def main():
#Dataset
dataset = ImageNet(DATASET_PATH)
#Model
model, checkpoint = model_utils.make_and_restore_model(
arch='vgg19', dataset=dataset, resume_path=CHECKPOINT_PATH)
#DataLoader
train_loader, val_loader = dataset.make_loaders(batch_size=BATCH_SIZE,
workers=NUM_WORKERS)
if 'module' in dir(model): model = model.module
# Create a cox store for logging
out_store = cox.store.Store(OUT_DIR)
# Hard-coded base parameters
train_kwargs = {
'out_dir': "train_out",
'adv_train': 1,
'constraint': '2',
'eps': 3.0,
'attack_lr': 0.5,
'attack_steps': 7,
'save_ckpt_iters': 5
}
train_args = Parameters(train_kwargs)
# Fill whatever parameters are missing from the defaults
train_args = defaults.check_and_fill_args(train_args,
defaults.TRAINING_ARGS, ImageNet)
train_args = defaults.check_and_fill_args(train_args, defaults.PGD_ARGS,
ImageNet)
# Train a model
train.train_model(train_args,
model, (train_loader, val_loader),
store=out_store,
checkpoint=checkpoint)
from cox.utils import Parameters
from robustness.main import setup_args, setup_store_with_metadata
EPS = 1.0
eps_str = f'{EPS}'.replace('.', 'p')
exp_name = f'fast_adv_l2_{eps_str}'
train_kwargs = {
'out_dir': "logs",
'exp_name': exp_name,
'dataset': 'cifar',
'data': 'data/cifar10',
'arch': 'resnet50',
'adv_train': 1,
'constraint': '2',
'eps': EPS,
'attack_steps': 20,
'attack_lr': 2.5 * EPS / 20,
'fast_attack_lr': 1.5 * EPS
}
train_args = Parameters(train_kwargs)
# Fill whatever parameters are missing from the defaults
train_args = setup_args(train_args)
print(train_args)
store = setup_store_with_metadata(train_args)
final_model = main(train_args, store=store)
exp_name = f'fast_adv_l2_{eps_str}'
for e in [0, 0.25, 0.5, 1.0, 2.0]:
e_str = f'{e}'.replace('.', 'p')
eval_exp_name = f'fast_adv_l2_{eps_str}_{e_str}'
eval_kwargs = {
'eval_only': 1,
'resume': os.path.join('logs', exp_name, 'checkpoint.pt.best'),
'out_dir': "logs_eval",
'exp_name': eval_exp_name,
'dataset': 'cifar',
'data': 'data/cifar10',
'arch': 'resnet50',
'adv_train': 0,
'adv_eval': 1,
'constraint': '2',
'eps': e,
'attack_steps': 20,
'attack_lr': 2.5 * e / 20,
'fast_attack_lr': 1.5 * e
}
train_args = Parameters(eval_kwargs)
# Fill whatever parameters are missing from the defaults
train_args = setup_args(train_args)
print(train_args)
store = setup_store_with_metadata(train_args)
final_model = main(train_args, store=store)
m, _ = model_utils.make_and_restore_model(arch='resnet50', dataset=ds)
train_loader, val_loader = ds.make_loaders(batch_size=128, workers=8)
# train_loader = LambdaLoader(train_loader, label_permutate)
# val_loader = LambdaLoader(val_loader, label_permutate)
# Create a cox store for logging
out_store = cox.store.Store('coxx')
# Hard-coded base parameters
train_kwargs = {
'out_dir': "train_out",
'adv_train': 0,
'lr': .01,
'epochs': 300,
'save-ckpt-iters': 10
}
train_args = Parameters(train_kwargs)
idx = np.arange(label_dim)
order = np.load('../data/rnd_label_c10_5.npy')[idx].T
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())