the initial model to the clients.')
initial_model_state = aggregator.model.state_dict()
success_threshold = max(int(len(participants) * client_fraction),
1) if client_fraction < 1.0 else -1
participant_fraction = sample_randomly_by_fraction(participants,
client_fraction)
logger.debug(
f'starting training round with {len(participant_fraction)}/{len(participants)}.'
)
trained_participants = run_fedavg_train_round(initial_model_state,
participant_fraction,
training_args,
success_threshold=-1)
logger.debug('starting aggregation.')
num_train_samples = [p.num_train_samples for p in trained_participants]
aggregator.aggregate(trained_participants,
num_train_samples=num_train_samples)
logger.debug('distribute the aggregated global model to clients')
resulting_model_state = aggregator.model.state_dict()
overwrite_participants_models(resulting_model_state, participants)
def reptile_train_step(aggregator: ReptileServer,
participants: List[ReptileClient],
inner_training_args: TrainArgs,
meta_training_args: TrainArgs = None,
evaluation_mode: bool = False,
*args, **kwargs):
"""
Routine to run a Reptile training step
:param aggregator: aggregator participant that will aggregate the resulting training models
:param participants: training participants in this round
:param inner_training_args: training arguments for participant models
:param meta_training_args: training arguments for meta model
:param evaluation_mode: is evaluation step
:return:
"""
logger.debug('distribute the initial model to the clients.')
initial_model_state = copy.deepcopy(aggregator.model.state_dict())
overwrite_participants_models(initial_model_state, participants)
logger.debug('starting training round.')
run_train_round(participants, inner_training_args)
# Aggregate only when not in evaluation mode
if not evaluation_mode:
assert meta_training_args is not None, ('Argument meta_training_args '
'must not be None when not in evaluation_mode')
logger.debug(
(f"Starting aggregation: num_participants={len(participants)}, "
f"meta_learning_rate={meta_training_args.kwargs['meta_learning_rate']}")
)
aggregator.aggregate(
participants=participants,
meta_learning_rate=meta_training_args.kwargs['meta_learning_rate']
)
def run_fedavg_train_round(
initial_model_state: Dict[str, Tensor],
participants: List['BaseTrainingParticipant'],
training_args: TrainArgs,
success_threshold=-1) -> List['BaseTrainingParticipant']:
"""
Routine to run a single round of training on the clients and return the results additional args are passed to the
clients training routines.
:param initial_model_state: model state to communicate before training
:param participants: participants to train in this round
:param training_args: arguments passed for training
:param success_threshold: threshold for how many clients should at least participate in the round
:return:
"""
overwrite_participants_models(initial_model_state, participants)
successful_participants = []
for participant in participants:
try:
logger.debug(
f'invoking training on participant {participant._name}')
participant.train(training_args)
successful_participants.append(participant)
if success_threshold != -1 and success_threshold <= len(
successful_participants):
break
except GradientExplodingError as gradient_exception:
logger.error(
f'participant {participant._name} failed due to exploding gradients',
gradient_exception)
except Exception as e:
logger.error(f'training on participant {participant._name} failed',
e)
if success_threshold != -1 and len(
successful_participants) < success_threshold:
raise ExecutionError(
'Failed to execute training round, not enough clients participated successfully'
)
return successful_participants
def run_hierarchical_clustering(local_evaluation_steps,
seed,
lr,
name,
total_fedavg_rounds,
cluster_initialization_rounds,
client_fraction,
local_epochs,
