def run():
experiments_raw = json.loads(args.hp)
hp_dicts = [
hp for x in experiments_raw for hp in xpm.get_all_hp_combinations(x)
][args.start:args.end]
if args.reverse_order:
hp_dicts = hp_dicts[::-1]
experiments = [xpm.Experiment(hyperparameters=hp) for hp in hp_dicts]
print("Running {} Experiments..\n".format(len(experiments)))
for xp_count, experiment in enumerate(experiments):
run_experiment(experiment, xp_count, len(experiments))
def run():
with open('federated_learning.json') as data_file:
experiments_raw = json.load(data_file)[args.schedule]
hp_dicts = [
hp for x in experiments_raw for hp in xpm.get_all_hp_combinations(x)
][args.start:args.end]
if args.reverse_order:
hp_dicts = hp_dicts[::-1]
experiments = [xpm.Experiment(hyperparameters=hp) for hp in hp_dicts]
print("Running {} Experiments..\n".format(len(experiments)))
for xp_count, experiment in enumerate(experiments):
run_experiment(experiment, xp_count, len(experiments))
print("Torch Version: ", torch.__version__)
device = 'cuda' if torch.cuda.is_available() else 'cpu'
args = parser.parse_args()
# Load the Hyperparameters of all Experiments to be performed and set up the Experiments
with open('federated_learning.json') as data_file:
experiments_raw = json.load(data_file)[args.schedule]
hp_dicts = [
hp for x in experiments_raw for hp in xpm.get_all_hp_combinations(x)
][args.start:args.end]
if args.reverse_order:
hp_dicts = hp_dicts[::-1]
experiments = [xpm.Experiment(hyperparameters=hp) for hp in hp_dicts]
def run_experiments(experiments):
print("Running {} Experiments..\n".format(len(experiments)))
for xp_count, xp in enumerate(experiments):
hp = dhp.get_hp(xp.hyperparameters)
xp.prepare(hp)
print(xp)
# Load the Data and split it among the Clients
client_loaders, train_loader, test_loader, stats = data_utils.get_data_loaders(
hp)
# Instantiate Clients and Server with Neural Net
net = getattr(neural_nets, hp['net'])
def experiment_wrapper(cfg, seed, instance, budget, **kwargs):
model_name = hashlib.sha1(
json.dumps(cfg._values, sort_keys=True).encode()).hexdigest()
files = glob.glob(f"{args.CHECKPOINT_PATH}/*-{model_name}.pth")
if files:
oldbudget = int(files[0].split('/')[-1].split('-')[0])
else:
oldbudget = 0
budget = max(0, int(budget) - oldbudget)
hp = {
"dataset":
"mnist",
"distill_dataset":
"emnist",
"net":
"lenet_mnist",
"n_clients":
20,
"classes_per_client":
0.01,
"communication_rounds":
budget,
"participation_rate":
0.4,
"local_epochs":
cfg["local_epochs"],
"distill_epochs":
cfg["distill_epochs"],
"n_distill":
100000,
"local_optimizer": [
cfg["local_optimizer"], {
"lr": cfg[f"{cfg['local_optimizer'].lower()}_lr"]
}
],
"distill_optimizer": ["Adam", {
"lr": 0.001
}],
"fallback":
cfg["fallback"],
"lambda_outlier":
cfg["lambda_outlier"],
"lambda_fedprox":
cfg["lambda_fedprox"],
"only_train_final_outlier_layer":
False,
"warmup_type":
"constant",
"mixture_coefficients": {
"base": cfg["mixture_coefficients_base"],
"public": 1 - cfg["mixture_coefficients_base"]
},
"batch_size":
256,
"distill_mode":
"logits_weighted_with_deep_outlier_score",
"aggregation_mode":
"FAD+P+S",
"pretrained":
f"{oldbudget}-{model_name}.pth",
"save_model":
f"{budget}-{model_name}.pth",
"log_frequency":
-100,
"log_path":
"trash/",
"job_id": ['']
}
# load already trained model if available
#hp.update(kwargs)
with warnings.catch_warnings():
warnings.filterwarnings('ignore')
experiment = xpm.Experiment(hyperparameters=hp)
return run_experiment(experiment, 0, 1, cmdargs, seed=seed)