SOLVED: ARM / Raspberry Pi install guide and inference benchmarks

[idata]

I would love to see an ARM / Raspberry Pi install guide and inference benchmarks. Anyone have a link for this? Thanks! Paul Krush

[idata]

I am working on an autonomous vehicle with a Raspberry Pi at the brain, and I would love to use this to apply deep learning to it. Will keep an eye on this thread.

[idata]

me too, same hw solutions but for health problems

[idata]

I am looking to do small part sorting. Here is a current rig of mine: https://www.youtube.com/watch?v=cwcn3tuXrv4

I ordered 3. Getting a speed using LeNet will be my first task. Working on the RPi will be next.

I might do a workshop/meetup on them in the Chicago area.

[idata]

What is the status on ARM processor ?

SDK require an X86, but deploiement can be done on an USB3 capable ARM, like the ODROID XU4 ?There is a lot of confusion about this at the moment.

[idata]

HI to all, I came here with the same doubt, If it's possible to use it with the Raspberry Pi, with my cNET I reach a time of classification of 161ms per image however I would like to use the Raspberry Pi in a real-time application

[idata]

Hi all,

The SDK that is provided with the Neural Compute Stick currently provides support for Ubuntu 16.04 Desktop (x86 64bit) Support for other operating systems may be added. I would urge everyone interested in NCS support on other platforms to check back in the coming weeks :-)

Neal

[idata]

@neal, that would be really helpful to get support for ARM/RPi. Including me, many have already some development in place with RPi. Please provide support for RPi/ARM.

[idata]

@neal I'm on Arch Linux - I took a look at the install script and it looks like it would be possible but tedious to port from Ubuntu 16.04 to Arch. The real gotcha is going to be GPU drivers for training - NVidia GPU drivers are proprietary and getting them to work on anything untested is hacky and poorly supported. They do exist in the Arch User Repository, but I think it's a waste of time.

[idata]

Ubuntu 16.04 Desktop (x86 64bit)

Does this mean it should work on the LattePanda which has an Intel Atom x5-Z83004GB RAM, 64GB?

[idata]

Yeah but the RPI isn't an industrial board (isnt ISO temp rated).

So its OK for development but not much else. I am hoping for TensorFlow Object Dectection API support which can be loaded on several platforms including windows. Getting YOLO results on any low powered device would be incredible.

[idata]

@chicagobob123, I'm able to run the entire MvNC SDK suite (profiling, checking, compiling, API framework) on my HP pavilion 10-n113dx with 2GB RAM and 32GB storage. You shouldn't have any issues running the SDK on the LattePanda as long as Ubuntu 16.04 is stable on it.

Please do share your results if you try it out.

[idata]

@Shailesh @josejacomeb @blakesresearch @pkrush @chicagobob123 @patrickpoirier @abhishekkd

FYI: https://ncsforum.movidius.com/discussion/118/movidius-nc-sdk-1-07-07-with-raspberry-pi-support

[idata]

Support for RPi has been released here https://ncsforum.movidius.com/discussion/118/movidius-nc-sdk-1-07-07-with-raspberry-pi-support#latest

[idata]

Cheers, if anyone gets sooner to a benchmark with a RPi than me, don't hesitate to post :)

[idata]

Also are you aware of any other RPI + Caffe + Squeezenet install tutorial beside this one? http://cv-blog.ru/?p=105

[idata]

@soobrosa I use Caffe's Ubuntu installation instructions (http://caffe.berkeleyvision.org/install_apt_debian.html) on RPI Jessie, it works fine for me. Try this if you need OpenCV installation instructions - http://www.pyimagesearch.com/2016/04/18/install-guide-raspberry-pi-3-raspbian-jessie-opencv-3/. (skip the virtualenv part, and install OpenCV natively)

NOTE: You don't really need Caffe to run NCS on RPI. See "Important Notes" section on https://ncsforum.movidius.com/discussion/118/movidius-nc-sdk-1-07-07-with-raspberry-pi-support for more details.

PS: Look forward to your benchmark results!

[idata]

I imagine that since this works for Debian that it should work on the ODROID line as well? Like the

ODROID-XU4? It has USB 3.0 connections more ram faster CPU etc. Or is there a tie to the hardware?

[idata]

@chicagobob123 We recently released MvNC SDK for RPI 3, which is ARMv8 architecture. The Odroid-XU4 is based on ARMv7 architecture, so there's no guarantee it'll work on the Odroid-XU4.

[idata]

@pkrush I'd like to close this thread as solved since @neal gave you the RPI installation instructions you originally asked. Please feel free to start a new discussion if you run into issues running NCS on RPI.

[idata]

@chicagobob123 please try the RPI package on your Odroid-XU4, and start a new discussion if you run into issues with it.

@soobrosa I look forward to your benchmarking results on RPI, please feel free to start a new discussion when you have the results.