def compute_estimated_cut(G, partition_dict, marked_nodes):
"""Compute overestimated cut value on a partial marked graph"""
buffer_dict = dict(partition_dict)
# cut marked nodes using strong minority
gc.marked_nodes_could_be_cut(G, buffer_dict, marked_nodes)
bk = set() # decided nodes in partition_dict
m = set() # marked nodes in partition_dict
for node in partition_dict:
if partition_dict[node] == gc.MARKED:
m.add(node)
else:
bk.add(node)
b1k1 = set() # decided and undecided nodes in buffer_dict
for node in m:
if buffer_dict[node] != gc.MARKED:
b1k1.add(node)
# nx.set_node_attributes(G, gc.PARTITION, buffer_dict)
nx.set_node_attributes(G, buffer_dict, gc.PARTITION)
result = (G.subgraph(m).number_of_edges() +
gc.cut_edges(G.subgraph(bk | b1k1)) -
gc.cut_edges(G.subgraph(b1k1)))
return result
def complete_cut(G, partition_dict, nodes_stack):
"""Compute the maximum cut from a partial partitioned graph."""
marked_nodes_stack = []
while gc.could_be_cut(G, partition_dict):
pass
# filter away the decided nodes
for i in nodes_stack:
if partition_dict[i] is gc.UNDECIDED or partition_dict[i] is gc.MARKED:
marked_nodes_stack.append(i)
if not marked_nodes_stack:
return partition_dict, gc.cut_edges(G, partition_dict)
candidate = marked_nodes_stack.pop()
# Complete cut
blue_cut, blue_cut_val = aux_local_consistent_max_cut(
G, dict(partition_dict), list(marked_nodes_stack), int(candidate),
gc.BLUE)
black_cut, black_cut_val = aux_local_consistent_max_cut(
G, dict(partition_dict), list(marked_nodes_stack), int(candidate),
gc.BLACK)
if blue_cut_val > black_cut_val:
return blue_cut, blue_cut_val
return black_cut, black_cut_val
def aux_pruning_local_consistent_max_cut(G, partition_dict, degree_node_seq,
candidate, partition_attribute):
"""Helper function to pruning_local_consistent_max_cut"""
partition_dict[candidate] = partition_attribute
# check consistency in the candidate and its neighbors only
check_nodes = nx.neighbors(G, candidate)
if not gc.is_cut_consistent(G, partition_dict, check_nodes):
return None, 0
if not degree_node_seq:
return partition_dict, gc.cut_edges(G, partition_dict)
while (gc.could_be_cut(G, partition_dict)):
pass
# pick a new candidate
candidate = choose_new_candidate(partition_dict, degree_node_seq)
blue_cut, blue_cut_val = aux_pruning_local_consistent_max_cut(
G, dict(partition_dict), list(degree_node_seq), candidate, gc.BLUE)
black_cut, black_cut_val = aux_pruning_local_consistent_max_cut(
G, dict(partition_dict), list(degree_node_seq), candidate, gc.BLACK)
if blue_cut is None and black_cut is None:
return None, 0
# Choose best cut according to the effective cut value
if blue_cut_val > black_cut_val:
return blue_cut, blue_cut_val
return black_cut, black_cut_val
def greedy_choice(G, candidate, blue_nodes, black_nodes, visited):
"""Helper function to greedy cut"""
G.node[candidate][gc.PARTITION] = gc.BLUE
blue_cut_val = gc.cut_edges(nx.subgraph(G, visited))
G.node[candidate][gc.PARTITION] = gc.BLACK
black_cut_val = gc.cut_edges(nx.subgraph(G, visited))
if blue_cut_val > black_cut_val:
G.node[candidate][gc.PARTITION] = gc.BLUE
blue_nodes.add(candidate)
else:
black_nodes.add(candidate)
return blue_nodes, black_nodes
def ising_ground_truth(cf, info_mtx, fig_save_path=""):
print("Running the Ground Truth...")
ising_model = Ising_model(cf, info_mtx)
dim = info_mtx.shape[0]
if cf.pb_type == "maxcut":
# 1/2\sum_edges 1-node1*node2 = 1/2*num_edges - 1/2*\sum_edges node1*node2
graph = ising_model.graph
start_time = time.time()
result = ising.search(graph,
num_states=3,
show_progress=True,
chunk_size=0)
end_time = time.time()
energy = result.energies[0]
state = decode_state(result.states[0], dim, list(range(dim)))
num_edges = info_mtx.sum() / 2
score = num_edges / 2 - energy
score1 = ising_model(state)
if abs((score - score1) / score) > 1e-2:
raise ("Mismatched energy - result1={}, result2={}".format(
score, score1))
