def main():
name = 'trades'
# set up the experiment
args, logger, device, \
(train_loader, test_loader), \
(model_factory, optimizer_factory, scheduler_factory) \
= init_experiment(args_factory=get_args, name=name)
model = model_factory()
opt = optimizer_factory(model)
scheduler = scheduler_factory(opt)
### training adversary config
assert args.geometry == 'linf', \
'l2 adversary not supported yet'
std = STD.to(device)
upper_limit = UPPER_LIMIT.to(device)
lower_limit = LOWER_LIMIT.to(device)
epsilon = (args.epsilon / 255.) / std
# parameters for PGD training
pgd_train_kwargs = dict(reg_weight=args.reg_weight,
geometry=args.geometry,
epsilon=epsilon,
alpha=(args.alpha / 255.) / std,
lower_limit=lower_limit,
upper_limit=upper_limit,
attack_iters=args.attack_iters,
criterion=nn.CrossEntropyLoss())
# parameters for PGD early stopping
pgd_kwargs = dict(epsilon=epsilon,
alpha=(2 / 255.) / std,
lower_limit=lower_limit,
upper_limit=upper_limit,
attack_iters=5,
restarts=1)
# training
model, best_state_dict = fit(
step=functools.partial(train_step, **pgd_train_kwargs),
epochs=args.epochs,
model=model,
optimizer=opt,
scheduler=scheduler,
data_loader=train_loader,
model_path=os.path.join(args.out_dir,
f'model_preact_resnet18_{name}.pt'),
logger=logger,
early_stop=args.early_stop,
pgd_kwargs=pgd_kwargs)
# eval
if not args.no_eval:
model_test = model_factory()
model_test.load_state_dict(best_state_dict)
evaluate(model=model_test,
test_loader=test_loader,
upper_limit=upper_limit,
lower_limit=lower_limit,
verbose=args.no_verbose)
def main():
name = 'distilled'
# set up the experiment
args, logger, device, \
(train_loader, test_loader), \
(model_factory, optimizer_factory, scheduler_factory) \
= init_experiment(args_factory=get_args, name=name)
assert not args.early_stop, \
'Model distillation is non-robust, do not early stop wrt adversarial accuracy.'
if args.pretrained_model is None:
logger.info('No pretrained model specified... pretraining model via `train_standard.py`.')
process = subprocess.run([
'python', 'train_standard.py',
'--epochs', str(args.epochs),
'--no-eval',
'--temperature', str(args.softmax_temperature)])
pretrained_model_path = os.path.join(args.out_dir, 'model_preact_resnet18_standard.pt')
else:
logger.info(f'Pretrained model specified as `{args.pretrained_model}`... skipping pretraining.')
pretrained_model_path = args.pretrained_model
pretrained_model = load_model(
path =pretrained_model_path,
model_factory=model_factory,
device =device)
model = model_factory()
opt = optimizer_factory(model)
scheduler = scheduler_factory(opt)
### train model
model, best_state_dict = fit(
step =functools.partial(train_step,
pretrained_model =pretrained_model,
softmax_temperature=args.softmax_temperature),
epochs =args.epochs,
model =model,
optimizer =opt,
scheduler =scheduler,
data_loader=train_loader,
model_path =os.path.join(args.out_dir, f'model_preact_resnet18_{name}.pt'),
logger =logger,
early_stop =args.early_stop)
### evaluate model
if not args.no_eval:
model_test = model_factory()
model_test.load_state_dict(best_state_dict)
evaluate(
model =model_test,
test_loader=test_loader,
upper_limit=UPPER_LIMIT,
lower_limit=LOWER_LIMIT,
verbose =args.no_verbose)
def main():
name = 'label_smoothing'
# set up the experiment
args, logger, device, \
(train_loader, test_loader), \
(model_factory, optimizer_factory, scheduler_factory) \
= init_experiment(args_factory=get_args, name=name)
assert not args.early_stop, \
'Model distillation is non-robust, do not early stop wrt adversarial accuracy.'
model = model_factory()
opt = optimizer_factory(model)
scheduler = scheduler_factory(opt)
model, best_state_dict = fit(
step=functools.partial(train_step, smoothing=args.smoothing),
epochs=args.epochs,
model=model,
optimizer=opt,
scheduler=scheduler,
data_loader=train_loader,
model_path=os.path.join(args.out_dir,
f'model_preact_resnet18_{name}.pt'),
logger=logger,
early_stop=args.early_stop)
if not args.no_eval:
model_test = model_factory()
model_test.load_state_dict(best_state_dict)
evaluate(model=model_test,
test_loader=test_loader,
upper_limit=UPPER_LIMIT,
lower_limit=LOWER_LIMIT,
verbose=args.no_verbose)