def main():
args = parse_args()
print ("Configured paths")
# if os.path.exists(args.embedding_path):
# os._exit(0)
graph = nx.read_weighted_edgelist(args.edgelist, delimiter=" ", nodetype=None,create_using=nx.Graph())
graph_int = nx.read_weighted_edgelist(args.edgelist, delimiter=" ", nodetype=int,create_using=nx.Graph())
model = Struc2Vec(graph.to_directed(), walk_length=10, num_walks=8,workers=8, verbose=40 )
walks=model.return_walk_list()
walks_int=[]
for one_walk in walks:
walks_int.append( [int(i) for i in one_walk])
walks=walks_int
graph=graph_int
#print(walks)
assert not (args.visualise and args.embedding_dim > 2), "Can only visualise two dimensions"
assert args.embedding_path is not None, "you must specify a path to save embedding"
configure_paths(args)
# build model
num_nodes = len(graph)
model = build_model(num_nodes, args)
model, initial_epoch = load_weights(model, args)
optimizer = ExponentialMappingOptimizer(lr=args.lr)
loss = hyperbolic_softmax_loss(sigma=args.sigma)
model.compile(optimizer=optimizer,
loss=loss,
target_tensors=[tf.placeholder(dtype=tf.int32)])
model.summary()
callbacks = [
TerminateOnNaN(),
EarlyStopping(monitor="loss",
patience=args.patience,
verbose=True),
Checkpointer(epoch=initial_epoch,
nodes=sorted(graph.nodes()),
embedding_directory=args.embedding_path)
]
positive_samples, negative_samples, probs = \
determine_positive_and_negative_samples(graph,walks,args)
if args.use_generator:
print ("Training with data generator with {} worker threads".format(args.workers))
training_generator = TrainingDataGenerator(positive_samples,
probs,
model,
args)
model.fit_generator(training_generator,
workers=args.workers,
max_queue_size=10,
use_multiprocessing=args.workers>0,
epochs=args.num_epochs,
steps_per_epoch=len(training_generator),
initial_epoch=initial_epoch,
verbose=args.verbose,
callbacks=callbacks
)
else:
print ("Training without data generator")
train_x = np.append(positive_samples, negative_samples, axis=-1)
train_y = np.zeros([len(train_x), 1, 1], dtype=np.int32 )
model.fit(train_x, train_y,
shuffle=True,
batch_size=args.batch_size,
epochs=args.num_epochs,
initial_epoch=initial_epoch,
verbose=args.verbose,
callbacks=callbacks
)
print ("Training complete")
embedding = model.get_weights()[0]
embedding = hyperboloid_to_poincare_ball(embedding)
print(embedding)
ax=plot(embedding)
theta = np.linspace(0, 2 * np.pi, 200)
x = np.cos(theta)
y = np.sin(theta)
ax.plot(x, y, color="black", linewidth=2)
ax.axis("equal")
ax.figure.savefig("Hyper.pdf",bbox_inches='tight')
def main():
args = parse_args()
assert not (args.visualise and args.embedding_dim > 2), "Can only visualise two dimensions"
assert args.embedding_path is not None, "you must specify a path to save embedding"
if not args.no_walks:
assert args.walk_path is not None, "you must specify a path to save walks"
random.seed(args.seed)
np.random.seed(args.seed)
tf.set_random_seed(args.seed)
graph, features, node_labels = load_data(args)
print ("Loaded dataset")
if False:
plot_degree_dist(graph, "degree distribution")
configure_paths(args)
print ("Configured paths")
# build model
num_nodes = len(graph)
model = build_model(num_nodes, args)
model, initial_epoch = load_weights(model, args)
optimizer = ExponentialMappingOptimizer(lr=args.lr)
loss = hyperbolic_softmax_loss(sigma=args.sigma)
model.compile(optimizer=optimizer,
loss=loss,
target_tensors=[tf.placeholder(dtype=tf.int32)])
model.summary()
callbacks = [
TerminateOnNaN(),
EarlyStopping(monitor="loss",
patience=args.patience,
verbose=True),
Checkpointer(epoch=initial_epoch,
nodes=sorted(graph.nodes()),
embedding_directory=args.embedding_path)
]
positive_samples, negative_samples, probs = \
determine_positive_and_negative_samples(graph,
features, args)