# plot the graph
# Laplacian matrices are real and symmetric, so we can use eigh,
# the variation on eig specialized for Hermetian matrices.
G = laplacian_to_graph(info_mtx)
# Laplacian matrices are real and symmetric, so we can use eigh,
# the variation on eig specialized for Hermetian matrices.
# method from maxCutPy
gt.execution_time(mc.local_consistent_max_cut, 1, G)
score1 = gc.cut_edges(G)
if abs((score - score1)) != 0:
raise ("Mismatched maxcut - result1={}, result2={}".format(
score, score1))
nbunch = G.nodes()
for i in nbunch:
G.node[i]['partition'] = state[i]
gd.draw_cut_graph(G)
plt.savefig(fig_save_path)
plt.close()
if cf.pb_type == "spinglass":
# \sum_edges J_ij*node1*node2
graph = ising_model.graph
start_time = time.time()
result = ising.search(graph,
num_states=3,
show_progress=True,
chunk_size=0)
end_time = time.time()
energy = result.energies[0]
state = decode_state(result.states[0], dim, list(range(dim)))
energy1 = ising_model(state)
if abs((energy - energy1) / energy) > 1e-2:
raise ("Mismatched energy - result1={}, result2={}".format(
energy, energy1))
score = energy
time_elapsed = end_time - start_time
return score, state, time_elapsed
def second_lemma(G):
"""Compute second lemma."""
cut_edges = gc.cut_edges(G)
uncut_edges = G.number_of_edges() - cut_edges
numerator = float(uncut_edges - cut_edges)
denominator = float(G.number_of_edges())
return numerator / denominator
def compare_cut_algorithm_results(cut_alg, results_graphs, test_name=''):
graphs_dict = gg.read_from_file(results_graphs + '.dat')
location = cut_alg.__name__
error_list = []
results_dict = defaultdict(list)
print('Algorithm: ' + str(cut_alg.__name__))
print('Benchamark: ' + results_graphs)
print('Starting..:)')
print('Graph\Result Epsilon\tCurrent Epsilon')
i = 0
for opt_eps in sorted(graphs_dict.iterkeys()):
for A, c in graphs_dict[opt_eps]:
G = nx.from_numpy_matrix(A)
cut_alg(G)
if gc.cut_edges(G) < G.number_of_edges() / 2:
print('####### ERROR 1 #######')
error_list.append(G)
if gc.are_undecided_nodes(G):
print('####### ERROR 2 #######')
error_list.append(G)
this_eps = gc.compute_epsilon(G)
print(str(i) + '\t' + str(opt_eps) + '\t' + str(this_eps))
results_dict[opt_eps].append(this_eps)
i += 1
gg.write_to_file(results_dict,
test_name + 'results_' + results_graphs + '.dat',
location)
if len(error_list) != 0:
gg.write_to_file(error_list,
test_name + 'errors_' + results_graphs + '.dat',
location)
def brute_force_max_cut(G):
"""Compute maximum cut of a graph considering all the possible cuts."""
max_cut_value = 0
max_cut_ind = 0
n = G.number_of_nodes()
for i in range(1, 2**(n - 1)):
cut_graph = nx.Graph(G)
gc.binary_cut(cut_graph, i)
value = gc.cut_edges(cut_graph)
if value > max_cut_value:
max_cut_value = value
max_cut_ind = i
gc.binary_cut(G, max_cut_ind)
return gc.partition_dictionary(G), max_cut_value
def plot_graph(cf, L, save_path):
if cf.pb_type == "maxcut":
seed = cf.random_seed
# Laplacian matrices are real and symmetric, so we can use eigh,
# the variation on eig specialized for Hermetian matrices.
# method from maxCutPy
time = gt.execution_time(mc.local_consistent_max_cut, 1, G)
result = gc.cut_edges(G)
return time, result
if cf.pb_type == "spinglass":
def decode_state(state_repr, no_spins, labels):
state = {}
for i in range(no_spins):
state[labels[no_spins - (i + 1)]] = 1 if state_repr % 2 else -1
state_repr //= 2
return state
def check(dic, graph):
total = 0
for edge, energy in graph.items():
total += energy * dic[edge[0]] * dic[edge[1]]
return total
J = L
graph = {}
shape = J.shape
size = int(shape[0] * shape[1])
for i in range(shape[0]):
for j in range(shape[1]):
graph[(i, j)] = J[i, j]
result = ising.search(graph, num_states=4)
print(result.energies, result.states)
print(decode_state(result.states[0], size, list(range(size))))
print(
check(decode_state(result.states[0], size, list(range(size))),
graph))
def trevisan_approximation_alg(G):
B, K = recursive_spectral_cut(G)
# set blue and black nodes in graph G
gc.set_partitions(G, B, K)
return gc.cut_edges(G)
