shuffle=True,
**kwargs)
testset = torchvision.datasets.CIFAR10(root='./data',
train=False,
download=True,
transform=transform_test)
test_loader = torch.utils.data.DataLoader(testset,
batch_size=args.test_batch_size,
shuffle=False,
**kwargs)
testloader = torch.utils.data.DataLoader(testset,
batch_size=args.test_size,
shuffle=True,
**kwargs)
# testloader = torch.utils.data.DataLoader(testset, batch_size=len(testset), shuffle=False, **kwargs)
loss_fn = nn.CrossEntropyLoss()
linear, conv, bn = True, True, False # not to regularize BN, since it has only little effect on BN norms
mmt = 0.5
betas = [0.00001, 0.0001, 0.001, 0.01, 0.1]
for beta in betas:
name = f'{args.model}_inf_nd_proj_bn_{beta}'
print(f'Using {name}')
recorder = Recorder(name, img_dir)
main()
print()
def main():
# init model, ResNet18() can be also used here for training
if args.loss == 'alp' or args.loss == 'trades' or args.loss == 'pgd':
print("normalize False")
model = nets[args.model]().to(device)
else:
print("normalize True")
model = nets[args.model](use_FNandWN=True).to(device)
bind_lip(model, norm='1-norm', verbose=False)
recorder = Recorder(f'{name}')
optimizer = optim.SGD(model.parameters(),
lr=args.lr,
momentum=args.momentum,
weight_decay=args.weight_decay)
for epoch in range(1, args.epochs + 1):
# adjust learning rate for SGD
print(f'Epoch: {epoch:3d}', end=' ')
adjust_learning_rate(optimizer, epoch)
# adversarial training
train(args, model, device, train_loader, optimizer, recorder, epoch)
# evaluation on natural examples
# print('==============')
# eval_train(model, device, train_loader, recorder)
eval_test(model, device, test_loader, recorder)
# print('==============')
# save checkpoint
if (epoch >= args.start_freq) and (epoch % args.save_freq == 0):
torch.save(model.state_dict(),
os.path.join(model_dir, f'{name}-epoch{epoch}.pt'))
recorder.step()
torch.save(model.state_dict(), os.path.join(model_dir, f'{name}.pt'))
with open(f'{log_dir}/{name}_record.pkl', 'wb') as file:
pickle.dump(recorder, file)
recorder.draw('lip_sum')
recorder.draw_many('lip')
recorder.draw('train_acc')
recorder.draw('test_acc')
adversary = AutoAttack(model,
norm='Linf',
eps=8 / 255,
version='standard',
verbose=False)
adversary.attacks_to_run = ['apgd-ce']
model.eval()
with torch.no_grad():
for batch_idx, (inputs, targets) in enumerate(testloader):
inputs, targets = inputs.to(device), targets.to(device)
# print(inputs.max(), inputs.min())
x_adv, robust_accuracy = adversary.run_standard_evaluation(inputs,
targets,
bs=128)
print(f'robust_accuracy: {robust_accuracy}')
break
recorder.record('robust_accuracy', robust_accuracy)
with open(f'{log_dir}/{name}_record.pkl', 'wb') as file:
pickle.dump(recorder, file)
def main():
model = nets[args.model]().to(device)
bind_lip(model, norm='1-norm', verbose=False)
recorder = Recorder(f'{name}', img_dir)
optimizer = optim.SGD(model.parameters(),
lr=args.lr,
momentum=args.momentum,
weight_decay=args.weight_decay)
for epoch in range(1, args.epochs + 1):
# adjust learning rate for SGD
print(f'Epoch: {epoch:3d}', end=' ')
adjust_learning_rate(optimizer, epoch)
train(args, model, device, train_loader, optimizer, recorder, epoch)
eval_test(model, device, test_loader, recorder)
# save checkpoint
# if (epoch >= args.start_freq) and (epoch % args.save_freq == 0):
# torch.save(model.state_dict(),
# os.path.join(model_dir, f'{name}-epoch{epoch}.pt'))
recorder.step()
torch.save(model.state_dict(), os.path.join(model_dir, f'{name}.pt'))
with open(f'{log_dir}/{name}_record.pkl', 'wb') as file:
pickle.dump(recorder, file)
# recorder.draw('lip_sum')
recorder.draw_many('lip')
recorder.draw('train_acc')
recorder.draw('test_acc')
adversary = AutoAttack(model,
norm='Linf',
eps=8 / 255,
version='standard',
verbose=True)
adversary.attacks_to_run = ['apgd-ce', 'apgd-t']
model.eval()
with torch.no_grad():
for batch_idx, (inputs, targets) in enumerate(testloader):
inputs, targets = inputs.to(device), targets.to(device)
# print(inputs.max(), inputs.min())
x_adv, robust_accuracy = adversary.run_standard_evaluation(inputs,
targets,
bs=128)
print(f'robust_accuracy: {robust_accuracy}')
break
recorder.record('robust_accuracy', robust_accuracy)
with open(f'{log_dir}/{name}_record.pkl', 'wb') as file:
pickle.dump(recorder, file)
clean_acc = recorder.test_acc[-1][1]
write2text(log_dir,
'log',
name,
clean_acc=clean_acc,
robust_accuracy=robust_accuracy)