def code_test():
from base_model.cifar_resnet18 import cifar_resnet18
from dataset import Dataset
from my_snip.base import EpochDataset
ds_val = Dataset(dataset_name = 'val')
ds_val.load()
ds_val = EpochDataset(ds_val)
model_path = '../exps/exp0/checkpoint.pth.tar'
net = cifar_resnet18()
checkpoint = torch.load(model_path)
net.load_state_dict(checkpoint['state_dict'])
net.cuda()
epoch = next(ds_val.epoch_generator())
evalRoboustness(net, epoch)
def test_IPGD():
model_path = './exps/exp0/model_best.pth.tar'
from base_model.cifar_resnet18 import cifar_resnet18
from dataset import Dataset
from my_snip.base import EpochDataset
pgd = IPGD()
net = cifar_resnet18()
net.load_state_dict(torch.load(model_path)['state_dict'])
net.cuda()
ds_train = Dataset(dataset_name='train')
ds_train.load()
ds_train = EpochDataset(ds_train)
epoch = next(ds_train.epoch_generator())
count = 0
for mn_batch in epoch:
count += 1
if count > 100:
break
data = mn_batch['data']
label = mn_batch['label']
#print(data.shape)
data = torch.tensor(data, dtype=torch.float32).cuda()
label = torch.tensor(label, dtype=torch.int64).cuda()
acc = pgd.get_batch_accuracy(net, data, label)
print('acc: {:.2f}%'.format(acc))
[100, 0.01],
[100, 0.001]
]
get_learing_rate = MultiStageLearningRatePolicy(learning_rate_policy)
def adjust_learning_rate(optimzier, epoch):
#global get_lea
lr = get_learing_rate(epoch)
for param_group in optimizer.param_groups:
param_group['lr'] = lr
torch.backends.cudnn.benchmark = True
ds_train = Dataset(dataset_name = 'train')
ds_train.load()
ds_train = EpochDataset(ds_train)
ds_val = Dataset(dataset_name = 'val')
ds_val.load()
ds_val = EpochDataset(ds_val)
net = cifar_resnet18()
net.cuda()
criterion = nn.CrossEntropyLoss().cuda()
optimizer = optim.SGD(net.parameters(), lr = get_learing_rate(0), momentum = 0.9, weight_decay=args.weight_decay)
PgdAttack = IPGD(eps = args.eps, sigma = args.eps // 2, nb_iter = args.iter, norm = np.inf)
best_prec = 0.0
if args.resume: