install tensorflow == 1.4.0

[idata]

Tried the following on a raspberrypi3 to obtain a full SDK installation

1) installed ubunuMate. Result Release: 16.04.4 LTS code: xenial installed without complaint2) ran script to a)git clone https://github.com/movidius/ncsdkb)sudo make install3) Errors reported as follows:No matching distribution found for tensorflow==1.4.0You are using pip version 8.1.1 however version 9.0.1 is available4)pip -V returns 9.0.1pip2 -V returns 9.0.1pip 3 -V returns 8.1.1Presumably pip3 is executing. Any help to resolve this appreciated. I tried several apt get install methods to update pip2 and pip35) Movidius environment spec mentions 64 bit machines. My Rasperrypi3 is 64 bitbut does the Ubuntu installation run in 64 bitmode and does this matter?6) Can I build the ncsdk on a 32bit intel pc or must it be a 64 bit pc running Ubuntu (which flavour?)

[idata]

Hi Again. I tried a clean cdrom install on an old intel32bit with Ubuntu16.04 and proceeded with the ncsdk as before. Result exactly the same error as per my attempt above using the Rasperrypi3 setup.

32/64 bit problem? Further clarification very much appreciated!

[idata]

Hello, @jslawton

Updating pip.

wget https://bootstrap.pypa.io/get-pip.pysudo python3 get-pip.pysudo rebootsudo python2 get-pip.pysudo reboot

Install Tensorflow 1.4.0.

wget https://github.com/lhelontra/tensorflow-on-arm/releases/download/v1.4.0/tensorflow-1.4.0-cp35-none-linux_armv7l.whlsudo pip3 install tensorflow-1.4.0-cp35-none-linux_armv7l.whlsudo reboot

[idata]

Hi Pinto

Thank you for your help. Tensorflow is now installed and the 'sudo make install' has proceeded OK after detecting the tensorflow installation. Note I needed to execute sudo with the -H flag as suggested by the install process. Many thanks

[idata]

Hi . The Raspberyrpi3 Ubuntu 16.04 did not finish the ncsdk make install. It reached the Caffee Build and hangs with the memory usage showing 100% ( and of course noswap file available!).

1) I guess this means the complete install cannot be achieved on Raspberrypi3 without a hard disk to provide a swap file support.2) I retried the install method on my old intel32bit Ubuntu 16.04 machine and it fails as previously to find Tensorflow == 1.4.0.Any help with a suitable Tensorflow install for intel32bit please.3) I have now installed installed Virtualbox on an ASUS I3 (intel64bit) Laptop running Windows10. Please can anyone recommend the correct Ubuntu 16.04 64Bit installation DVD iso as I can only find 64 bit releases for AMD devices.

[idata]

Hi, @jslawton

I think that setting up SWAP is essential.

And at least 2GB.I'm sorry if I could not understand the meaning of the question.

sudo apt install dphys-swapfile

sudo rebootsudo nano /etc/dphys-swapfile CONF_SWAPSIZE=2048

sudo /etc/init.d/dphys-swapfile restart swapon -s

free -h

[idata]

Hi again Pinto

Thank you for your observations on the necessity of having a SWAP file. I have taken your key point on board but was wary to allocate swapfile space on my USB memory stick because of its likely failure with repeated write cycles as a flash memory device. I tried an old external USB drive and used instruction as per the following appended reference.The result for 'sudo make install ' is 'Setup is complete'.Sorry I did not make my other points clear in my previous comment.To reiterate:1) The Raspberry Pi 3 cannot install the NC SDK under Ubuntu 16.04 without a SWAP file. The SWAP file mechanism should ideally be implemented on an external (USB) hard disk.

2) I have also tried to install the NCSDK under Ubuntu 16.04 on an old Intel 32bit base Unit. Even with your pip fix this still returned the original error

'No matching distribution found for tensorflow==1.4.0' and so I am curious to know where to locate the correct tensorflow distribution.

3) I am also trying to install the NC SDK within VirtualBox on a Windows10 ASUS I3 64 bit environment. I cannot find an Ubuntu 16.04 (64Bit) iso distribution. There are distros explicitly listed for AMD . Am I to infer they are inherently compatible with Intel 64 bit cores?

Note. Since I understand my quad core 64bit Raspberry Pi3 is actually running in 32 bit mode under Ubuntu 16.04 and I now have the NC SDK installed perhaps the 64 bit concern ( implied in the overall Movidius laptop spec) is not an absolute requirement.Do hope this clarifies for you.Many thanks indeed for your help.

https://www.digitalocean.com/community/tutorials/how-to-add-swap-space-on-ubuntu-16-04

[idata]

Hi, @jslawton

Thank you for the detailed explanation.

I understood the situation well.I will investigate. However, it may take time.

[idata]

3)There are distros explicitly listed for AMD . Am I to infer they are inherently compatible with Intel 64 bit cores?

