class Gcat(NodeEdgeModel):
"""1 iteration Graph Convolution Attention Model."""
def __init__(self, config):
"""Initializes GCAT model."""
super(Gcat, self).__init__(config)
self.edge_model = Gae(config)
self.node_model = Gcn(config)
def compute_inference(self, node_features_in, sp_adj_matrix, is_training):
"""Forward pass for GAT model."""
adj_matrix_pred = self.edge_model.compute_inference(
node_features_in, sp_adj_matrix, is_training)
sp_adj_mask = tf.SparseTensor(indices=sp_adj_matrix.indices,
values=tf.ones_like(
sp_adj_matrix.values),
dense_shape=sp_adj_matrix.dense_shape)
sp_adj_att = sp_adj_mask * adj_matrix_pred
sp_adj_att = tf.SparseTensor(indices=sp_adj_att.indices,
values=tf.nn.leaky_relu(
sp_adj_att.values),
dense_shape=sp_adj_att.dense_shape)
sp_adj_att = tf.sparse_softmax(sp_adj_att)
logits = self.node_model.compute_inference(node_features_in,
sp_adj_att, is_training)
return logits, adj_matrix_pred
class GaeGcn(NodeEdgeModel):
"""GAE for link prediction and GCN for node classification."""
def __init__(self, config):
"""Initializes EGCNGCN model."""
super(GaeGcn, self).__init__(config)
self.edge_model = Gae(config)
self.node_model = Gcn(config)
def compute_inference(self, node_features_in, sp_adj_matrix, is_training):
adj_matrix_pred = self.edge_model.compute_inference(
node_features_in, sp_adj_matrix, is_training)
self.adj_matrix_pred = adj_matrix_pred
adj_mask = get_sp_topk(adj_matrix_pred, sp_adj_matrix, self.nb_nodes,
self.topk)
sp_adj_pred = tf.contrib.layers.dense_to_sparse(
tf.multiply(adj_mask, tf.nn.leaky_relu(adj_matrix_pred)))
sp_adj_pred = tf.sparse_softmax(sp_adj_pred)
logits = self.node_model.compute_inference(node_features_in,
sp_adj_pred, is_training)
return logits, adj_matrix_pred
class GaeGat(NodeEdgeModel):
"""GAE for link prediction and GAT for node classification."""
def __init__(self, config):
"""Initializes EGCNGAT model."""
super(GaeGat, self).__init__(config)
self.edge_model = Gae(config)
self.node_model = Gat(config)
def compute_inference(self, node_features_in, sp_adj_matrix, is_training):
adj_matrix_pred = self.edge_model.compute_inference(
node_features_in, sp_adj_matrix, is_training)
self.adj_matrix_pred = adj_matrix_pred
adj_mask = get_sp_topk(adj_matrix_pred, sp_adj_matrix, self.nb_nodes,
self.topk)
self.adj_mask = adj_mask
# masked_adj_matrix_pred = tf.multiply(adj_mask,
# tf.nn.sigmoid(adj_matrix_pred))
masked_adj_matrix_pred = mask_edges(tf.nn.sigmoid(adj_matrix_pred),
adj_mask)
sp_adj_pred = tf.contrib.layers.dense_to_sparse(masked_adj_matrix_pred)
logits = self.node_model.compute_inference(node_features_in,
sp_adj_pred, is_training)
return logits, adj_matrix_pred
def __init__(self, config):
"""Initializes EGCNGCN model."""
super(GaeGcn, self).__init__(config)
self.edge_model = Gae(config)
self.node_model = Gcn(config)
def __init__(self, config):
"""Initializes GCAT model."""
super(Gcat, self).__init__(config)
self.edge_model = Gae(config)
self.node_model = Gcn(config)
def __init__(self, config):
"""Initializes EGCN_GAT model."""
super(GaeGatConcat, self).__init__(config)
self.edge_model = Gae(config)
self.node_model = Gat(config)