del features # remove features reference to free up memory
if args.use_generator:
print ("Training with data generator with {} worker threads".format(args.workers))
training_generator = TrainingDataGenerator(positive_samples,
probs,
model,
args)
model.fit_generator(training_generator,
workers=args.workers,
max_queue_size=10,
use_multiprocessing=args.workers>0,
epochs=args.num_epochs,
steps_per_epoch=len(training_generator),
initial_epoch=initial_epoch,
verbose=args.verbose,
callbacks=callbacks
)
else:
print ("Training without data generator")
train_x = np.append(positive_samples, negative_samples, axis=-1)
train_y = np.zeros([len(train_x), 1, 1], dtype=np.int32 )
model.fit(train_x, train_y,
shuffle=True,
batch_size=args.batch_size,
epochs=args.num_epochs,
initial_epoch=initial_epoch,
verbose=args.verbose,
callbacks=callbacks
)
print ("Training complete")
if args.visualise:
embedding = model.get_weights()[0]
if embedding.shape[1] == 3:
print ("projecting to poincare ball")
embedding = hyperboloid_to_poincare_ball(embedding)
draw_graph(graph,
embedding,
node_labels,
path="2d-poincare-disk-visualisation.png")
def main():
args = parse_args()
test_results_dir = args.test_results_dir
if not os.path.exists(test_results_dir):
os.makedirs(test_results_dir, exist_ok=True)
test_results_filename = os.path.join(test_results_dir,
"test_results.csv")
if check_complete(test_results_filename, args.seed):
return
test_results_lock_filename = os.path.join(test_results_dir,
"test_results.lock")
touch(test_results_lock_filename)
_, _, node_labels = load_data(args)
print ("Loaded dataset")
embedding = load_embedding(args.dist_fn, args.embedding_directory)
min_count = 10
if node_labels.shape[1] == 1: # remove any node belonging to an under-represented class
label_counts = Counter(node_labels.flatten())
mask = np.array([label_counts[l] >= min_count
for l in node_labels.flatten()])
embedding = embedding[mask]
node_labels = node_labels[mask]
else:
assert node_labels.shape[1] > 1
idx = node_labels.sum(0) >= min_count
node_labels = node_labels[:, idx]
idx = node_labels.any(-1)
embedding = embedding[idx]
node_labels = node_labels[idx]
if args.dist_fn == "hyperboloid":
print ("loaded a hyperboloid embedding")
# print ("projecting from hyperboloid to klein")
# embedding = hyperboloid_to_klein(embedding)
print ("projecting from hyperboloid to poincare")
embedding = hyperboloid_to_poincare_ball(embedding)
print ("projecting from poincare to klein")
embedding = poincare_ball_to_klein(embedding)
elif args.dist_fn == "poincare":
print ("loaded a poincare embedding")
# print ("projecting from poincare to klein")
# embedding = poincare_ball_to_hyperboloid(embedding)
# embedding = hyperboloid_to_klein(embedding)
print ("projecting from poincare to klein")
embedding = poincare_ball_to_klein(embedding)
test_results = {}
label_percentages, f1_micros, f1_macros = \
evaluate_node_classification(embedding, node_labels)
for label_percentage, f1_micro, f1_macro in zip(label_percentages, f1_micros, f1_macros):
print ("{:.2f}".format(label_percentage),
"micro = {:.2f}".format(f1_micro),
"macro = {:.2f}".format(f1_macro) )
test_results.update({"{:.2f}_micro".format(label_percentage): f1_micro})
test_results.update({"{:.2f}_macro".format(label_percentage): f1_macro})
k = 10
k_fold_roc, k_fold_f1, k_fold_precision, k_fold_recall = \
evaluate_kfold_label_classification(embedding, node_labels, k=k)
test_results.update({
"{}-fold-roc".format(k): k_fold_roc,
"{}-fold-f1".format(k): k_fold_f1,
"{}-fold-precision".format(k): k_fold_precision,
"{}-fold-recall".format(k): k_fold_recall,
})
print ("saving test results to {}".format(test_results_filename))
threadsafe_save_test_results(test_results_lock_filename, test_results_filename, args.seed, data=test_results )