def get_model(dataset, arch):
# arch2 = dict_arch2deephunter.get(arch,arch)
model = get_net(name=arch, dt_name=dataset)
if arch == "convmnist":
layer_names = ["0/relu1", "1/relu2", "2/relu3"]
num_layer = 3
elif arch == "convcifar10":
layer_names = [
"0/relu1", "1/relu2", "2/relu3", "3/relu4", "4/relu5", "5/relu6"
]
num_layer = 6
elif arch == "alexnet":
layer_names = [
'0/features/1', '1/features/4', '2/features/7', '3/features/9',
'4/features/11'
]
num_layer = 5
elif arch == "vgg":
# layer_names = ["5/relu6"]
layer_names = [
'0/relu1', '1/relu2', '2/relu3', '3/relu4', '4/relu5', '5/relu6',
'6/relu7', '7/relu8', '8/relu9', '9/relu10'
]
num_layer = 10
elif arch == "vgg16_bn":
# layer_names = ["5/relu6"]
layer_names = [
'0/features/2', '1/features/5', '2/features/9', '3/features/12',
'4/features/16', '5/features/19', '6/features/22', '7/features/26',
'8/features/29', '9/features/32'
]
num_layer = 10
return model, layer_names, num_layer
def test_get_net():
name_list = [
"convmnist",
# "convcifar10",
# "vgg",
# "alexnet",
# "alexnet",
]
for name in name_list:
model = get_net(name=name)
assert model is not None
def load_npc_lnpc(fileid, ):
file_info = gdrive_fileids.rq3_fileid_list[args_fid]
model_name = file_info["arch"]
mode = file_info["attack"]
dataset = file_info["dataset"]
sample_threshold, cluster_threshold, cluster_num = calc_sadl_utils_data.get_cluster_para(
dataset, model_name)
#print(model_name, mode)
intra = {}
layer_intra = {}
m_name = f"{model_name},{mode}"
# models[model_name]
nn_model = get_net(name=model_name)
# test_set = DatasetAdv(file_id_list[m_name])
test_set = dataloader.get_dataloader(fileid)
fetch_func = lambda x: x["your_adv"]
time_collect_list = []
for index, datax in enumerate(test_set):
covered_10 = set()
covered_100 = set()
total_100 = total_10 = 0
keys = datax["key"]
# print("keys:", keys)
x_test = datax["your_adv"]
y_test = torch.rand((x_test.shape[0]))
#print ("x_test:",x_test.mean(),x_test.std(),"y_test:",y_test.mean(),y_test.std())
test_loader1 = torch.torch.utils.data.DataLoader(
torch.utils.data.TensorDataset(x_test, y_test),
batch_size=BATCH_SIZE)
for step, (x, y) in enumerate(test_loader1):
# print("step", step)
x = x.to(device)
# models[model_name] = models[model_name].to(device)
# cover = Coverager(models[model_name], model_name, cluster_threshold, num_classes=num_classes, num_cluster=cluster_num)
cover = Coverager(nn_model,
model_name,
cluster_threshold,
num_classes=num_classes,
num_cluster=cluster_num)
#print ("model_name",model_name,"cluster_threshold",cluster_threshold,"num_classes",num_classes,"clust",cluster_num)
start_time1 = time.time()
covered1, total1 = cover.Intra_NPC(x,
y,
bucket_m,
sample_threshold,
mode=mode,
simi_soft=False,
arc=model_name)
start_time2 = time.time()
covered2, total2 = cover.Layer_Intra_NPC(x,
y,
bucket_m,
sample_threshold,
mode=mode,
simi_soft=False,
useOldPaths_X=True,
arc=model_name)
start_time3 = time.time()
total_10 += total1
total_100 += total2
covered_10 = covered_10 | covered1
covered_100 = covered_100 | covered2
time_collect_list.append((start_time1, start_time2, start_time3))
# print(cover.get_simi(x, y, bucket_m, single_threshold, mode=mode, simi_soft=False))
intra[keys] = round(len(covered_10) / total1, 5)
layer_intra[keys] = round(len(covered_100) / total2, 5)
#print(m_name, intra[keys],"<---intra")
#print(m_name, layer_intra[keys],"<---layer_intra")
return intra, layer_intra, time_collect_list
])
dataset = torchvision.datasets.MNIST(root='~/.torch/',
train=args.data_train,
download=True,
transform=transform_test)
# data_loader = torch.utils.data.DataLoader(
# testset, batch_size=batch_size, shuffle=False)
else:
pass
if args.dataset == "mnist":
# ori_model = ConvnetMnist()
# ori_model.load_state_dict(torch.load("trained_models/mnist_mixup_acc_99.28_ckpt.pth")["net"])
# net = mask_ConvnetMnist()
# net.load_state_dict(torch.load("trained_models/mnist_mixup_acc_99.28_ckpt.pth")["net"])
ori_model = get_net(name="mnist")
net = get_masked_net(name="mnist")
ori_model = ori_model.cuda()
ori_model.eval()
net = net.cuda()
net.eval()
grids = args.grids
right_count = [0 for i in range(grids)]
prob_count = [0 for i in range(grids)]
remedy_count = [0 for i in range(grids)]
num = [0 for i in range(grids)]
results = []