def attackerOracle(G, coverage_probs, phi, omega=4, num_paths=100):
N = nx.number_of_nodes(G)
coverage_prob_matrix = torch.zeros((N, N))
for i, e in enumerate(list(G.edges())):
coverage_prob_matrix[e[0]][e[1]] = coverage_probs[i]
coverage_prob_matrix[e[1]][e[0]] = coverage_probs[
i] # for undirected graph only
# EXACT EDGE PROBS
unbiased_probs = phi2prob(G, phi)
biased_probs = prob2unbiased(
G, -coverage_probs, unbiased_probs,
omega=omega) # feeding negative coverage to be biased
# EMPIRICAL EDGE PROBS
path_list = []
simulated_defender_utility_list = []
for _ in range(num_paths):
path = getMarkovianWalk(G, biased_probs)
# path = getSimplePath(G, path) # TODO
path_list.append(path)
defender_utility = -G.node[path[-1][1]]['utility']
for e in path:
defender_utility *= (1 - coverage_prob_matrix[e[0]][e[1]])
simulated_defender_utility_list.append(defender_utility)
simulated_defender_utility = np.mean(simulated_defender_utility_list)
return path_list, simulated_defender_utility
def train_model(train_data,
validate_data,
test_data,
lr=0.1,
learning_model='random_walk_distribution',
block_selection='coverage',
n_epochs=150,
batch_size=100,
optimizer='adam',
omega=4,
training_method='surrogate-decision-focused',
max_norm=0.1,
block_cut_size=0.5,
T_size=10):
net2 = GCNPredictionNet2(feature_size)
net2.train()
sample_graph = train_data[0][0]
init_T, init_s = torch.rand(sample_graph.number_of_edges(),
T_size), torch.zeros(
sample_graph.number_of_edges())
T, s = torch.tensor(
normalize_matrix_positive(init_T), requires_grad=True
), torch.tensor(
init_s, requires_grad=False
) # bias term s can cause infeasibility. It is not yet known how to resolve it.
full_T, full_s = torch.eye(sample_graph.number_of_edges(),
requires_grad=False), torch.zeros(
sample_graph.number_of_edges(),
requires_grad=False)
T_lr = lr
# ================ Optimizer ================
if optimizer == 'adam':
optimizer = optim.Adam(net2.parameters(), lr=lr)
T_optimizer = optim.Adam([T, s], lr=T_lr)
# optimizer=optim.Adam(list(net2.parameters()) + [T], lr=lr)
elif optimizer == 'sgd':
optimizer = optim.SGD(net2.parameters(), lr=lr)
T_optimizer = optim.SGD([T, s], lr=T_lr)
elif optimizer == 'adamax':
optimizer = optim.Adamax(net2.parameters(), lr=lr)
T_optimizer = optim.Adamax([T, s], lr=T_lr)
# scheduler = ReduceLROnPlateau(optimizer, 'min')
scheduler = ReduceLROnPlateau(optimizer, 'min')
T_scheduler = ReduceLROnPlateau(T_optimizer, 'min')
training_loss_list, validating_loss_list, testing_loss_list = [], [], []
training_defender_utility_list, validating_defender_utility_list, testing_defender_utility_list = [], [], []
print("Training...")
