How is the projection vector p trainable?

[Yfhu1103]

Hi, In your "ops.py" line 89~line 107, you define graph pool. But it seems to be that you have missed the gate step, i.e. scores should be multiplied to X, but I do not see this step in your code. Could do please explain how the projection vector is trainable? Thank you!

[HongyangGao]

Hi, thank you for pointing it out. It seems I uploaded the wrong version. I have updated the code. Thank you.

[Yfhu1103]

I used the previous codes and new codes to do the graph classification task and I get the similar results as shown in the paper, I do not understand why, is it make sense?

[HongyangGao]

Hi, I think it's interesting. But since the p is not trainable, it should stay the same all the time. That means it can be considered as a fixed composition pattern.

For example, c = 2a + 3b, where 2 and 3 are fixed weights. This means the network may train the GCN part to make the resulting ranking scores useful in pooling.

Also, the architecture of the U-Net should mostly benefits the performance instead of just relying a projection vector p.

Best,

Hongyang

On Fri, Jul 12, 2019 at 3:28 AM Mrslock notifications@github.com wrote:

I used the previous codes and new codes to do the graph classification task and I get the similar results as shown in the paper, I do not understand why, is it make sense?

— You are receiving this because you commented. Reply to this email directly, view it on GitHub https://github.com/HongyangGao/gunet/issues/4?email_source=notifications&email_token=ABQZZL7KD5AYDQHTDN7BE5LP7BMETA5CNFSM4ICDNNFKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGODZZMB6Q#issuecomment-510836986, or mute the thread https://github.com/notifications/unsubscribe-auth/ABQZZL4Y2VMYFFVMMRWZ2PDP7BMETANCNFSM4ICDNNFA .

[Yfhu1103]

Thanks for explaination