def build(self, hidden_layers=[16], activations=['relu'], dropout=0.5,
learning_rate=0.01, l2_norm=5e-4, p1=1., p2=1.,
n_power_iterations=1, epsilon=0.03, xi=1e-6):
with self.device:
x = Input(batch_shape=[self.n_nodes, self.n_features], dtype=tf.float32, name='features')
adj = Input(batch_shape=[self.n_nodes, self.n_nodes], dtype=tf.float32, sparse=True, name='adj_matrix')
index = Input(batch_shape=[None], dtype=tf.int32, name='index')
self.GCN_layers = [GraphConvolution(hidden_layers[0],
activation=activations[0],
kernel_regularizer=regularizers.l2(l2_norm)),
GraphConvolution(self.n_classes)]
self.dropout_layer = Dropout(dropout)
logit = self.propagation(x, adj)
output = tf.gather(logit, index)
output = Softmax()(output)
model = Model(inputs=[x, adj, index], outputs=output)
self.model = model
self.train_metric = SparseCategoricalAccuracy()
self.test_metric = SparseCategoricalAccuracy()
self.optimizer = Adam(lr=learning_rate)
self.built = True
self.p1 = p1 # Alpha
self.p2 = p2 # Beta
self.xi = xi # Small constant for finite difference
self.epsilon = epsilon # Norm length for (virtual) adversarial training
self.n_power_iterations = n_power_iterations # Number of power iterations
def build(self, hidden_layers=[16], activations=['relu'], dropout=0.5, learning_rate=0.01, l2_norm=5e-4, p1=1.4, p2=0.7, epsilon=0.01):
with self.device:
x = Input(batch_shape=[self.n_nodes, self.n_features], dtype=tf.float32, name='features')
adj = Input(batch_shape=[self.n_nodes, self.n_nodes], dtype=tf.float32, sparse=True, name='adj_matrix')
index = Input(batch_shape=[None], dtype=tf.int32, name='index')
self.GCN_layers = [GraphConvolution(hidden_layers[0], activation=activations[0],
kernel_regularizer=regularizers.l2(l2_norm)),
GraphConvolution(self.n_classes)]
self.dropout_layer = Dropout(rate=dropout)
logit = self.propagation(x, adj)
logit = tf.ensure_shape(logit, (self.n_nodes, self.n_classes))
output = tf.gather(logit, index)
output = Softmax()(output)
model = Model(inputs=[x, adj, index], outputs=output)
model.compile(loss='sparse_categorical_crossentropy', optimizer=Adam(lr=learning_rate), metrics=['accuracy'])
entropy_loss = entropy_y_x(logit)
vat_loss = self.virtual_adversarial_loss(x, adj, logit, epsilon)
model.add_loss(p1 * vat_loss + p2 * entropy_loss)
self.model = model
self.adv_optimizer = Adam(lr=learning_rate/10)
self.built = True
def build(self, hidden_layers=[16], activations=['relu'], dropout=0.5,
learning_rate=0.01, l2_norm=5e-4, use_bias=False):
with self.device:
x = Input(batch_shape=[self.n_nodes, self.n_features], dtype=tf.float32, name='features')
adj = Input(batch_shape=[self.n_nodes, self.n_nodes], dtype=tf.float32, sparse=True, name='adj_matrix')
index = Input(batch_shape=[None], dtype=tf.int32, name='index')
h = x
for hid, activation in zip(hidden_layers, activations):
h = GraphConvolution(hid, use_bias=use_bias,
activation=activation,
kernel_regularizer=regularizers.l2(l2_norm))([h, adj])
h = Dropout(rate=dropout)(h)
h = GraphConvolution(self.n_classes, use_bias=use_bias)([h, adj])
h = tf.ensure_shape(h, [self.n_nodes, self.n_classes])
h = tf.gather(h, index)
output = Softmax()(h)
model = Model(inputs=[x, adj, index], outputs=output)
model.compile(loss='sparse_categorical_crossentropy', optimizer=Adam(lr=learning_rate), metrics=['accuracy'])
self.model = model
self.built = True
def build(self,
hidden_layers=[32],
activations=['relu'],
dropout=0.5,
learning_rate=0.01,
l2_norm=1e-5):
with self.device:
x = Input(batch_shape=[None, self.n_features],
dtype=tf.float32,
name='features')
adj = Input(batch_shape=[None, None],
dtype=tf.float32,
sparse=True,
name='adj_matrix')
mask = Input(batch_shape=[None], dtype=tf.bool, name='mask')
h = Dropout(rate=dropout)(x)
for hid, activation in zip(hidden_layers, activations):
h = GraphConvolution(
hid,
activation=activation,
kernel_regularizer=regularizers.l2(l2_norm))([h, adj])
h = Dropout(rate=dropout)(h)
h = GraphConvolution(self.n_classes)([h, adj])
h = tf.boolean_mask(h, mask)
output = Softmax()(h)
model = Model(inputs=[x, adj, mask], outputs=output)
model.compile(loss='sparse_categorical_crossentropy',
optimizer=Adam(lr=learning_rate),
metrics=['accuracy'],
experimental_run_tf_function=False)
self.model = model
self.built = True
def build_GCN(x):
h = Dropout(rate=dropout)(x)
for hid, activation in zip(hidden_layers, activations):
h = GraphConvolution(
hid,
use_bias=use_bias,
activation=activation,
kernel_regularizer=regularizers.l2(l2_norm))([h, adj])
# h = Dropout(rate=dropout)(h)
h = GraphConvolution(self.n_classes,
use_bias=use_bias)([h, adj])
h = tf.ensure_shape(h, [self.n_nodes, self.n_classes])
h = tf.gather(h, index)
output = Softmax()(h)
model = Model(inputs=[x, adj, index], outputs=output)
model.compile(loss='categorical_crossentropy',
optimizer=RMSprop(lr=learning_rate),
metrics=['accuracy'])
return model