forward_time, qp_time, backward_time = 0, 0, 0
pretrain_epochs = 0
decay_rate = 0.95
for epoch in range(-1, n_epochs):
epoch_forward_time, epoch_qp_time, epoch_backward_time = 0, 0, 0
if epoch <= pretrain_epochs:
ts_weight = 1
df_weight = 0
else:
ts_weight = decay_rate**(epoch - pretrain_epochs)
df_weight = 1 - ts_weight
for mode in ["training", "validating", "testing"]:
if mode == "training":
dataset = train_data
epoch_loss_list = training_loss_list
epoch_def_list = training_defender_utility_list
if epoch > 0:
net2.train()
else:
net2.eval()
elif mode == "validating":
dataset = validate_data
epoch_loss_list = validating_loss_list
epoch_def_list = validating_defender_utility_list
net2.eval()
elif mode == "testing":
dataset = test_data
epoch_loss_list = testing_loss_list
epoch_def_list = testing_defender_utility_list
net2.eval()
else:
raise TypeError("Not valid mode: {}".format(mode))
loss_list, def_obj_list = [], []
for iter_n in tqdm.trange(len(dataset)):
G, Fv, coverage_prob, phi_true, path_list, cut, log_prob, unbiased_probs_true, previous_gradient = dataset[
iter_n]
n, m = G.number_of_nodes(), G.number_of_edges()
budget = G.graph['budget']
# ==================== Visualization ===================
# if iter_n == 0 and mode == 'training':
# from plot_utils import plot_graph, reduce_dimension
# T_reduced = T.detach().numpy() # reduce_dimension(T.detach().numpy())
# plot_graph(G, T_reduced, epoch)
# =============== Compute edge probabilities ===========
Fv_torch = torch.as_tensor(Fv, dtype=torch.float)
edge_index = torch.Tensor(list(
nx.DiGraph(G).edges())).long().t()
phi_pred = net2(Fv_torch, edge_index).view(
-1
) if epoch >= 0 else phi_true # when epoch < 0, testing the optimal loss and defender utility
# phi_pred.require_grad = True
unbiased_probs_pred = phi2prob(
G, phi_pred) if epoch >= 0 else unbiased_probs_true
biased_probs_pred = prob2unbiased(
G, -coverage_prob, unbiased_probs_pred,
omega=omega) # feeding negative coverage to be biased
# =================== Compute loss =====================
log_prob_pred = torch.zeros(1)
for path in path_list:
for e in path:
log_prob_pred -= torch.log(
biased_probs_pred[e[0]][e[1]])
log_prob_pred /= len(path_list)
loss = (log_prob_pred - log_prob)[0]
# ============== COMPUTE DEFENDER UTILITY ==============
single_data = dataset[iter_n]
if epoch == -1: # optimal solution
cut_size = m
def_obj, def_coverage, (
single_forward_time, single_qp_time
) = getDefUtility(
single_data,
full_T,
full_s,
unbiased_probs_pred,
learning_model,
cut_size=cut_size,
omega=omega,
verbose=False,
training_mode=False,
training_method=training_method,
block_selection=block_selection) # feed forward only
single_forward_time, single_qp_time = 0, 0 # testing epoch so not counting the computation time
elif mode == 'testing' or mode == "validating" or epoch <= 0:
cut_size = m
def_obj, def_coverage, (
single_forward_time, single_qp_time
) = getDefUtility(
single_data,
T,
s,
unbiased_probs_pred,
learning_model,
cut_size=cut_size,
omega=omega,
verbose=False,
training_mode=False,
training_method=training_method,
block_selection=block_selection) # feed forward only
single_forward_time, single_qp_time = 0, 0 # testing epoch so not counting the computation time
else:
if training_method == 'decision-focused' or training_method == 'surrogate-decision-focused':
cut_size = m
else:
raise TypeError('Not defined method')
def_obj, def_coverage, (
single_forward_time, single_qp_time) = getDefUtility(
single_data,
T,
s,
unbiased_probs_pred,
learning_model,
cut_size=cut_size,
omega=omega,
verbose=False,
training_mode=True,
training_method=training_method,
block_selection=block_selection
) # most time-consuming part
epoch_forward_time += single_forward_time
epoch_qp_time += single_qp_time
def_obj_list.append(def_obj.item())
loss_list.append(loss.item())
if (iter_n % batch_size == (batch_size - 1)) and (
epoch > 0) and (mode == "training"):
backward_start_time = time.time()
optimizer.zero_grad()
T_optimizer.zero_grad()
try:
if training_method == "decision-focused" or training_method == "surrogate-decision-focused":
(-def_obj).backward()
# (-def_obj * df_weight + loss * ts_weight).backward()
else:
raise TypeError("Not Implemented Method")
# torch.nn.utils.clip_grad_norm_(net2.parameters(), max_norm=max_norm) # gradient clipping
for parameter in net2.parameters():
parameter.grad = torch.clamp(parameter.grad,
min=-max_norm,
max=max_norm)
T.grad = torch.clamp(T.grad,
min=-max_norm,
max=max_norm)
optimizer.step()
T_optimizer.step()
except:
print("no grad is backpropagated...")
