What is different from the original code?

[hkthirano]

As I trained, the result of this repository is more accurate than the original paper (SEMI-SUPERVISED CLASSIFICATION WITH GRAPH CONVOLUTIONAL NETWORKS) in cora dateset.

What is different from the original code?

[tkipf]

The data splits are different, the normalization of the adjacency matrix is slightly different and there is no dropout on the first layer.

On Tue 25. Sep 2018 at 10:02 hokuto_HIRANO notifications@github.com wrote:

As I trained, the result of this repository is more accurate than the original paper (SEMI-SUPERVISED CLASSIFICATION WITH GRAPH CONVOLUTIONAL NETWORKS) in cora dateset.

What is different from the original code?

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/tkipf/pygcn/issues/20, or mute the thread https://github.com/notifications/unsubscribe-auth/AHAcYEuO_TElTlJUjoBe3gc4rS13e0dtks5uefE3gaJpZM4W4Kyk .

[kkteru]

I have been trying to understand what you mean by 'there is no dropout on the first layer' and 'sparse dropout'.

def forward(self, x, adj):
        x = F.relu(self.gc1(x, adj))
--->    x = F.dropout(x, self.dropout, training=self.training)
        x = self.gc2(x, adj)
        return F.log_softmax(x, dim=1)

^ I am assuming this to be the dropout on first layer. Please let me know what I am missing.

Thanks!

[tkipf]

Yes, this is correct- my wording was a bit ambiguous: the original TensorFlow-based implementation also uses Dropout on the input features directly (which is what I meant by first layer).

On Sun, Feb 3, 2019 at 12:34 AM Komal Kumar notifications@github.com wrote:

I have been trying to understand what you mean by 'there is no dropout on the first layer' and 'sparse dropout'.

def forward(self, x, adj): x = F.relu(self.gc1(x, adj)) ---> x = F.dropout(x, self.dropout, training=self.training) x = self.gc2(x, adj) return F.log_softmax(x, dim=1)

^ I am assuming this to be the dropout on first layer. Please let me know what I am missing.

Thanks!

— You are receiving this because you commented. Reply to this email directly, view it on GitHub https://github.com/tkipf/pygcn/issues/20#issuecomment-460008872, or mute the thread https://github.com/notifications/unsubscribe-auth/AHAcYDxYVubyD_N1o8CNAPEAtZN1hJOzks5vJiCggaJpZM4W4Kyk .

[kkteru]

I see. Thanks for the clarification!

[chenzhao]

The data splits are different, the normalization of the adjacency matrix is slightly different and there is no dropout on the first layer. … On Tue 25. Sep 2018 at 10:02 hokuto_HIRANO @.***> wrote: As I trained, the result of this repository is more accurate than the original paper (SEMI-SUPERVISED CLASSIFICATION WITH GRAPH CONVOLUTIONAL NETWORKS) in cora dateset. What is different from the original code? — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#20>, or mute the thread https://github.com/notifications/unsubscribe-auth/AHAcYEuO_TElTlJUjoBe3gc4rS13e0dtks5uefE3gaJpZM4W4Kyk .

Hi, this is some very clean code. Good job. I compared the accuracy on Cora with this repo and GCN sample codes from PyG and DGL. Surprisingly, result from this one is about 2 to 3 points better (0.83 v.s. 0.81). Is it because of the slightly different adjacency matrix? Thanks a lot.