args.devices = list(f"cuda:{i}"
for i in range(torch.cuda.device_count()))
else:
args.devices = ["cpu"]
else:
args.devices = list(name.strip() for name in args.devices.split(","))
# ---------------------------------------------------------------------------- #
# Serial preloading of the dataset
tools.success("Pre-downloading datasets...")
# Pre-load the datasets to prevent the first parallel runs from downloading them several times
with tools.Context("dataset", "info"):
for name in ("cifar10", ):
with tools.Context(name, "info"):
experiments.make_datasets(name, 1, 1)
# ---------------------------------------------------------------------------- #
# Run (missing) experiments
tools.success("Running experiments...")
# GAR to use
gars = ("krum", "median", "bulyan")
# Command maker helper
def make_command(params):
cmd = ["python3", "-OO", "attack.py"]
cmd += tools.dict_to_cmdlist(params)
return tools.Command(cmd)
# Defense
defense = aggregators.gars.get(args.gar)
if defense is None:
tools.fatal_unavailable(aggregators.gars, args.gar, what="aggregation rule")
# Attack
attack = attacks.attacks.get(args.attack)
if attack is None:
tools.fatal_unavailable(attacks.attacks, args.attack, what="attack")
# Model
model = experiments.Model(args.model, config, **args.model_args)
# Datasets
if args.no_transform:
train_transforms = test_transforms = torchvision.transforms.ToTensor()
else:
train_transforms = test_transforms = None # Let default values
trainset, testset = experiments.make_datasets(args.dataset, args.batch_size, args.batch_size_test, train_transforms=train_transforms, test_transforms=test_transforms, **args.dataset_args)
model.default("trainset", trainset)
model.default("testset", testset)
# Loss and criterion
loss = experiments.Loss(args.loss, **args.loss_args)
if args.l1_regularize is not None:
loss += args.l1_regularize * experiments.Loss("l1")
if args.l2_regularize is not None:
loss += args.l2_regularize * experiments.Loss("l2")
criterion = experiments.Criterion(args.criterion, **args.criterion_args)
model.default("loss", loss)
model.default("criterion", criterion)
# Optimizer
optimizer = experiments.Optimizer("sgd", model, lr=args.learning_rate, momentum=args.momentum, dampening=args.dampening, weight_decay=args.weight_decay)
model.default("optimizer", optimizer)
# Privacy noise distribution