epoch_backward_time += time.time() - backward_start_time
# ============== normalize T matrix =================
T.data = normalize_matrix_positive(T.data)
# T.data = normalize_matrix_qr(T.data)
# s.data = normalize_vector(s.data, max_value=budget)
# print(s.data)
# ========= scheduler using validation set ==========
if (epoch > 0) and (mode == "validating"):
if training_method == "decision-focused" or training_method == "surrogate-decision-focused":
scheduler.step(-np.mean(def_obj_list))
T_scheduler.step(-np.mean(def_obj_list))
else:
raise TypeError("Not Implemented Method")
# ======= Storing loss and defender utility =========
epoch_loss_list.append(np.mean(loss_list))
epoch_def_list.append(np.mean(def_obj_list))
# ========== Print stuff after every epoch ==========
np.random.shuffle(dataset)
print("Mode: {}/ Epoch number: {}/ Loss: {}/ DefU: {}".format(
mode, epoch, np.mean(loss_list), np.mean(def_obj_list)))
print('Forward time for this epoch: {}'.format(epoch_forward_time))
print('QP time for this epoch: {}'.format(epoch_qp_time))
print('Backward time for this epoch: {}'.format(epoch_backward_time))
if epoch >= 0:
forward_time += epoch_forward_time
qp_time += epoch_qp_time
backward_time += epoch_backward_time
# ============= early stopping criteria =============
kk = 3
if epoch >= kk * 2 - 1:
GE_counts = np.sum(
np.array(validating_defender_utility_list[1:][-kk:]) <=
np.array(validating_defender_utility_list[1:][-2 * kk:-kk]) +
1e-4)
print(
'Generalization error increases counts: {}'.format(GE_counts))
if GE_counts == kk:
break
average_nodes = np.mean([x[0].number_of_nodes() for x in train_data] +
[x[0].number_of_nodes() for x in validate_data] +
[x[0].number_of_nodes() for x in test_data])
average_edges = np.mean([x[0].number_of_edges() for x in train_data] +
[x[0].number_of_edges() for x in validate_data] +
[x[0].number_of_edges() for x in test_data])
print('Total forward time: {}'.format(forward_time))
print('Total qp time: {}'.format(qp_time))
print('Total backward time: {}'.format(backward_time))
return net2, training_loss_list, validating_loss_list, testing_loss_list, training_defender_utility_list, validating_defender_utility_list, testing_defender_utility_list, (
forward_time, qp_time, backward_time), epoch
def train_model(train_data, validate_data, test_data, lr=0.1, learning_model='random_walk_distribution', block_selection='coverage',
n_epochs=150, batch_size=100, optimizer='adam', omega=4, training_method='two-stage', max_norm=0.1, block_cut_size=0.5):
net2= GCNPredictionNet2(feature_size)
net2.train()
if optimizer=='adam':
optimizer=optim.Adam(net2.parameters(), lr=lr)
elif optimizer=='sgd':
optimizer=optim.SGD(net2.parameters(), lr=lr)
elif optimizer=='adamax':
optimizer=optim.Adamax(net2.parameters(), lr=lr)
# scheduler = ReduceLROnPlateau(optimizer, 'min')
scheduler = ReduceLROnPlateau(optimizer, 'min', factor=0.5)
training_loss_list, validating_loss_list, testing_loss_list = [], [], []
training_defender_utility_list, validating_defender_utility_list, testing_defender_utility_list = [], [], []
print ("Training...")