batch_size,
num_clients,
sample_threshold,
num_label_limit,
train_args,
dataset,
partitioner_class,
linkage_mech,
criterion,
dis_metric,
max_value_criterion,
reallocate_clients,
threshold_min_client_cluster,
use_colored_images,
use_pattern,
train_cluster_args=None,
mean=None,
std=None):
fix_random_seeds(seed)
global_tag = 'global_performance'
global_tag_local = 'global_performance_personalized'
initialize_clients_fn = DEFAULT_CLIENT_INIT_FN
if dataset == 'ham10k':
fed_dataset = load_ham10k_federated(partitions=num_clients,
batch_size=batch_size,
mean=mean,
std=std)
initialize_clients_fn = initialize_ham10k_clients
else:
raise ValueError(f'dataset "{dataset}" unknown')
if not hasattr(max_value_criterion, '__iter__'):
max_value_criterion = [max_value_criterion]
if not hasattr(lr, '__iter__'):
lr = [lr]
for cf in client_fraction:
for lr_i in lr:
optimizer_args = OptimizerArgs(optim.SGD, lr=lr_i)
model_args = ModelArgs(MobileNetV2Lightning,
optimizer_args=optimizer_args,
num_classes=7)
fedavg_context = FedAvgExperimentContext(
name=name,
client_fraction=cf,
local_epochs=local_epochs,
lr=lr_i,
batch_size=batch_size,
optimizer_args=optimizer_args,
model_args=model_args,
train_args=train_args,
dataset_name=dataset)
experiment_specification = f'{fedavg_context}'
experiment_logger = create_tensorboard_logger(
fedavg_context.name, experiment_specification)
fedavg_context.experiment_logger = experiment_logger
for init_rounds, max_value in generate_configuration(
cluster_initialization_rounds, max_value_criterion):
# load the model state
round_model_state = load_fedavg_state(fedavg_context,
init_rounds)
server = FedAvgServer('initial_server',
fedavg_context.model_args,
fedavg_context)
server.overwrite_model_state(round_model_state)
logger.info('initializing clients ...')
clients = initialize_clients_fn(fedavg_context, fed_dataset,
server.model.state_dict())
overwrite_participants_models(round_model_state, clients)
# initialize the cluster configuration
round_configuration = {
'num_rounds_init': init_rounds,
'num_rounds_cluster': total_fedavg_rounds - init_rounds
}
if partitioner_class == DatadependentPartitioner:
clustering_dataset = load_femnist_colored_dataset(
str((REPO_ROOT / 'data').absolute()),
num_clients=num_clients,
batch_size=batch_size,
sample_threshold=sample_threshold)
dataloader = load_n_of_each_class(
clustering_dataset,
n=5,
tabu=list(fed_dataset.train_data_local_dict.keys()))
cluster_args = ClusterArgs(
partitioner_class,
linkage_mech=linkage_mech,
criterion=criterion,
dis_metric=dis_metric,
max_value_criterion=max_value,
plot_dendrogram=False,
reallocate_clients=reallocate_clients,
threshold_min_client_cluster=
threshold_min_client_cluster,
dataloader=dataloader,
**round_configuration)
else:
cluster_args = ClusterArgs(
partitioner_class,
linkage_mech=linkage_mech,
criterion=criterion,
dis_metric=dis_metric,
max_value_criterion=max_value,
plot_dendrogram=False,
reallocate_clients=reallocate_clients,
threshold_min_client_cluster=
threshold_min_client_cluster,
**round_configuration)
# create new logger for cluster experiment
experiment_specification = f'{fedavg_context}_{cluster_args}'
experiment_logger = create_tensorboard_logger(
fedavg_context.name, experiment_specification)
fedavg_context.experiment_logger = experiment_logger
initial_train_fn = partial(run_fedavg_train_round,
round_model_state,
training_args=train_cluster_args)
create_aggregator_fn = partial(FedAvgServer,
model_args=model_args,
context=fedavg_context)
federated_round_fn = partial(run_fedavg_round,
training_args=train_args,
client_fraction=cf)
after_post_clustering_evaluation = [
partial(log_after_round_evaluation, experiment_logger,
'post_clustering')
]
after_clustering_round_evaluation = [
partial(log_after_round_evaluation, experiment_logger)
]
after_federated_round_evaluation = [
