def main(argv):
if len(argv) > 1:
raise app.UsageError('Too many command-line arguments.')
print('Bröther may i have some self-lööps')
n_nodes = FLAGS.n_nodes
n_clusters = FLAGS.n_clusters
train_size = FLAGS.train_size
data_clean, data_dirty, labels = overlapping_gaussians(n_nodes, n_clusters)
graph_clean = construct_knn_graph(data_clean).todense().A1.reshape(
n_nodes, n_nodes)
train_mask = np.zeros(n_nodes, dtype=np.bool)
train_mask[np.random.choice(np.arange(n_nodes),
int(n_nodes * train_size),
replace=False)] = True
test_mask = ~train_mask
print(f'Data shape: {data_clean.shape}, graph shape: {graph_clean.shape}')
print(f'Train size: {train_mask.sum()}, test size: {test_mask.sum()}')
input_features = tf.keras.layers.Input(shape=(2, ))
input_features_corrupted = tf.keras.layers.Input(shape=(2, ))
input_graph = tf.keras.layers.Input((n_nodes, ))
encoder = [GCN(64), GCN(32)]
model = deep_graph_infomax(
[input_features, input_features_corrupted, input_graph], encoder)
def loss(model, x, y, training):
_, y_ = model(x, training=training)
return loss_object(y_true=y, y_pred=y_)
def grad(model, inputs, targets):
with tf.GradientTape() as tape:
loss_value = loss(model, inputs, targets, training=True)
for loss_internal in model.losses:
loss_value += loss_internal
return loss_value, tape.gradient(loss_value, model.trainable_variables)
labels_dgi = tf.concat([tf.zeros([n_nodes, 1]), tf.ones([n_nodes, 1])], 0)
loss_object = tf.keras.losses.BinaryCrossentropy(from_logits=True)
optimizer = tf.keras.optimizers.Adam(FLAGS.learning_rate)
for epoch in range(FLAGS.n_epochs):
data_corrupted = data_dirty.copy()
perc_shuffle = np.linspace(1, 0.05, FLAGS.n_epochs)[epoch]
# perc_shuffle = 1
rows_shuffle = np.random.choice(np.arange(n_nodes),
int(n_nodes * perc_shuffle))
data_corrupted_tmp = data_corrupted[rows_shuffle]
np.random.shuffle(data_corrupted_tmp)
data_corrupted[rows_shuffle] = data_corrupted_tmp
loss_value, grads = grad(model,
[data_dirty, data_corrupted, graph_clean],
labels_dgi)
optimizer.apply_gradients(zip(grads, model.trainable_variables))
print('epoch %d, loss: %0.4f, shuffle %0.2f%%' %
(epoch, loss_value.numpy(), 100 * perc_shuffle))
representations, _ = model([data_dirty, data_corrupted, graph_clean],
training=False)
representations = representations.numpy()
clf = LogisticRegression(solver='lbfgs', multi_class='multinomial')
clf.fit(representations[train_mask], labels[train_mask])
clusters = clf.predict(representations[test_mask])
print(
'NMI:',
normalized_mutual_info_score(labels[test_mask],
clusters,
average_method='arithmetic'))
print('Accuracy:', 100 * accuracy_score(labels[test_mask], clusters))
def main(argv):
if len(argv) > 1:
raise app.UsageError('Too many command-line arguments.')
print('Starting', format_filename())
if FLAGS.load_strategy == 'schur':
adjacency, features, labels, label_mask = load_npz_to_sparse_graph(
FLAGS.graph_path)
elif FLAGS.load_strategy == 'kipf':
adjacency, features, labels, label_mask = load_kipf_data(
*os.path.split(FLAGS.graph_path))
else:
raise Exception('Unknown loading strategy!')
