def compute_inference(self, node_features_in, sp_adj_matrix, is_training):
with tf.variable_scope('edge-model'):
z_latent = gcn_module(node_features_in, sp_adj_matrix,
self.n_hidden_edge, self.p_drop_edge,
is_training, self.input_dim,
self.sparse_features)
adj_matrix_pred = compute_adj(z_latent, self.att_mechanism,
self.p_drop_edge, is_training)
self.adj_matrix_pred = adj_matrix_pred
with tf.variable_scope('node-model'):
z_latent = tf.sparse_concat(
axis=1,
sp_inputs=[
tf.contrib.layers.dense_to_sparse(z_latent),
node_features_in
])
sparse_features = True
input_dim = self.n_hidden_edge[-1] + self.input_dim
logits = gcn_module(z_latent,
sp_adj_matrix,
self.n_hidden_node,
self.p_drop_node,
is_training,
input_dim,
sparse_features=sparse_features)
return logits, adj_matrix_pred
def compute_inference(self, node_features, adj_matrix, is_training):
"""Forward step for GAE model."""
sparse = self.sparse_features
in_dim = self.input_dim
with tf.variable_scope('edge-model'):
h0 = gcn_module(node_features, adj_matrix, self.n_hidden[:-1],
self.p_drop, is_training, in_dim, sparse)
# N x F
with tf.variable_scope('mean'):
z_mean = gcn_module(h0, adj_matrix, self.n_hidden[-1:], self.p_drop,
is_training, self.n_hidden[-2], False)
self.z_mean = z_mean
with tf.variable_scope('std'):
# N x F
z_log_std = gcn_module(h0, adj_matrix, self.n_hidden[-1:], self.p_drop,
is_training, self.n_hidden[-2], False)
self.z_log_std = z_log_std
# add noise during training
noise = tf.random_normal([self.nb_nodes, self.n_hidden[-1]
]) * tf.exp(z_log_std)
z = tf.cond(is_training, lambda: tf.add(z_mean, noise),
lambda: z_mean)
# N x N
adj_matrix_pred = compute_adj(z, self.att_mechanism, self.p_drop,
is_training)
self.adj_matrix_pred = tf.nn.sigmoid(adj_matrix_pred)
return adj_matrix_pred
def compute_inference(self, node_features_in, sp_adj_matrix, is_training):
with tf.variable_scope('edge-model'):
z_latent = gcn_module(node_features_in, sp_adj_matrix,
self.n_hidden_edge, self.p_drop_edge,
is_training, self.input_dim,
self.sparse_features)
adj_matrix_pred = compute_adj(z_latent, self.att_mechanism,
self.p_drop_edge, is_training)
self.adj_matrix_pred = adj_matrix_pred
with tf.variable_scope('node-model'):
z_latent = tf.sparse_concat(
axis=1,
sp_inputs=[
tf.contrib.layers.dense_to_sparse(z_latent),
node_features_in
])
sparse_features = True
input_dim = self.n_hidden_edge[-1] + self.input_dim
sp_adj_train = tf.SparseTensor(
indices=sp_adj_matrix.indices,
values=tf.ones_like(sp_adj_matrix.values),
dense_shape=sp_adj_matrix.dense_shape)
logits = gat_module(z_latent,
sp_adj_train,
self.n_hidden_node,
self.n_att_node,
self.p_drop_node,
is_training,
input_dim,
sparse_features=sparse_features,
average_last=True)
return logits, adj_matrix_pred
def compute_inference(self, node_features, adj_matrix, is_training):
"""Forward step for GAE model."""
sparse = self.sparse_features
in_dim = self.input_dim
with tf.variable_scope('edge-model'):
h0 = gcn_module(node_features, adj_matrix, self.n_hidden, self.p_drop,
is_training, in_dim, sparse)
adj_matrix_pred = compute_adj(h0, self.att_mechanism, self.p_drop,
is_training)
self.adj_matrix_pred = tf.nn.sigmoid(adj_matrix_pred)
return adj_matrix_pred
def compute_inference(self, node_features, adj_matrix, is_training):
with tf.variable_scope('node-model'):
hidden_repr = mlp_module(node_features,
self.n_hidden,
self.p_drop,
is_training,
self.input_dim,
self.sparse_features,
use_bias=True,
return_hidden=True)
logits = hidden_repr[-1]
hidden_repr_reg = hidden_repr[self.semi_emb_k]
l2_scores = compute_adj(hidden_repr_reg,
self.att_mechanism,
self.p_drop,
is_training=False)
self.l2_scores = tf.gather_nd(l2_scores, adj_matrix.indices)
return logits