partial(log_after_round_evaluation, experiment_logger,
['final hierarchical', global_tag])
]
after_clustering_fn = [
partial(log_cluster_distribution,
experiment_logger,
num_classes=fed_dataset.class_num),
partial(log_sample_images_from_each_client,
experiment_logger)
]
after_federated_round_fn = [
partial(
log_personalized_global_cluster_performance,
experiment_logger,
['final hierarchical personalized', global_tag_local],
local_evaluation_steps)
]
run_fedavg_hierarchical(
server,
clients,
cluster_args,
initial_train_fn,
federated_round_fn,
create_aggregator_fn,
after_post_clustering_evaluation,
after_clustering_round_evaluation,
after_federated_round_evaluation,
after_clustering_fn,
after_federated_round=after_federated_round_fn)
def clustering_test(mean, std, seed, lr, local_epochs, client_fraction,
optimizer_args, total_fedavg_rounds, batch_size,
num_clients, model_args, train_args, train_cluster_args,
initialization_rounds, partitioner_class, linkage_mech,
criterion, dis_metric, max_value_criterion):
fix_random_seeds(seed)
fed_dataset = load_ham10k_federated(partitions=num_clients,
batch_size=batch_size,
mean=mean,
std=std)
initialize_clients_fn = initialize_ham10k_clients
fedavg_context = FedAvgExperimentContext(name='ham10k_clustering',
client_fraction=client_fraction,
local_epochs=local_epochs,
lr=lr,
batch_size=batch_size,
optimizer_args=optimizer_args,
model_args=model_args,
train_args=train_args,
dataset_name='ham10k')
experiment_specification = f'{fedavg_context}'
experiment_logger = create_tensorboard_logger(fedavg_context.name,
experiment_specification)
log_dataset_distribution(experiment_logger, 'full dataset', fed_dataset)
server, clients = run_fedavg(context=fedavg_context,
num_rounds=total_fedavg_rounds,
dataset=fed_dataset,
save_states=True,
restore_state=True,
evaluate_rounds=False,
initialize_clients_fn=initialize_clients_fn)
for init_rounds in initialization_rounds:
# load the model state
round_model_state = load_fedavg_state(fedavg_context, init_rounds)
overwrite_participants_models(round_model_state, clients)
run_fedavg_train_round(round_model_state,
training_args=train_cluster_args,
participants=clients)
for max_value in max_value_criterion:
# initialize the cluster configuration
round_configuration = {
'num_rounds_init': init_rounds,
'num_rounds_cluster': total_fedavg_rounds - init_rounds
}
cluster_args = ClusterArgs(partitioner_class,
linkage_mech=linkage_mech,
criterion=criterion,
dis_metric=dis_metric,
max_value_criterion=max_value,
plot_dendrogram=False,
reallocate_clients=False,
threshold_min_client_cluster=-1,
**round_configuration)
experiment_logger = create_tensorboard_logger(
fedavg_context.name,
f'{experiment_specification}{cluster_args}')
partitioner = cluster_args()
cluster_clients_dic = partitioner.cluster(clients, server)
log_cluster_distribution(experiment_logger, cluster_clients_dic, 7)
def run_hierarchical_clustering_reptile(
seed,
name,
dataset,
num_clients,
batch_size,
num_label_limit,
use_colored_images,
sample_threshold,
hc_lr,
hc_cluster_initialization_rounds,
hc_client_fraction,
hc_local_epochs,
hc_train_args,
hc_partitioner_class,
hc_linkage_mech,
hc_criterion,
hc_dis_metric,
hc_max_value_criterion, # distance threshold
hc_reallocate_clients, #
hc_threshold_min_client_cluster, # only with hc_reallocate_clients = True,
# results in clusters having at least this number of clients
hc_train_cluster_args,
rp_sgd, # True -> Use SGD as inner optimizer; False -> Use Adam
rp_adam_betas, # Used only if sgd = False
rp_meta_batch_size,
rp_num_meta_steps,
rp_meta_learning_rate_initial,
rp_meta_learning_rate_final,
rp_eval_interval,
rp_inner_learning_rate,
rp_num_inner_steps,
rp_num_inner_steps_eval):
fix_random_seeds(seed)
global_tag = 'global_performance'
if dataset == 'femnist':
if use_colored_images:
fed_dataset = load_femnist_colored_dataset(
