MORP: Movidius On Raspberry Pi This repo about using Movidius neural compute sticks(NCS) on the Raspberry Pi.
It is the ongoing result of a series Meetups I have been hosting for Chicago Area Embedded Deep Learning: https://www.meetup.com/Chicago-Area-Embedded-Deep-Learning
As of 12/19/17 Movidius is saying you can have the NCS or TensorFlow, but not both:
- Build directions for a microSD card image for a Raspberry Pi with a Movidius Neural Compute Stick(NCS)
- Link to 6GB SD Card Image - RPI with NCS root password: 45tom*
- Build directions for a microSD card image with Raspberry Pi with Tensorflow
Slides from a very short demo I gave:
FAQ: What are the Movidius links?:
How do I use the Pi Camera:
What hardwhere do I need to get into this?
- Really you need a Ubuntu 16.04 PC, RPi, and NCS.
Does it have to be Ubuntu 16.04 PC?
- Only if you value your time! Preferably it should be a modern PC with a GPU running Ubuntu directly(not virtual). This has been a major stumpling block both in the meetup group and on the movidous forum.
What does it do?
- Infers(runs the computational graph) a Tensorflow or caffe model.
What are some NCS+RPI pros and cons?
- Power usage(2 Watts NCS+RPI)
- Size and wieght
- It's the first ASIC out so there is nothing to compare it to at this point
- It's true edge device (no internet needed & appliance like)
- RPI is very popular
- 15x quicker than RPI along on Deep Learning inferance.
- First one out: so limited support & could have much more power
- You can't access the Myraid 2 VPU with all the computer vision functionality
- Cost is high for what your getting (It's less than one thread on a CPU)
- If power and size are not an issue an old PC is more cost effective and easier to use.
How does the NCS compare to the Google AIY Vision Kit?
- Not sure yet, I ordered three of them
- The Google AIY Vision kit is a hat/bonnet that sits on top of a RPi Zero
- The Vision Kit is a hobby/learning product only, becuase you can't buy RPI Zeros in quanity and need to solder the headers on(correctly) to use it.
**More FAQ's to Come
Thanks! Paul Krush 630-830-6640