Cntk: Deploy Model on Mobile

[Piso]

I have been using CNTK and it has been great so far! However, many computer vision tasks are really meaningful on mobile.

I would be very useful to be able to deploy a trained model on iOS/Android/Windows Phone.

[zhouwangzw]

Currently we do not have support for servicing model on mobile platform yet, but it is on our roadmap, and we hope that we can provide it in near future.

[sskgit]

Can anyone provide more details on the availability of mobile option for CNTK?

[fwaris]

I am also interested in embedded and mobile deployment scenarios.

[asvany]

+1

[buptkang]

+1

[JimSEOW]

@buptkang @asvany @fwaris @sskgit @Piso Please come to this discussion for effort to do .NET deep learning through Xamarin

[wolfma61]

there is no official announcment which can be made at this time. .net is disucssed as mentioned above, different other implementations/scenarios are also being evaluated (for example you could take a look at Halide - see https://docs.microsoft.com/en-us/cognitive-toolkit/releasenotes/cntk_2_0_release_notes)

[JimSEOW]

@wolfma61 my background is pharma. The best source of use cases for deep learning for pharma is DeepChem using Tensorflow.

I have been lobbying this community to look into Win 10 OS for the development, instead of LINUX, for many benefits I have presented to them.

Pharma is drown with big data, not just the standard blah blah, but how to bring the whole multi-disciplinary teams, from pre-clinical, clinical, and sale/marketing people dealing with Real World in ONE SINGLE PLACE.

For that, .MS AI strategy MUST NOT depends on CLOUD Only. Democratization of AI as advocated by Microsoft Build2017 must spread to Smart UI on different OS and mobile platforms. Xamarin Form (One GUI to rule all OS, with .NET standard 2.0, .NET core 2.0, and XAML standard) is another MS ammunition to truly democratize AI in ways CNTK no longer has to do any more catch up game.

For Smart Reactive UI, there could be a business model of adapting to user responses and training locally within client [NOT just take an existing trained model from python etc]. .NET client deep learning, not python, not c++ is key to this untouched business opportunity. This is why I am pushing .NET deep Learning to be first class as python.

[wolfma61]

I didn't want to appear disagreeing with you or not taking your serious ...
all the thing you mention, but in this topic mobile or client side intelligence, learning, and inference are important to me, the team and Microsoft. This is actively being discussed, I don't think we can make an announcement yet what CNTK will address next in this area. But it is going to get some love ;)

[JimSEOW]

@wolfma61 Some LOVE from you will MEAN so much for us :-)

[peachthiefmedia]

+1

[douglas125]

Agree:

I would be very useful to be able to deploy a trained model on iOS/Android/Windows Phone.

Or at least give us a simple way to host a trained model in the cloud (using Azure Functions would be perfect).