data_dir=str((REPO_ROOT / 'data').absolute()),
num_clients=num_clients,
batch_size=batch_size,
sample_threshold=sample_threshold)
else:
fed_dataset = load_femnist_dataset(
data_dir=str((REPO_ROOT / 'data').absolute()),
num_clients=num_clients,
batch_size=batch_size,
sample_threshold=sample_threshold)
if num_label_limit != -1:
fed_dataset = scratch_labels(fed_dataset, num_label_limit)
else:
raise ValueError(f'dataset "{dataset}" unknown')
if not hasattr(hc_max_value_criterion, '__iter__'):
hc_max_value_criterion = [hc_max_value_criterion]
if not hasattr(hc_lr, '__iter__'):
hc_lr = [hc_lr]
input_channels = 3 if use_colored_images else 1
data_distribution_logged = False
for cf in hc_client_fraction:
for lr_i in hc_lr:
# Initialize experiment context parameters
fedavg_optimizer_args = OptimizerArgs(optim.SGD, lr=lr_i)
fedavg_model_args = ModelArgs(CNNLightning,
optimizer_args=fedavg_optimizer_args,
input_channels=input_channels,
only_digits=False)
fedavg_context = FedAvgExperimentContext(
name=name,
client_fraction=cf,
local_epochs=hc_local_epochs,
lr=lr_i,
batch_size=batch_size,
optimizer_args=fedavg_optimizer_args,
model_args=fedavg_model_args,
train_args=hc_train_args,
dataset_name=dataset)
reptile_context = ReptileExperimentContext(
name=name,
dataset_name=dataset,
swap_labels=False,
num_classes_per_client=0,
num_shots_per_class=0,
seed=seed,
model_class=CNNLightning,
sgd=rp_sgd,
adam_betas=rp_adam_betas,
num_clients_train=num_clients,
num_clients_test=0,
meta_batch_size=rp_meta_batch_size,
num_meta_steps=rp_num_meta_steps,
meta_learning_rate_initial=rp_meta_learning_rate_initial,
meta_learning_rate_final=rp_meta_learning_rate_final,
eval_interval=rp_eval_interval,
num_eval_clients_training=-1,
do_final_evaluation=True,
num_eval_clients_final=-1,
inner_batch_size=batch_size,
inner_learning_rate=rp_inner_learning_rate,
num_inner_steps=rp_num_inner_steps,
num_inner_steps_eval=rp_num_inner_steps_eval)
experiment_specification = f'{fedavg_context}'
experiment_logger = create_tensorboard_logger(
name, experiment_specification)
if not data_distribution_logged:
log_dataset_distribution(experiment_logger, 'full dataset',
fed_dataset)
data_distribution_logged = True
log_after_round_evaluation_fns = [
partial(log_after_round_evaluation, experiment_logger,
'fedavg'),
partial(log_after_round_evaluation, experiment_logger,
global_tag)
]
server, clients = run_fedavg(
context=fedavg_context,
num_rounds=max(hc_cluster_initialization_rounds),
dataset=fed_dataset,
save_states=True,
restore_state=True,
after_round_evaluation=log_after_round_evaluation_fns)
for init_rounds, max_value in generate_configuration(
hc_cluster_initialization_rounds, hc_max_value_criterion):
# load the model state
round_model_state = load_fedavg_state(fedavg_context,
init_rounds)
overwrite_participants_models(round_model_state, clients)
# initialize the cluster configuration
round_configuration = {
'num_rounds_init': init_rounds,
'num_rounds_cluster': 0
}
cluster_args = ClusterArgs(
hc_partitioner_class,
linkage_mech=hc_linkage_mech,
criterion=hc_criterion,
dis_metric=hc_dis_metric,
max_value_criterion=max_value,
plot_dendrogram=False,
reallocate_clients=hc_reallocate_clients,
threshold_min_client_cluster=
hc_threshold_min_client_cluster,
**round_configuration)
# create new logger for cluster experiment
experiment_specification = f'{fedavg_context}_{cluster_args}_{reptile_context}'
experiment_logger = create_tensorboard_logger(
name, experiment_specification)
fedavg_context.experiment_logger = experiment_logger
initial_train_fn = partial(run_fedavg_train_round,
round_model_state,
training_args=hc_train_cluster_args)
create_aggregator_fn = partial(FedAvgServer,
model_args=fedavg_model_args,
context=fedavg_context)
# HIERARCHICAL CLUSTERING
logger.debug('starting local training before clustering.')
