scheduler = lr_scheduler.MultiStepLR(optimizer,
milestones=[60, 120, 160],
gamma=0.2)
attacker = DDN(steps=args.steps, device=DEVICE)
max_loss = torch.log(torch.tensor(10.)).item() # for callback
best_acc = 0
best_epoch = 0
for epoch in range(args.epochs):
scheduler.step()
cudnn.benchmark = True
model.train()
requires_grad_(m, True)
accs = AverageMeter()
losses = AverageMeter()
attack_norms = AverageMeter()
length = len(train_loader)
for i, (images, labels) in enumerate(tqdm.tqdm(train_loader, ncols=80)):
images, labels = images.to(DEVICE), labels.to(DEVICE)
#原图loss
#logits_clean = model.forward(images)
logits_clean, feature_clean = model.forward(images)
#loss = F.cross_entropy(logits_clean, labels)
if args.adv is not None and epoch >= args.adv:
model.eval()
print('Loading data')
dataset = datasets.MNIST(args.data_path, train=False,
transform=transforms.ToTensor(),
download=True)
loader = data.DataLoader(dataset, shuffle=False, batch_size=16)
x, y = next(iter(loader))
x = x.to(device)
y = y.to(device)
print('Loading model')
model = SmallCNN()
model.load_state_dict(torch.load(args.model_path))
model.eval().to(device)
requires_grad_(model, False)
print('Running DDN attack')
attacker = DDN(steps=100, device=device)
start = time.time()
ddn_atk = attacker.attack(model, x, labels=y, targeted=False)
ddn_time = time.time() - start
print('Running C&W attack')
cwattacker = CarliniWagnerL2(device=device,
image_constraints=(0, 1),
num_classes=10)
start = time.time()
cw_atk = cwattacker.attack(model, x, labels=y, targeted=False)
cw_time = time.time() - start