tkipf
Why I got the 0 feature data for all nodes?
Hi, thank you again for your great work. Now I have another problem that I need your help.
I process many graphs one by one. Because the value of features_nonzero is different, and I can only initialize one model, so I did the following:
The embedding of each node like [0.0, 0.3024, 2.034 ... 0.0, 0.005, 0.03, 1.03], the len() is 2048, it contains many 0 value. In order to ensure features_nonzero is 2048, I replace 0 value with 1e-15, and I tried to get all node embeddings in a GAE model, but I got 0 value matrix
[[0. 0. 0. ... 0. 0. 0.]
[0. 0. 0. ... 0. 0. 0.]
[0. 0. 0. ... 0. 0. 0.]
...
[0. 0. 0. ... 0. 0. 0.]
[0. 0. 0. ... 0. 0. 0.]
[0. 0. 0. ... 0. 0. 0.]]
Why is that?If I don't, how do I process many graphs with one model?
What's more, I can't do like this:
I recommend using a dense matrix multiplication in the first layer if you have a somewhat dense feature vector.
On Thu, May 30, 2019 at 12:38 AM Wangzhen [email protected] wrote:
@tkipf https://github.com/tkipf Sorry, can you help me... Thank you very much!
— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/tkipf/gae/issues/36?email_source=notifications&email_token=ABYBYYHH5ZHUAH47VBBUJ2DPX576VA5CNFSM4HQBKSLKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGODWRTXCA#issuecomment-497236872, or mute the thread https://github.com/notifications/unsubscribe-auth/ABYBYYHKKXWHDQOWYXDEAQDPX576VANCNFSM4HQBKSLA .
@tkipf So I should replace the GraphConvolutionSparse with GraphConvolution, right?
What's more, if I want one step away from the central node, I can only use one time GraphConvolution in model GCNModelAE, right? If so, how to set the act, act=tf.nn.relu or act=lambda x: x?
Finally, thank you again for your patience!
Sounds correct to me. Simply don’t use an activation function on the last layer (as this is handled separately).
On Sat 1. Jun 2019 at 05:46 Wangzhen [email protected] wrote:
@tkipf https://github.com/tkipf So I should replace the GraphConvolutionSparse with GraphConvolution, right? What's more, if I want one step away from the central node, I can only use one time GraphConvolution in model GCNModelAE, right? If so, how to set the act, act=tf.nn.relu or act=lambda x: x? Finally, thank you again for your patience!
— You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub https://github.com/tkipf/gae/issues/36?email_source=notifications&email_token=ABYBYYBRK5BT5NZU4CAO463PYJVUFA5CNFSM4HQBKSLKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGODWW77LI#issuecomment-497942445, or mute the thread https://github.com/notifications/unsubscribe-auth/ABYBYYAFYO7N5WHNPT7EVRLPYJVUFANCNFSM4HQBKSLA .