Discussion for the proposal

[tngan]

It's published here with current progress. Feel free to take a look and discuss here. If you encounter problem during installation, please create a new issue for that.

https://github.com/node-tensorflow/node-tensorflow/tree/1.0.0
https://github.com/node-tensorflow/node-tensorflow/blob/master/README.md
https://github.com/node-tensorflow/node-tensorflow/blob/master/doc/PROPOSAL_v1.md

[mprentic-harris]

@tngan link gives a 404. I guess you are just referring to the README.md in master?

[tngan]

Thanks @mprentic-harris.
Link is updated. We have merged it back to master.

[peterbraden]

I didn't see any discussion of using SWIG to generate the native nodejs integrations - it seems like this would be a good starting point? Also proposed here as the best way to create an interface: https://www.tensorflow.org/versions/0.6.0/api_docs/index.html

[tngan]

@peterbraden At the very beginning, we have discussed about using SWIG in slack channel. There is a simple C-based client API that we can use. Therefore we decide to get started with a simple one first. What do you think about that ?

[peterbraden]

I think that using the C client to begin with is fine, but I'd suggest trying to use SWIG as soon as possible afterwards. With node-opencv, most of the issues people have with the API are due to the bindings not having support for a specific method etc, it's a lot of work to make them match up. SWIG solves this nicely.

[tngan]

I am not willing to disagree with using SWIG. How do you handle the version of source files in node-opencv ?

[peterbraden]

Sadly OpenCV doesn't use SWIG so we have no way other than to manually track each function. One of my collaborators is exploring generating bindings with libclang, but sadly there isn't anything as simple as SWIG available.

[tngan]

It's a bit complicated in the tensorflow right now. @ivanseidel , one of our collaborators, tried to work on the bindings and track the official version of source files, but I am not sure whether it still works or not.

If we want to use SWIG, we can refer to the python's one. The tensorflow.i is a SWIG wrapper for all of TensorFlow native functionality.

@peterbraden I've added you as the collaborator, feel free to create a new issue and branch for discussion of SWIG interface and get started on it. I am not sure to track the source files manually is a good idea right now, but it's simplest to kickstart. I am also available to participate in.

[jeffwhelpley]

@tngan you mentioned a slack channel. what is the name of that channel? is it open to anyone?