def generate_point(net, testloader, prob): # DEPRECATED
"""
This function is deprecated because of the use of clusters
Handlers now store tensor information themselves
"""
net_pert = copy.deepcopy(net)
pert_pert = P.Zeros(prob)
c_pert = C.Cluster([pert_pert], networks=[net_pert])
handler_pert = H.Handler(net_pert, [c_pert])
net_acti = copy.deepcopy(net)
pert_acti = P.Zeros(prob)
c_acti = C.Cluster([pert_acti], network_activations=[net_acti])
handler_acti = H.Handler(net_acti, [c_acti])
net_both = copy.deepcopy(net)
pert_both = P.Zeros(prob)
c_both = C.Cluster([pert_both],
networks=[net_both],
network_activations=[net_both])
handler_both = H.Handler(net_both, [c_both])
clean_accuracy = test_accuracy(net, testloader)
pert_accuracy = test_accuracy(handler_pert, testloader)
acti_accuracy = test_accuracy(handler_acti, testloader)
both_accuracy = test_accuracy(handler_both, testloader)
return clean_accuracy, pert_accuracy, acti_accuracy, both_accuracy
"Pert",
"Repr",
"Init acc",
"Handler",
"Init Trained",
"Handler trained",
]]
for net in net_list:
for repr in repr_list:
for pert in pert_list:
criterion = nn.CrossEntropyLoss()
optimizer = optim.SGD(net.parameters(), lr=0.001, momentum=0.9)
net_copy = copy.deepcopy(net)
handler = H.Handler(net_copy)
handler.device = device
pert.width = repr.width
handler.add_net_parameters(representation=repr, perturb=pert)
mse_list = handler.compute_MSE()
min_mse_index = mse_list.index(min(mse_list))
net_copy.eval()
init_acc = utils.test_accuracy(net_copy, testloader)
handler_acc = utils.test_accuracy(handler, testloader)
net_copy.train()
utils.train_net(handler,
optimizer,
root='./data', train=False, download=True, transform=transform)
testloader = torch.utils.data.DataLoader(
testset, batch_size=512, shuffle=False, num_workers=2)
# init params: depth=28, width=10
net = McDo.WRN_McDonnell(
depth=28, width=10, num_classes=10, dropit=False, actprec=3).to(device)
state_dict = torch.load(PATH, map_location=device)[
'model_state_dict']
lanmax_state_space = torch.load(PATH, map_location=device)[
'lanmax_state_space']
net.load_state_dict(state_dict)
net.device = device
net.eval()
handler = H.Handler(net)
handler.from_json('./profiles/McDo.json')
wm = BinaryWeightMemory(net, p=0.01, scaling=SCALING)
wm.pis = lanmax_state_space
def iter_weights():
weights = []
for m in net.modules():
if isinstance(m, nn.Conv2d) or isinstance(m, nn.Linear):
y = np.bincount(m.weight.data.numpy().flatten().astype(int) + 2)
ii = np.nonzero(y)[0]
print(list(zip(ii, y[ii])))
weights.append(copy.deepcopy(m.weight.data.numpy()))
return weights
x = F.relu(self.fc2(x))
x = self.fc3(x)
return x
net = McDo.WRN_McDonnell(depth=28,
width=10,
num_classes=10,
dropit=False,
actprec=3)
# net = SimpleNet()
nb_clusters = 4
cluster_list = [C.Cluster() for _ in range(nb_clusters)]
print(cluster_list)
handler = H.Handler(net, cluster_list)
handler.from_json('./profiles/McDo.json')
handler.assign_clusters()
# handler.to_json('./profiles/saved_handler.json')
exit()
inp = torch.tensor([1.])
out = handler(inp)
print(handler)
print('out: ', out)
"""
def confint_mean_testset():