def __init__(self,
model_name="vgg16",
MODEL_DIR="",
EPOCHS=[],
P=[1, 2, 3],
SUBBANDS=["d", "h", "v"],
IMAGE_SIZE=32,
N_CHANELS=3,
use_cuda=True,
is_prefilt=False):
super(MultiChannel, self).__init__()
if model_name == "vgg16":
net = modelvgg.VGG('VGG16')
elif model_name == "resnet18":
net = modelresnet.ResNet18()
self.image_size = IMAGE_SIZE
self.n_channel = N_CHANELS
self.use_cuda = use_cuda
self.is_prefilt = is_prefilt
self.epochs = EPOCHS
self.p = P
self.subbands = SUBBANDS
self.NET = []
self.permutation = []
i = -1
for p in P:
for sb in SUBBANDS:
i += 1
# --- load model ----
FROM_DIR = MODEL_DIR % (sb, p)
net.load_state_dict(
torch.load(FROM_DIR + "/%s_%d.pth" %
(model_name, EPOCHS[i]),
map_location=lambda storage, loc: storage))
if use_cuda:
net.cuda()
else:
net.cpu()
net.eval()
self.NET.append(net)
# --- end load model ----
# --- load permutation ----
self.permutation.append(
np.load(
FROM_DIR +
"/permutation_cifar_%s_dct_sign_permutation_subband_%s_%d.npy"
% (model_name, sb, p)))
R = 10
if args.permut == args.channel_n or not args.is_random_channels:
R = 1
data_dir_ = "data/attacked/cifar/multi_channel/%s/permutations_%d/p%d" % (
args.model, args.permut, args.pixels)
model_dir_ = "checkpoints/multi_channel/%s/permutations_%d" % (args.model,
args.permut)
SUBBANDS = ["d", "h", "v"]
EPOCHS = [100, 100, 100, 100, 100, 100, 100, 100, 100]
P = np.asarray(range(1, round(args.permut / 3) + 1))
if args.model == "vgg16":
Net = modelvgg.VGG('VGG16')
elif args.model == "resnet18":
Net = modelresnet.ResNet18()
E = []
for r in range(R):
if args.is_random_channels:
error = testClasifierRandom(Net,
EPOCHS,
model_dir_,
SUBBANDS=SUBBANDS,
P=P,
channel_n=args.channel_n,
is_vanila=args.is_vanila,
is_prefilt=args.is_prefilt,