tensorflow-gpu

[ryandawsonuk]

The seldon-core microservice library has tensorflow (cpu version) as a dependency. This leads to either dependency issues if anyone wants to use tensorflow-gpu or being able to build and then getting a 'CPU supports instructions that this TensorFlow binary was not compiled to use' message. This has been encountered before and came up again recently on slack.

We should look at ways to make it easier to use tensorflow-gpu with seldon. Some ideas:

• Will it work to install seldon-core, uninstall tensorflow and then install tensorflow-gpu? Slack discussion currently suggests not but if we did get that working would it be adequate to document that?
• Could we just rely on the tensorflow protos (which is what we really need) without bringing in all of tensorflow?
• Could we use another library to handle detection?
• Provide two versions of seldon-core?

[ukclivecox]

The seldon-core python libraries just need to convert a tensorProto to ndarray but so far not found a way to isolate that code easily and maintainably.

[ukclivecox]

Still hard to see any easy solution to this without creating separating out the tensorflow part.

[ukclivecox]

Fixed - you can specify tensorflow-gpu in python wrapper setup