def args_preprocess(args):
if args.adv_train and eval(args.eps) == 0:
print('[Switching to standard training since eps = 0]')
args.adv_train = 0
if args.pytorch_pretrained:
assert not args.model_path, 'You can either specify pytorch_pretrained or model_path, not together.'
## CIFAR10 to CIFAR10 assertions
if args.cifar10_cifar10:
assert args.dataset == 'cifar10'
if args.data != '':
cs.PETS_PATH = cs.CARS_PATH = cs.FGVC_PATH = cs.FLOWERS_PATH = cs.DTD_PATH = cs.SUN_PATH = cs.FOOD_PATH = cs.BIRDS_PATH = args.data
ALL_DS = list(transfer_datasets.DS_TO_FUNC.keys()) + [
'imagenet', 'breeds_living_9', 'stylized_imagenet'
]
assert args.dataset in ALL_DS
# Important for automatic job retries on the cluster in case of premptions. Avoid uuids.
assert args.exp_name != None
# Preprocess args
args = defaults.check_and_fill_args(args, defaults.CONFIG_ARGS, None)
if not args.eval_only:
args = defaults.check_and_fill_args(args, defaults.TRAINING_ARGS, None)
if args.adv_train or args.adv_eval:
args = defaults.check_and_fill_args(args, defaults.PGD_ARGS, None)
args = defaults.check_and_fill_args(args, defaults.MODEL_LOADER_ARGS, None)
return args
def setup_args(args):
'''
Fill the args object with reasonable defaults from
:mod:`robustness.defaults`, and also perform a sanity check to make sure no
args are missing.
'''
# override non-None values with optional config_path
args.adv_train = (args.classifier_loss == 'robust') or \
(args.estimator_loss == 'worst')
if args.config_path:
args = cox.utils.override_json(args, args.config_path)
ds_class = DATASETS[args.dataset]
args = check_and_fill_args(args, defaults.CONFIG_ARGS, ds_class)
if not args.eval_only:
args = check_and_fill_args(args, defaults.TRAINING_ARGS, ds_class)
if args.adv_train or args.adv_eval:
args = check_and_fill_args(args, defaults.PGD_ARGS, ds_class)
args = check_and_fill_args(args, defaults.MODEL_LOADER_ARGS, ds_class)
if args.eval_only: assert args.resume is not None, \
"Must provide a resume path if only evaluating"
return args
out_store = cox.store.Store(args.out_dir)
# Hard-coded base parameters
train_kwargs = {
'out_dir': args.out_dir,
'adv_train': 0,
'constraint': GaussianNoise,
'eps': args.eps,
'attack_lr': args.eps,
'lr': args.lr,
'attack_steps': 1,
'step_lr': 2000,
'random_start': 1,
'use_best': False,
'epochs': args.epochs,
'save_ckpt_iters': -1, # best and last
'eps_fadein_epochs': args.fade_in
}
train_args = Parameters(train_kwargs)
# Fill whatever parameters are missing from the defaults
# use Imagnet defaults
ds_class = datasets.DATASETS['imagenet']
train_args = defaults.check_and_fill_args(train_args,
defaults.TRAINING_ARGS, ds_class)
train_args = defaults.check_and_fill_args(train_args,
defaults.PGD_ARGS, ds_class)
# Train a model
train.train_model(train_args, model, (train_loader, val_loader), store=out_store)
for k in new_args), set(new_args.keys()) - set(vars(args).keys())
for k in new_args:
setattr(args, k, new_args[k])
assert not args.adv_train, 'not supported yet slatta dog'
assert args.training_mode is not None, "training_mode is required"
