Intel NCSM2450.DK1 Movidius Neural Compute Stick

Running this on VMware with Ubuntu 16.04.The movidius github samples are working just fine.On vmware, make sure to add usb controller and specify usb 3.0 compatibility.Also, note the device defaults as a usb 2.0 device but switches to usb 3.0 when being used to run inferences/predictions.Don't be surprised if you see this behavior.Vmware will notify you of disconnects and reconnects when using the device so just allow it. When you encounter it, disable the prompting as it may interfere with the operation of the device -- causing time-out.Note that you may need to allow some delay before running the sample codes. I get occassional error that the device is not present or stale connections. Retrying seems to solve the issue. Or reconnecting the usb to the VM if it didn't do so automatically.I had one case when running the multistick sample code that caused the devices to be not detected even restarting the VM did not help. For that case, I rebooted my machine which help reset everything back to normal.This was said to work on RPI 3, but I was hoping to try this on another SBC(RPI alternative) then integrate it to my custom SBC cluster for docker swarm.

The bad news first:1. The device is only usable on a a Raspberry Pi or a PC running Ubuntu 16.04 LTS. Some people have reported success accessing it from virtual machines running Ubuntu 16.04 LTS but I have not been able to make that work repeatably.2. The device and its SDK currently support a limited subset of the caffe and TensorFlow frameworks. The device itself is quite general, but the software is geared towards state-of-the-art computer vision algorithms.3. The device and SDK currently only do inference, not training.4. The models run with 16-bit floating point for speed.The good news:1. You can host it on a Raspberry Pi!2. Low power - it gets it's power from the USB port.3. Very fast - the examples that ship with it are well-known deep learning test cases and they run on the device in milliseconds!4. It works out of the box if you have the hardware and operating system required.5. There is a growing collection of pre-trained working open-source models (a "model zoo").Summary: if you want to learn state-of-the-art computer vision algorithms and put them to work, this is the device you need. I'm hoping Intel will open up the SDK and on-device software and expand the generality of the tools for other applications.

I originally thought this had a lot of potential, however after playing around with it I found it has a lot of limitations.The biggest issue is that its SDK is not mature. You can run the examples provided and they work, but once you try to do something of your own things get complicated quickly. You end up searching all over the web for examples due to the lack of documentation. If you are someone who enjoys hacking around this isn't a problem, but if you are someone who is used to purchasing something that is production ready then I would recommend saving your money. I don't look for this product to be around long. This is an intel product and its pretty obvious theres not a lot of effort being put into it.Also you pretty much have to use Linux or a VM with linux on other OS's. Not that I mind but I think it should be clear to anyone else who is going to purchase this product.Overall I wish this product was better. I was very hopeful, but after about a week of trying to get my custom tensorflow model work on it I am disappointed at the level of effort to takes to do something simple.

I really wanted to have fun with this. The getting started guide looked simple. The video looked cool. Then I bought it. The device only seems to work with one specific version of Ubuntu. The getting started guide (looks so simple) leaves out numerous details of version conflicts of libraries (e.g. numpy), additional install requirements, etc. The forums don't have a lot of help around current distributions. The example applications don't appear to work with the native rPi camera. Overall, it's like the platform is stuck a year or two ago in the past. Hey, I got the hello world program to talk to the stick! But that's about it. I am not the sharpest of the bunch by any means, but I would expect official getting started guides to work. Overall, a very poor experience.Edit: Just to update, I did a second environment with Ubuntu 16.04 VM, new USB camera, etc. Similar results. nothing works. Little in the forums. Avoid unless you like a cool looking blue USB stick. Edit #2: So I thought I would be clever and bought the Neural Compute Stick 2. I figured as a recent device, it would have better support. WRONG. It does not even work on ARM yet. So for "deep learning" computing on the edge, you have to have a full Intel device dedicated to it.

Worked great on my desktop connected to a VMware workstation virtual machine running Ubuntu 16.04 LTS. It does disconnect and reconnect when compiling and running examples, but this appears to be from the application enabling and disabling the interface via open/close commands to the interface platform. You do need to maintain the guest virtual machine as the primary focus or it will reconnect to the host machine when it closes the interface. There are examples where you can run the multi interface setup, but this requires 2 of these devices to work properly.