trained_participants = initial_train_fn(clients)
if len(trained_participants) != len(clients):
raise ValueError(
'not all clients successfully participated in the clustering round'
)
# Clustering of participants by model updates
partitioner = cluster_args()
cluster_clients_dic = partitioner.cluster(clients, server)
_cluster_clients_dic = dict()
for cluster_id, participants in cluster_clients_dic.items():
_cluster_clients_dic[cluster_id] = [
c._name for c in participants
]
log_cluster_distribution(experiment_logger,
cluster_clients_dic, 62)
# Initialize cluster models
cluster_server_dic = {}
for cluster_id, participants in cluster_clients_dic.items():
intermediate_cluster_server = create_aggregator_fn(
'cluster_server' + cluster_id)
intermediate_cluster_server.aggregate(participants)
cluster_server = ReptileServer(
participant_name=f'cluster_server{cluster_id}',
model_args=reptile_context.meta_model_args,
context=reptile_context,
initial_model_state=intermediate_cluster_server.model.
state_dict())
#create_aggregator_fn('cluster_server' + cluster_id)
#cluster_server.aggregate(participants)
cluster_server_dic[cluster_id] = cluster_server
# REPTILE TRAINING INSIDE CLUSTERS
after_round_evaluation = [log_after_round_evaluation]
RANDOM = random.Random(seed)
# Perform training
for i in range(reptile_context.num_meta_steps):
for cluster_id, participants in cluster_clients_dic.items(
):
if reptile_context.meta_batch_size == -1:
meta_batch = participants
else:
meta_batch = [
participants[k] for k in cyclerange(
start=i * reptile_context.meta_batch_size %
len(participants),
interval=reptile_context.meta_batch_size,
total_len=len(participants))
]
# Meta training step
reptile_train_step(
aggregator=cluster_server_dic[cluster_id],
participants=meta_batch,
inner_training_args=reptile_context.
get_inner_training_args(),
meta_training_args=reptile_context.
get_meta_training_args(
frac_done=i / reptile_context.num_meta_steps))
# Evaluation on train and test clients
if i % reptile_context.eval_interval == 0:
global_step = init_rounds + i
global_loss, global_acc = [], []
for cluster_id, participants in cluster_clients_dic.items(
):
# Test on all clients inside clusters
reptile_train_step(
aggregator=cluster_server_dic[cluster_id],
participants=participants,
inner_training_args=reptile_context.
get_inner_training_args(eval=True),
evaluation_mode=True)
result = evaluate_local_models(
participants=participants)
loss = result.get('test/loss')
acc = result.get('test/acc')
# Log
if after_round_evaluation is not None:
for c in after_round_evaluation:
c(experiment_logger,
f'cluster_{cluster_id}', loss, acc,
global_step)
loss_list = loss.tolist()
acc_list = acc.tolist()
global_loss.extend(loss_list if isinstance(
loss_list, list) else [loss_list])
global_acc.extend(acc_list if isinstance(
acc_list, list) else [acc_list])
if after_round_evaluation is not None:
for c in after_round_evaluation:
c(experiment_logger, 'mean_over_all_clients',
Tensor(global_loss), Tensor(global_acc),
global_step)
logger.info(f'Finished Reptile training round {i}')
# Final evaluation at end of training
if reptile_context.do_final_evaluation:
global_loss, global_acc = [], []
for cluster_id, participants in cluster_clients_dic.items(
):
# Final evaluation on train and test clients
# Test on all clients inside clusters
reptile_train_step(
aggregator=cluster_server_dic[cluster_id],
participants=participants,
inner_training_args=reptile_context.
get_inner_training_args(eval=True),
evaluation_mode=True)
result = evaluate_local_models(
participants=participants)
loss = result.get('test/loss')
acc = result.get('test/acc')
print(
f'Cluster {cluster_id} ({len(participants)} part.): loss = {loss}, acc = {acc}'
)
loss_list = loss.tolist()
acc_list = acc.tolist()
global_loss.extend(loss_list if isinstance(
loss_list, list) else [loss_list])
global_acc.extend(acc_list if isinstance(
acc_list, list) else [acc_list])
# Log
if after_round_evaluation is not None:
for c in after_round_evaluation:
c(experiment_logger, f'cluster_{cluster_id}',
loss, acc, reptile_context.num_meta_steps)
log_loss_and_acc('overall_mean', Tensor(global_loss),
Tensor(global_acc), experiment_logger, 0)