forward_time, qp_time, backward_time = 0, 0, 0
evaluate = False
pretrain_epochs = 0
decay_rate = 0.95
for epoch in range(-1, n_epochs):
if epoch == n_epochs - 1:
evaluate = True
epoch_forward_time, epoch_qp_time, epoch_backward_time = 0, 0, 0
if epoch <= pretrain_epochs:
ts_weight = 1
df_weight = 0
else:
ts_weight = decay_rate ** (epoch - pretrain_epochs)
df_weight = 1 - ts_weight
for mode in ["training", "validating", "testing"]:
if mode == "training":
dataset = train_data
epoch_loss_list = training_loss_list
epoch_def_list = training_defender_utility_list
if epoch > 0:
net2.train()
else:
net2.eval()
elif mode == "validating":
dataset = validate_data
epoch_loss_list = validating_loss_list
epoch_def_list = validating_defender_utility_list
net2.eval()
elif mode == "testing":
dataset = test_data
epoch_loss_list = testing_loss_list
epoch_def_list = testing_defender_utility_list
net2.eval()
else:
raise TypeError("Not valid mode: {}".format(mode))
loss_list, def_obj_list = [], []
for iter_n in tqdm.trange(len(dataset)):
G, Fv, coverage_prob, phi_true, path_list, cut, log_prob, unbiased_probs_true, previous_gradient = dataset[iter_n]
n, m = G.number_of_nodes(), G.number_of_edges()
# =============== Compute edge probabilities ===========
Fv_torch = torch.as_tensor(Fv, dtype=torch.float)
edge_index = torch.Tensor(list(nx.DiGraph(G).edges())).long().t()
phi_pred = net2(Fv_torch, edge_index).view(-1) if epoch >= 0 else phi_true # when epoch < 0, testing the optimal loss and defender utility
# phi_pred.require_grad = True
unbiased_probs_pred = phi2prob(G, phi_pred) if epoch >= 0 else unbiased_probs_true
biased_probs_pred = prob2unbiased(G, -coverage_prob, unbiased_probs_pred, omega=omega) # feeding negative coverage to be biased
# =================== Compute loss =====================
log_prob_pred = torch.zeros(1)
for path in path_list:
for e in path:
log_prob_pred -= torch.log(biased_probs_pred[e[0]][e[1]])
log_prob_pred /= len(path_list)
loss = (log_prob_pred - log_prob)[0]
# ============== COMPUTE DEFENDER UTILITY ==============
single_data = dataset[iter_n]
if mode == 'testing' or mode == "validating" or epoch <= 0: # or training_method == "two-stage" or epoch <= 0:
cut_size = m
if training_method == "two-stage" and not evaluate:
def_obj, def_coverage, single_forward_time, single_qp_time = torch.Tensor([-float('Inf')]), None, 0, 0
else:
def_obj, def_coverage, (single_forward_time, single_qp_time) = getDefUtility(single_data, unbiased_probs_pred, learning_model, cut_size=cut_size, omega=omega, verbose=False, training_mode=False, training_method=training_method, block_selection=block_selection) # feed forward only
single_foward_time, single_qp_time = 0, 0
else:
if training_method == "two-stage" or epoch <= pretrain_epochs:
cut_size = m
if evaluate:
def_obj, def_coverage, (single_forward_time, single_qp_time) = getDefUtility(single_data, unbiased_probs_pred, learning_model, cut_size=cut_size, omega=omega, verbose=False, training_mode=False, training_method=training_method, block_selection=block_selection) # most time-consuming part
else
def_obj, def_coverage, single_forward_time, single_qp_time = torch.Tensor([-float('Inf')]), None, 0, 0
# ignore the time of computing defender utility
else:
if training_method == 'decision-focused':
cut_size = m
elif training_method == 'block-decision-focused' or training_method == 'hybrid' or training_method == 'corrected-block-decision-focused':
if type(block_cut_size) == str and block_cut_size[-1] == 'n':
