if args.lr_schedule is None:
args.lr_schedule = '30,60,80'
if args.num_epochs is None:
args.num_epochs = 90
torch.manual_seed(args.seed)
np.random.seed(args.seed)
random.seed(args.seed)
dataset, model = get_dataset_model(args)
if isinstance(model, FeatureModel):
model.allow_train()
if torch.cuda.is_available():
model.cuda()
vae = get_vae_model(args.vae_dir, 32, 2048)
l2_distance = L2Distance()
linf_distance = LinfDistance()
l2_distance.cuda()
linf_distance.cuda()
train_loader, val_loader = dataset.make_loaders(
workers=4, batch_size=args.batch_size)
attacks = [eval(attack_str) for attack_str in args.attack]
validation_attacks = [
NoAttack(),
LinfAttack(model, dataset_name=args.dataset,
num_iterations=VAL_ITERS),
L2Attack(model, dataset_name=args.dataset,
num_iterations=VAL_ITERS),
# CIAttack2(model, vae, num_iterations=100, norm='linf',eps_max=8/255, loss='margin')
help='number of batches (default entire dataset)')
parser.add_argument('--per_example',
action='store_true',
default=False,
help='output per-example accuracy')
parser.add_argument('--output', type=str, help='output CSV')
parser.add_argument('attacks',
metavar='attack',
type=str,
nargs='+',
help='attack names')
args = parser.parse_args()
dist_models: List[Tuple[str, nn.Module]] = [
('l2', L2Distance()),
('linf', LinfDistance()),
('ssim', SSIM()),
]
dataset, model = get_dataset_model(args)
if not isinstance(model, FeatureModel):
raise TypeError('model must be a FeatureModel')
dist_models.append(('lpips_self', LPIPSDistance(model)))
alexnet_model_name: Literal['alexnet_cifar', 'alexnet']
if args.dataset.startswith('cifar'):
alexnet_model_name = 'alexnet_cifar'
else:
alexnet_model_name = 'alexnet'
dist_models.append((