Probably, there is almost no problem. The difference is a little.

I am looking at it gradually between jobs, so please wait for other answers a bit.

[idata]

Hello, @jslawton

Sorry, the pip3 command I introduced for installing tensorflow1.4.0 was for the Arm architecture.

About 2).

Ubuntu 16.04 32bit/64bit

https://www.tensorflow.org/install/install_linux#InstallingNativePip

$ sudo -H pip3 install --upgrade tensorflow # for Python 3.n

or$ sudo -H pip3 install --upgrade tensorflow-gpu # for Python 3.n and GPU

Or what about the following?

$ sudo -H pip3 install --upgrade https://storage.googleapis.com/tensorflow/linux/cpu/tensorflow-1.4.0-cp35-cp35m-linux_x86_64.whl

[idata]

Hi Pinto

Re point 3) and your reply. I have tried a Win10 VirtualBox Ubuntu 16.04 iso installation followed by theMovidius SDK installation. The make examples phase returned error -7 failing to find USD device.I then installed the 'guest extensions' from Virtual Box.org to provide the USB filters. My mistake here was this install was out of the order prescribed in the Movidius Docs.The hello world app then failed to detect the Movidius Stick no matter how the device was selected from the Virtual Box menus. Neither lsusb or dmesg reported the Movidius device though the Virtual Box did annunciate the device and its serial number. A Movidius Loop Back device was also (sometimes) reported.Any thoughts on need to install 'guest extensions' directly after the DVD iso install.Possibly install the latest version of Virtual Box, 5.2.8 ? I installed the recommended VirtualBox.org VitrualBox 5.1 build and the Extensions Pack listed for all platforms.

Re point 1) I tried The Raspberry Pi3, Ubuntu 16.04 with swap file on External harddrive. The installation proceeded until

failing the Opencv installation that was specified by the make examples phase. I used a build opencv from sources with contrib . This took several hours. After completion and again making examples I rebooted with the Movidius stick installed. the hello world app confirmed the device working and dmesg reports the device presence. I also intend to complete the Raspian Stretch installation with the SDK app on another 16MB SDcard.

Re point 2) I have abandoned the intel32bit installation as the other two installations now look more promising.

Thank you again for your help. Further comments appreciated

[idata]

Good morning. @jslawton

I think that England is late at night.

Installing OpenCV, it will take time and frustrating.

At raspberryPI + OpenCV, Express installation procedure.

https://ncsforum.movidius.com/discussion/678/ncsdk-installation-improvements-we-want-your-feedback

Neither lsusb or dmesg reported the Movidius device though the Virtual Box did annunciate the device and its serial number. A Movidius Loop Back device was also (sometimes) reported.

Unfortunately the same problem happens with 5.2.8. I myself have encountered the same problem. As for Virtulbox, exactly the same discussion has occurred, how about seeing this?

https://ncsforum.movidius.com/discussion/684/problem-with-movidius-neural-compute-stick

[idata]

Good morning Pinto ( in California?)

After reading your reference above to neural compute stick problems I experimented withselecting TWO USB filters ( I do not understand why two ).With settings much as below I have now managed to have the hello world program work!

Machine>Settings>USB

a) selected USB 3.0 (xHCI) Controllerb) USB Device Filters Selected are:Movidius Ltd. Movidius MA2X5X [001]Movidius VSC LoopbackDevice [0100]

Execution of the make run on the hello_ncs_app directory

returns: |ncs device workingdmesg returns:Product: Movidius MA2X5XManufacturer: Movidius Ltd.Serial Number: 03e72150

Thank you again for your help.

[idata]

Congratulations! @jslawton

It may be unnecessary information, but Tome_at_Intel tells why we need two filters.

https://ncsforum.movidius.com/discussion/comment/1870/#Comment_1870

Japan is advanced by 9 hours by time difference.

[idata]

Successfully identified a Coffee Mug and an oscilloscope generally using the

first example pgm with very helpful walkthru as described via the link.

https://www.rs-online.com/designspark/ai-powered-identification-with-the-pidentifier

Thank you for the help!

[idata]

long time no see, @jslawton

It is polite and a great article.

Thank you for introducing it to me.I had no idea to use the OLED module.It seems to be very good from the viewpoint of compactness and low power consumption.By the way, although the following link is RaspberryPi + motion analysis + object detection using Caffe base MobileNet-SSD, it is faster than the standard sample program of NC App Zoo.

https://github.com/PINTO0309/MobileNet-SSD

[idata]

Thank you for the link. Jaw dropping! I have been reading S.Haykin and C.Bishop and realise I have a 'some way' to go with my studies/learning before I can harvest all the insites from that work!

[idata]

Thank you very much for post on Tensorflow. It's very simple to understand and start with Tensorflow.

Install Tensorflow on windows

[idata]

Thanks, man. Recently I came across the following quote:

Data Science produces insights

Machine Learning produces predictionsArtificial Intelligence produces actions