cut_size = int(n * float(block_cut_size[:-1]))
elif block_cut_size <= 1:
cut_size = int(m * block_cut_size)
else:
cut_size = block_cut_size
else:
raise TypeError('Not defined method')
def_obj, def_coverage, (single_forward_time, single_qp_time) = getDefUtility(single_data, unbiased_probs_pred, learning_model, cut_size=cut_size, omega=omega, verbose=False, training_mode=True, training_method=training_method, block_selection=block_selection) # most time-consuming part
epoch_forward_time += single_forward_time
epoch_qp_time += single_qp_time
def_obj_list.append(def_obj.item())
loss_list.append(loss.item())
if (iter_n%batch_size == (batch_size-1)) and (epoch > 0) and (mode == "training"):
backward_start_time = time.time()
optimizer.zero_grad()
try:
if training_method == "two-stage" or epoch <= pretrain_epochs:
loss.backward()
elif training_method == "decision-focused" or training_method == "block-decision-focused" or training_method == 'corrected-block-decision-focused':
# (-def_obj).backward()
(-def_obj * m / cut_size).backward()
elif training_method == "hybrid":
# ((-def_obj) * df_weight + loss[0] * ts_weight).backward()
((-def_obj * m / cut_size) * df_weight + loss * ts_weight).backward()
else:
raise TypeError("Not Implemented Method")
torch.nn.utils.clip_grad_norm_(net2.parameters(), max_norm=max_norm) # gradient clipping
# print(torch.norm(net2.gcn1.weight.grad))
# print(torch.norm(net2.gcn2.weight.grad))
# print(torch.norm(net2.fc1.weight.grad))
optimizer.step()
except:
print("no grad is backpropagated...")
epoch_backward_time += time.time() - backward_start_time
if (epoch > 0) and (mode == "validating"):
if training_method == "two-stage":
scheduler.step(np.mean(loss_list))
elif training_method == "decision-focused" or training_method == "block-decision-focused" or training_method == 'corrected-block-decision-focused' or training_method == 'hybrid':
scheduler.step(-np.mean(def_obj_list))
else:
raise TypeError("Not Implemented Method")
# Storing loss and defender utility
epoch_loss_list.append(np.mean(loss_list))
epoch_def_list.append(np.mean(def_obj_list))
################################### Print stuff after every epoch
np.random.shuffle(dataset)
print("Mode: {}/ Epoch number: {}/ Loss: {}/ DefU: {}".format(
mode, epoch, np.mean(loss_list), np.mean(def_obj_list)))
print('Forward time for this epoch: {}'.format(epoch_forward_time))
print('QP time for this epoch: {}'.format(epoch_qp_time))
print('Backward time for this epoch: {}'.format(epoch_backward_time))
if epoch >= 0:
forward_time += epoch_forward_time
qp_time += epoch_qp_time
backward_time += epoch_backward_time
# ============= early stopping criteria =============
kk = 3
if epoch >= kk*2 -1:
if training_method == 'two-stage':
if evaluate:
break
GE_counts = np.sum(np.array(validating_loss_list[1:][-kk:]) >= np.array(validating_loss_list[1:][-2*kk:-kk]) - 1e-4)
print('Generalization error increases counts: {}'.format(GE_counts))
if GE_counts == kk:
evaluate = True
else: # surrogate or decision-focused
GE_counts = np.sum(np.array(validating_defender_utility_list[1:][-kk:]) <= np.array(validating_defender_utility_list[1:][-2*kk:-kk]) + 1e-4)
print('Generalization error increases counts: {}'.format(GE_counts))
if GE_counts == kk:
break
average_nodes = np.mean([x[0].number_of_nodes() for x in train_data] + [x[0].number_of_nodes() for x in validate_data] + [x[0].number_of_nodes() for x in test_data])
