shinkyuy / graph_neural_net Goto Github PK

Hi, I just found your repository and and it's really helpful to understand graph networks and it's variants.
I just gone through your notebook and it was really informative and maintained.
In your experiments you have only predicted node and presented node classification problem.
I have few doubts, can we use same networks for edge classification or do we have other powerful networks for edge classification instead of node classification?
If you can post a notebook with edge classification that would be very helpful though.

Thank you,
And keep posting and learning. :)

def loss(self, lr, active=tf.nn.relu,optim=tf.train.AdamOptimizer):
self.L=tf.identity(self.X)
for i in range(len(self.layer)):
self.L=self.graph_conv(self.L, self.adj, self.layer[i], activation=active)
self.L= tf.nn.dropout(self.L, self.dropout)

    self.hypothesis=tf.layers.dense(tf.matmul(self.adj,self.L), self.output_dim, activation=None)

def loss(self, lr, active=tf.nn.relu,optim=tf.train.AdamOptimizer):
self.L=tf.identity(self.X)
for i in range(len(self.layer)):
self.L=self.graph_conv(self.L, self.adj, self.layer[i], activation=active)
self.L= tf.nn.dropout(self.L, self.dropout)

    self.hypothesis=tf.layers.dense(tf.matmul(self.adj,self.L), self.output_dim, activation=None)

你每次调用loss的时候都会创建新的tf.layers.dense 那么每次网络的权重都是新的