n_nodes = adjacency.shape[0]
feature_size = features.shape[1]
architecture = [int(x) for x in FLAGS.architecture.strip('[]').split('_')]
graph_clean_normalized = scipy_to_tf(
normalize_graph(adjacency.copy(), normalized=True))
input_features = tf.keras.layers.Input(shape=(feature_size,))
input_features_corrupted = tf.keras.layers.Input(shape=(feature_size,))
input_graph = tf.keras.layers.Input((n_nodes,), sparse=True)
encoder = [GCN(512) for size in architecture]
model = deep_graph_infomax(
[input_features, input_features_corrupted, input_graph], encoder)
def loss(model, x, y, training):
_, y_ = model(x, training=training)
return loss_object(y_true=y, y_pred=y_)
def grad(model, inputs, targets):
with tf.GradientTape() as tape:
loss_value = loss(model, inputs, targets, training=True)
for loss_internal in model.losses:
loss_value += loss_internal
return loss_value, tape.gradient(loss_value, model.trainable_variables)
loss_object = tf.keras.losses.BinaryCrossentropy(from_logits=True)
optimizer = tf.keras.optimizers.Adam(FLAGS.learning_rate)
patience = 20
best_loss = 999
patience_counter = 0
for epoch in range(FLAGS.n_epochs):
features_corr = features.copy()
pseudolabels = tf.concat([tf.zeros([n_nodes, 1]), tf.ones([n_nodes, 1])], 0)
features_corr = features_corr.copy()
np.random.shuffle(features_corr)
loss_value, grads = grad(model,
[features, features_corr, graph_clean_normalized],
pseudolabels)
optimizer.apply_gradients(zip(grads, model.trainable_variables))
loss_value = loss_value.numpy()
print(epoch, loss_value)
if loss_value > best_loss:
patience_counter += 1
if patience_counter == patience:
break
else:
best_loss = loss_value
patience_counter = 0
representations = model([features, features, graph_clean_normalized],
training=False)[0].numpy()
clf = KMeans(n_clusters=FLAGS.n_clusters)
clf.fit(representations)
clusters = clf.labels_
print('Conductance:', conductance(adjacency, clusters))
print('Modularity:', modularity(adjacency, clusters))
print(
'NMI:',
normalized_mutual_info_score(
labels, clusters[label_mask], average_method='arithmetic'))
print('Precision:', precision(labels, clusters[label_mask]))
print('Recall:', recall(labels, clusters[label_mask]))
with open(format_filename(), 'w') as out_file:
print('Conductance:', conductance(adjacency, clusters), file=out_file)
print('Modularity:', modularity(adjacency, clusters), file=out_file)
print(
'NMI:',
normalized_mutual_info_score(
labels, clusters[label_mask], average_method='arithmetic'),
file=out_file)
print('Precision:', precision(labels, clusters[label_mask]), file=out_file)
print('Recall:', recall(labels, clusters[label_mask]), file=out_file)
def main(argv):
if len(argv) > 1:
raise app.UsageError('Too many command-line arguments.')
print('Bröther may i have some self-lööps')
n_nodes = FLAGS.n_nodes
n_clusters = FLAGS.n_clusters
train_size = FLAGS.train_size
batch_size = FLAGS.batch_size
data_clean, data_dirty, labels = line_gaussians(n_nodes, n_clusters)
graph_clean = construct_knn_graph(data_clean)
n_neighbors = [15, 10] # TODO(tsitsulin): move to FLAGS.
total_matrix_size = 1 + np.cumprod(n_neighbors).sum()
train_mask = np.zeros(n_nodes, dtype=np.bool)
train_mask[np.random.choice(np.arange(n_nodes),
int(n_nodes * train_size),
replace=False)] = True
test_mask = ~train_mask
print(f'Data shape: {data_clean.shape}, graph shape: {graph_clean.shape}')
print(f'Train size: {train_mask.sum()}, test size: {test_mask.sum()}')
input_features = tf.keras.layers.Input(shape=(
total_matrix_size,
2,
))
input_features_corrupted = tf.keras.layers.Input(shape=(
total_matrix_size,
2,
))
input_graph = tf.keras.layers.Input((
total_matrix_size,
total_matrix_size,
))
encoder = [
GCN(64),
GCN(32),
tf.keras.layers.Lambda(lambda x: x[0][:, 0, :])
]
model = deep_graph_infomax(
[input_features, input_features_corrupted, input_graph], encoder)
def loss(model, x, y, training):
_, y_ = model(x, training=training)
return loss_object(y_true=y, y_pred=y_)
def grad(model, inputs, targets):
with tf.GradientTape() as tape:
loss_value = loss(model, inputs, targets, training=True)
for loss_internal in model.losses:
loss_value += loss_internal
return loss_value, tape.gradient(loss_value, model.trainable_variables)
labels_dgi = tf.concat(
[tf.zeros([batch_size, 1]),
tf.ones([batch_size, 1])], 0)
loss_object = tf.keras.losses.BinaryCrossentropy(from_logits=True)
optimizer = tf.keras.optimizers.Adam(FLAGS.learning_rate)
for epoch in range(FLAGS.n_epochs):
subgraph_mat, features_mat, _, nonzero_indices = random_batch(
graph_clean, data_dirty, batch_size, n_neighbors)
perc_shuffle = 1 # np.linspace(1, 0.25, max_epoch)[epoch]
features_corrupted = shuffle_inbatch(features_mat, nonzero_indices,
perc_shuffle)
loss_value, grads = grad(
model, [features_mat, features_corrupted, subgraph_mat],
labels_dgi)
optimizer.apply_gradients(zip(grads, model.trainable_variables))
print(
f'epoch {epoch}, loss: {loss_value.numpy():.4f}, shuffle %: {100*perc_shuffle:.2f}'
)
subgraph_mat, features_mat, _ = make_batch(graph_clean, data_dirty,
np.arange(n_nodes), n_neighbors)
representations, _ = model([features_mat, features_mat, subgraph_mat],
training=False)
representations = representations.numpy()
clf = LogisticRegression(solver='lbfgs', multi_class='multinomial')
clf.fit(representations[train_mask], labels[train_mask])
clusters = clf.predict(representations[test_mask])
print(
'NMI:',
normalized_mutual_info_score(labels[test_mask],
clusters,
average_method='arithmetic'))
print('Accuracy:', 100 * accuracy_score(labels[test_mask], clusters))