average_edges = np.mean([x[0].number_of_edges() for x in train_data] + [x[0].number_of_edges() for x in validate_data] + [x[0].number_of_edges() for x in test_data])
print('Total forward time: {}'.format(forward_time))
print('Total qp time: {}'.format(qp_time))
print('Total backward time: {}'.format(backward_time))
return net2, training_loss_list, validating_loss_list, testing_loss_list, training_defender_utility_list, validating_defender_utility_list, testing_defender_utility_list, (forward_time, qp_time, backward_time), epoch
def generateSyntheticData(node_feature_size,
omega=4,
n_graphs=20,
samples_per_graph=100,
empirical_samples_per_instance=10,
fixed_graph=False,
path_type='random_walk',
N_low=16,
N_high=20,
e_low=0.6,
e_high=0.7,
budget=2,
train_test_split_ratio=(0.7, 0.1, 0.2),
n_sources=1,
n_targets=1,
random_seed=0,
noise_level=0):
# Random seed setting
print("Random seed: {}".format(random_seed))
if random_seed != 0:
torch.manual_seed(random_seed)
np.random.seed(random_seed)
random.seed(random_seed)
# initialization
data = [
] # aggregate all the data first then split into training and testing
generated_node_feature_size = node_feature_size
net3 = featureGenerationNet2(generated_node_feature_size)
# net3= featureGenerationNet2(node_feature_size)
n_samples = n_graphs * samples_per_graph
print("N_samples: ", n_samples)
for graph_number in range(n_graphs):
'''
# Pick the graph in cyclic fashion from the correct list of graphs
graph_index=0
if sample_number<n_training_samples:
G=training_graphs[sample_number%n_training_graphs]
graph_index=sample_number%n_training_graphs
else:
G=testing_graphs[sample_number%n_testing_graphs]
graph_index=sample_number%n_testing_graphs
'''
while True:
G = returnGraph(fixed_graph=fixed_graph,
n_sources=n_sources,
n_targets=n_targets,
N_low=N_low,
N_high=N_high,
e_low=e_low,
e_high=e_high,
budget=budget)
# PRECOMPUTE A MIN-CUT
m = G.number_of_edges()
edges = G.edges()
edge2index = {}
for idx, edge in enumerate(edges):
edge2index[edge] = idx
edge2index[(edge[1], edge[0])] = idx
dummyG = copy.deepcopy(G)
dummyG.add_nodes_from(['ds', 'dt'
]) # 1000 dummy source, 2000 dummy target
for x in dummyG.graph['sources']:
dummyG.add_edge('ds', x, capacity=100)
for x in dummyG.graph['targets']:
dummyG.add_edge(x, 'dt', capacity=100)
value, partition = nx.minimum_cut(dummyG, 'ds', 'dt')
print('cut size:', value)
partition0, partition1 = set(partition[0]), set(partition[1])
cut = []
for idx, edge in enumerate(G.edges()):
if edge[0] in partition0 and edge[1] in partition1:
cut.append(idx)
elif edge[0] in partition1 and edge[1] in partition0:
cut.append(idx)
print('cut:', cut)
if value > budget:
break
# COMPUTE ADJACENCY MATRIX
edge_index = torch.Tensor(list(nx.DiGraph(G).edges())).long().t()
N = nx.number_of_nodes(G)
m = nx.number_of_edges(G)
# Visualization
# from plot_utils import plot_graph
# colors = np.random.random((m, 3))
# plot_graph(G, colors)
'''
# Define node features for each of the n nodes
for node in list(G.nodes()):
node_features=np.random.randn(node_feature_size)
# TODO: Use a better feature computation for a given node
G.node[node]['node_features']=node_features
# Generate features
Fv=np.zeros((N,node_feature_size))
for node in list(G.nodes()):
Fv[node]=G.node[node]['node_features']
'''
random_feature_indices = np.random.choice(generated_node_feature_size,
node_feature_size,
replace=False)
for _ in range(samples_per_graph):
# Randomly assign coverage probability
private_coverage_prob = np.random.rand(m)
private_coverage_prob = (private_coverage_prob /
