Require Keras backend and version update if possible

[JunhengH]

As Theano is no longer updated, Keras automatically uses tensorflow as backend.

Keras also has been updated to 2.2, such as some module names (example: initializations => initializers. and _inbound_nodes error).

Also, the codes still support stated version theano (0.9.0) and keras(1.2.1).It could be better to fix some minor issues to up-to-date version if possible

[tkipf]

Thanks, I agree that updating the code base to work with the new keras API would be great. I am happy to accept pull requests on this, if you’d like to contribute.

A reason why Theano might still be preferable in this particular project is that Theano supports larger sparse matrices than TensorFlow. Last time I checked, TensorFlow had a hard limit on the number of non-zero elements which was too limiting for some of the experiments here. With some bookkeeping, one could implement the same mechanism with tf.gather operations, which doesn’t suffer from this limitation, however. On Tue 11. Sep 2018 at 02:37 Jeff Hao [email protected] wrote:

As Theano is no longer updated, Keras automatically uses tensorflow as backend.

Keras also has been updated to 2.2, such as some module names (example: initializations => initializers. and _inbound_nodes error).

Also, the codes still support stated version theano (0.9.0) and keras(1.2.1).It could be better to fix some minor issues to up-to-date version if possible

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/tkipf/relational-gcn/issues/5, or mute the thread https://github.com/notifications/unsubscribe-auth/AHAcYPvNAzhfe60nVKVhJbJBKrz7XIyDks5uZxPBgaJpZM4Wichm .

[wxwilcke]

I happen to have ported your code to the most recent versions of these libraries a few months ago while extending it with the ability to learn from additional features. It evolved a lot since, but I guess you can still see the necessary changes in my old commits.

[tkipf]

Thanks for sharing this! I’m sure this will be helpful for people who might stumble on this thread.

On Mon 17. Sep 2018 at 16:40 Xander Wilcke [email protected] wrote:

I happen to have ported your code to the most recent versions of these libraries a few months ago while extending it with the ability to learn from additional features. It evolved a lot since, but I guess you can still see the necessary changes in my old commits https://gitlab.com/wxwilcke/mrgcn/commit/9378791f5b62bf9665a9056a53867f5cd123af31 http://url

— You are receiving this because you commented.

Reply to this email directly, view it on GitHub https://github.com/tkipf/relational-gcn/issues/5#issuecomment-422065019, or mute the thread https://github.com/notifications/unsubscribe-auth/AHAcYClhxu_lOpLUlx2zbUPPIBkf_mufks5ub8KEgaJpZM4Wichm .