# Important for automatic job retries on the cluster in case of premptions. Avoid uuids.
if args.exp_name == 'random':
args.exp_name = str(uuid4())
print(f"Experiment name: {args.exp_name}")
assert args.exp_name != None
# Preprocess args
default_ds = args.dataset if args.dataset in datasets.DATASETS else "imagenet"
args = defaults.check_and_fill_args(args, defaults.CONFIG_ARGS,
datasets.DATASETS[default_ds])
if not args.eval_only:
args = defaults.check_and_fill_args(args, defaults.TRAINING_ARGS,
datasets.DATASETS[default_ds])
if False and (args.adv_train or args.adv_eval):
args = defaults.check_and_fill_args(args, defaults.PGD_ARGS,
datasets.DATASETS[default_ds])
args = defaults.check_and_fill_args(args, defaults.MODEL_LOADER_ARGS,
datasets.DATASETS[default_ds])
store = cox.store.Store(args.out_dir, args.exp_name)
if 'metadata' not in store.keys:
args_dict = args.__dict__
schema = cox.store.schema_from_dict(args_dict)
store.add_table('metadata', schema)
store['metadata'].append_row(args_dict)
loss.backward()
adv = self.clip(adv + step_size * torch.sign(adv.grad.data), inp,
eps) # gradient ASCENT
return adv.clone().detach()
ds = CIFAR('/scratch/raunakc/datasets/cifar10')
model, _ = make_and_restore_model(arch='resnet18', dataset=ds)
model.attacker = WhiteboxPGD(model.model, ds)
train_kwargs = {
'dataset': 'cifar',
'arch': 'resnet',
'out_dir': "train_out",
'adv_train': 1,
'adv_eval': 1,
'eps': 8 / 255,
'attack_lr': 2 / 255,
'attack_steps': 10,
'constraint': 'inf' # not required but arg checker requires it :(
}
args = utils.Parameters(train_kwargs)
args = check_and_fill_args(args, defaults.TRAINING_ARGS, ds.__class__)
if args.adv_train or args.adv_eval:
args = check_and_fill_args(args, defaults.PGD_ARGS, ds.__class__)
args = check_and_fill_args(args, defaults.MODEL_LOADER_ARGS, ds.__class__)
train_loader, val_loader = ds.make_loaders(batch_size=128, workers=8)
train.train_model(args, model, (train_loader, val_loader))
if args.adv_train and args.eps == 0:
print('[Switching to standard training since eps = 0]')
args.adv_train = 0
assert not args.adv_train, 'not supported yet slatta dog'
# Important for automatic job retries on the cluster in case of premptions. Avoid uuids.
assert args.exp_name != None
# Useful for evaluation QRCodes since they are not robust to occlustions at all
if args.no_translation:
constants.PATCH_TRANSFORMS['translate'] = (0., 0.)
# Preprocess args
args = defaults.check_and_fill_args(
args, defaults.CONFIG_ARGS, datasets.DATASETS[args.dataset])
args = defaults.check_and_fill_args(
args, defaults.MODEL_LOADER_ARGS, datasets.DATASETS[args.dataset])
store = cox.store.Store(args.out_dir, args.exp_name)
if args.args_from_store:
args_from_store = args.args_from_store.split(',')
df = store['metadata'].df
print(f'==>[Reading from existing store in {store.path}]')
for a in args_from_store:
if a not in df:
raise Exception(f'Did not find {a} in the store {store.path}')
setattr(args,a, df[a][0])
print(f'==>[Read {a} = ({getattr(args, a)}) ')
if 'metadata_eval' not in store.keys:
train_crit = torch.nn.BCELoss()
def custom_train_loss(logits, targ):
if torch.cuda.is_available():
targets = torch.from_numpy(order[targ.cpu().numpy()]).cuda()
else:
targets = torch.from_numpy(order[targ.numpy()])
outputs = torch.sigmoid(logits.float())
return train_crit(outputs.float(), targets.float())
train_args.custom_train_loss = custom_train_loss
# Fill whatever parameters are missing from the defaults
train_args = defaults.check_and_fill_args(train_args, defaults.TRAINING_ARGS,
CustomCIFAR)
# train_args = defaults.check_and_fill_args(train_args,
# defaults.PGD_ARGS, CIFAR)
# Train a model
train.train_model(train_args, m, (train_loader, val_loader), store=out_store)
# from robustness.datasets import DATASETS
# from robustness.model_utils import make_and_restore_model
# from robustness.train import train_model
# from robustness.defaults import check_and_fill_args
# from robustness.tools import constants, helpers
# from robustness import defaults
# from cox import utils
# from cox import store
'eps_fadein_epochs': 0,
'random_restarts': 0,
'lr': parsed_args.start_lr,
'use_adv_eval_criteria': 1,
'regularizer': regularizer,
'let_reg_handle_loss': True
}
ds_class = DATASETS[train_kwargs['dataset']]
train_args = cox.utils.Parameters(train_kwargs)
dx = utils.CIFAR10()
dataset = dx.get_dataset()
args = check_and_fill_args(train_args, defaults.TRAINING_ARGS, ds_class)
args = check_and_fill_args(train_args, defaults.MODEL_LOADER_ARGS,
ds_class)
model, _ = make_and_restore_model(arch='vgg19', dataset=dataset)
# Make the data loaders
train_loader, val_loader = dataset.make_loaders(args.workers,
args.batch_size,
data_aug=bool(
args.data_aug))
# Prefetches data to improve performance
train_loader = helpers.DataPrefetcher(train_loader)
val_loader = helpers.DataPrefetcher(val_loader)
