def test_lib(profile: str = './profiles/McDo_faultfree.json', nb_nets: int = 5):
acc_dict = {}
handler.from_json(profile)
for scaling in SCALINGS:
avg_list = []
for _ in range(nb_nets):
handler.perturb_tensors(scaling)
avg_list.append(utils.test_accuracy(net, testloader))
handler.restore()
acc_dict[scaling] = avg_list
return acc_dict
def test_legacy(profile: np.ndarray = np.zeros_like(lanmax_state_space), nb_nets: int = 5):
acc_dict = {}
wm.pis = profile
for scaling in SCALINGS:
avg_list = []
wm.scaling = scaling
for _ in range(nb_nets):
wm.binarize()
avg_list.append(utils.test_accuracy(net, testloader))
wm.restore()
acc_dict[scaling] = avg_list
return acc_dict
batch_size=5,
shuffle=True,
num_workers=2)
testset = utils.R2Dataset(2, 10000)
testloader = torch.utils.data.DataLoader(testset,
batch_size=4,
shuffle=False,
num_workers=2)
# Loading Net
net = utils.Xor()
net.load_state_dict(torch.load(PATH))
# Initial Test
init_accuracy = utils.test_accuracy(net, testloader)
print('Initial accuracy of the network: %3.2f %%' % (100 * init_accuracy))
# Defining loss and optim
criterion = nn.CrossEntropyLoss()
optimizer = optim.SGD(net.parameters(), lr=0.001, momentum=0.9)
utils.train_net(net,
optimizer,
criterion,
trainloader,
nb_epochs=100,
prt=False)
# Final test
accuracy = utils.test_accuracy(net, testloader)
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,
criterion,
trainloader,
5,
prt=False)
init_trained_acc = utils.test_accuracy(net_copy, testloader)
handler_trained_acc = utils.test_accuracy(handler, testloader)
csv_data.append([
net.name,
"activations":
{
"net": null,
"modules":null
}
}
To unwrap this, let's go step by step. The "weights" section dictates which
parameters (weights and biases) will be perturbed. You can specify an entire
network under the "net" field, specific modules or layers in the "modules" list
and specific parameters in the "tensors" list. These modules and parameters
need to be in the net.named_modules() and net.named_parameters() respectively
in order to be found by the constructor.
Specifying representations and perturbations is done by giving the name of the
desired class in the "name" field, and creating a field for each constructor
argument.
If no representation or perturbation is wanted, specifying null
respectively will void that field.
The same principles hold true for activations.
Note: the saved file from handler.to_json() is not guaranteed to match the one
used to load it, as there are multiple inputs that can give the same output,
and for the time being we adress everything by tensor instead of looking to
group tensors to simplify the saved file.
"""
import FaultyMemory.utils as utils
accuracy = utils.test_accuracy(my_handler, testloader)
config_path = './profiles/saved_handler.json'
my_handler.from_json(config_path)