Turicreate: Turi to CreateML

[alelordelo]

Is it possible to:

1- Create a Model in Turi
2- Export as CoreML
3- Load in CreateML
4- Keep training in CreateML

This workflow would be amazing!

Cheers

[alelordelo]

hey guys, any news on this?

[hoytak]

Hey @alelordelo,

Currently, CoreML models don't support training -- they're for inference only. However, we're interested in supporting something like this workflow at some point. What's your specific use case? Which models would you use?

[alelordelo]

Thanks @hoytak!

I'm interested in vid2vid:
https://github.com/NVIDIA/vid2vid

The idea would be to convert vid2vid to CoreML, and then train the exported CoreML model with my own datasets using CreateML.

What you think? This kind of workflow make sense?

Cheers

[mfriedel]

Hey @alelordelo. I don’t work for Turi (though definitely a fan of the people + the library), but I do spend a lot of time thinking about on-device machine learning, and I’m always interested to learn more about how people build models. I’m curious about your proposed workflow—if you’ve been able to build a model using Turi and export to CoreML, what is your goal with switching to CreateML? Is there a reason you couldn’t load your data into a Jupyter Notebook and continue to use Turi there to refine and/or build a custom model there?

[alelordelo]

@mfriedel, what I wanted is to use an existing coreML (created with Turi, Tensroflow, etc) model, and train with my own datasets. But as @hoytak mentioned, coreML doesn't currently support training.,,,