features_trainloader[str(l + 1)].append(norm_feat)
labels.append(target.item())
for l in range(4):
features_trainloader[str(l + 1)] = torch.tensor(
features_trainloader[str(l + 1)])
pickle.dump(features_trainloader,
open('features/features_trainloader_tensor_' + net + '.pickle',
'wb'),
protocol=4)
pickle.dump(labels, open('features/labels_' + net + '.pickle', 'wb'))
if extract:
if net == 'mnist':
libary_loader = get_mnist_train_loader(batch_size=1, shuffle=False)
elif net == 'fashion_mnist':
libary_loader = get_fashion_mnist_train_loader(batch_size=1,
shuffle=False)
elif net == 'svhn':
libary_loader = get_svhn_train_loader(batch_size=1, shuffle=False)
extract_libary(model1_, libary_loader, net)
with open('features/features_trainloader_tensor_' + net + '.pickle',
'rb') as handle:
features_trainloader = pickle.load(handle)
with open('features/labels_' + net + '.pickle', 'rb') as handle:
labels = pickle.load(handle)
for l in range(4):
features_trainloader[str(l + 1)] = features_trainloader[str(l +
np.save('train_meta/w_dist_' + str(epoch) + '.npy',
self.losses['w_dist'])
if __name__ == '__main__':
LAMBDA = 10
ITERS = 200
parser = argparse.ArgumentParser(description='Train MNIST')
parser.add_argument('--seed', default=0, type=int)
parser.add_argument('--train_batch_size', default=100, type=int)
parser.add_argument('--test_batch_size', default=1000, type=int)
parser.add_argument('--log_interval', default=10, type=int)
args = parser.parse_args()
model_filename = "mnist_wass_net.pt"
torch.manual_seed(args.seed)
train_loader = get_mnist_train_loader(batch_size=args.train_batch_size,
shuffle=True)
test_loader = get_mnist_test_loader(batch_size=args.test_batch_size,
shuffle=False)
filename = "mnist_lenet5_clntrained.pt"
model = LeNet5()
model.load_state_dict(
torch.load(os.path.join(TRAINED_MODEL_PATH, filename),
map_location='cpu'))
model.to(device)
model.eval()
adversary = L1BasicIterativeAttack(
model,
loss_fn=nn.CrossEntropyLoss(reduction="sum"),
eps=0.2,
nb_iter=10,
from advertorch.test_utils import LeNet5
from advertorch_examples.utils import get_mnist_train_loader
from advertorch_examples.utils import get_mnist_test_loader
import constopt
from constopt.adversary import Adversary
from constopt.optim import PGD, PGDMadry, FrankWolfe, MomentumFrankWolfe
# Setup
torch.manual_seed(0)
use_cuda = torch.cuda.is_available()
device = torch.device("cuda" if use_cuda else "cpu")
# Data Loaders
train_loader = get_mnist_train_loader(batch_size=128, shuffle=True)
test_loader = get_mnist_test_loader(batch_size=512, shuffle=True)
# Model setup
model = LeNet5()
model.to(device)
criterion = nn.CrossEntropyLoss()
# Outer optimization parameters
nb_epochs = 20
optimizer = torch.optim.SGD(model.parameters(), lr=.1, momentum=.9)
# Inner optimization parameters
eps = 0.3
constraint = constopt.constraints.make_LpBall(alpha=eps, p=np.inf)
inner_iter = 40