sum(private_coverage_prob)) * budget
coverage_prob_matrix = torch.zeros(N, N)
for i, e in enumerate(list(G.edges())):
coverage_prob_matrix[e[0]][e[1]] = private_coverage_prob[i]
coverage_prob_matrix[e[1]][e[0]] = private_coverage_prob[i]
# Randomly generate attractiveness
# phi is the attractiveness function, phi(v,f) for each of the N nodes, v
phi = generatePhi(G, fixed_phi=fixed_graph)
phi = phi - np.mean(phi)
phi = torch.as_tensor(phi, dtype=torch.float)
# Generate features from phi values
Fv_torch = net3.forward(phi.view(-1, 1), edge_index)
Fv = Fv_torch.detach().numpy()
Fv = Fv[:, random_feature_indices]
# EXACT EDGE PROBS
unbiased_probs = phi2prob(G, phi)
biased_probs = prob2unbiased(
G, -private_coverage_prob, unbiased_probs,
omega=omega) # feeding negative coverage to get biased
# Call Attacker Oracle
path_list, _ = attackerOracle(
G,
private_coverage_prob,
phi,
omega=omega,
num_paths=empirical_samples_per_instance)
# EMPIRICAL EDGE PROBS
empirical_transition_probs = torch.zeros((N, N))
for path in path_list:
for e in path:
empirical_transition_probs[e[0]][e[1]] += 1
# row_sum = torch.sum(empirical_transition_probs, dim=1)
adj = torch.Tensor(
nx.adjacency_matrix(G, nodelist=range(N)).toarray())
empirical_transition_probs = empirical_transition_probs / torch.sum(
empirical_transition_probs, dim=1, keepdim=True)
# empirical_transition_probs[row_sum == 0] = 0
empirical_transition_probs[torch.isnan(
empirical_transition_probs)] = 0
# empirical_transition_probs = empirical_transition_probs * adj
# print('biased:', empirical_transition_probs)
empirical_unbiased_probs = prob2unbiased(
G, private_coverage_prob, empirical_transition_probs, omega)
# print('unbiased:', empirical_unbiased_probs)
previous_gradient = torch.zeros(m, m)
# DATA POINT
if path_type == 'random_walk_distribution':
log_prob = torch.zeros(1)
for path in path_list:
for e in path:
log_prob -= torch.log(biased_probs[e[0]][e[1]])
log_prob /= len(path_list)
data_point = (G, Fv, private_coverage_prob, phi, path_list,
cut, log_prob, unbiased_probs, previous_gradient)
elif path_type == 'empirical_distribution':
log_prob = torch.zeros(1)
for path in path_list:
for e in path:
log_prob -= torch.log(
empirical_transition_probs[e[0]][e[1]])
log_prob /= len(path_list)
data_point = (G, Fv, private_coverage_prob, phi, path_list,
cut, log_prob, empirical_unbiased_probs,
previous_gradient)
else:
raise (TypeError)
data.append(data_point)
data = np.array(data)
np.random.shuffle(data)
print("average node size:",
np.mean([x[0].number_of_nodes() for x in data]))
print("average edge size:",
np.mean([x[0].number_of_edges() for x in data]))
train_size = int(train_test_split_ratio[0] * len(data))
validate_size = int(train_test_split_ratio[1] * len(data))
Fv_training_list = [data[i][1] for i in range(train_size)]
Fv_training_features = np.concatenate(
Fv_training_list, axis=0) # concatenate all the features
Fv_training_mean = np.mean(Fv_training_features)
Fv_training_std = np.std(Fv_training_features)
print('mean:', Fv_training_mean, 'std', Fv_training_std)
for i in range(len(data)): # normalizing based on the training set
data[i][1] = (
data[i][1] -
Fv_training_mean) / Fv_training_std + np.random.normal(
size=data[i][1].shape) * noise_level # maintain that var=1
training_data, validate_data, testing_data = data[:train_size], data[
train_size:train_size + validate_size], data[train_size +
validate_size:]
return np.array(training_data), np.array(validate_data), np.array(
testing_data)