RE Tensorflow

[idata]

I know this has been brought up a few times, but what are the plans for Tensorflow, personally it has been a long time since I used Caffe and would be a lot of extra work and effort to refamiliarise myself with it enough to reproduce my projects. I also feel that you would achieve a lot more interest supporting Tensorflow, I do not really know anyone that uses Caffe anymore. What is the likely hood of this being put in place?

[idata]

@AdamMiltonBarker We currently support only Caffe, but based on developer feedback we are evaluating other frameworks; Tensorflow being one of them.

[idata]

@Tome_at_Intel Would you have a plan of open source the mvNCCompile.pyc? I think that will make more and more developers transport their own models which created by other frameworks( like tf, torch, mxnet… ) to the graph format.

[idata]

@arlose At this moment, we don't have any plans to open source mvNCCompile.pyc, but what you are suggesting makes sense. Thanks for the input!

[idata]

Hi @Tome_at_Intel sorry for delayed reply, I have received the Movidius today. I notice Tensorflow is available but after the installation process told me to install Stretch on the Pi and doing so, it tells me Tensorflow is not supported ?

[idata]

Apparently this should work, I will try install tomorrow, if not will shift to Ubuntu

https://github.com/lhelontra/tensorflow-on-arm/releases/tag/v1.3.1

[idata]

I can confirm that you can install the above Tensorflow on Stretch, but so far unable to get TF working with Movidius on the RPI 3

https://ncsforum.movidius.com/discussion/628/issues-with-raspberry-pi-tensorflow#latest

[idata]

@AdamMiltonBarker Were you able to get TensorFlow working on the Raspberry Pi?

[idata]

but so far unable to get TF working with Movidius on the RPI 3

@AdamMiltonBarker , you should easily be able to run Tensorflow based Movidius graph files on the NCS, here's how:

There are 2 parts to NCSDK:

1. Toolkit - Consists of mvNCProfile, mvNCCompile, mvNCCheck and other helper libraries. These tools are used to profile, convert and validate (respectively) either a Caffe or Tensorflow based model.
2. API framework - This is primarily mvncapi, which should be used to deploy a Movidius graph file onto the NCS.

Unless you want the capability to profile, compile and validate TensorFlow models on the RPI, you are not really blocked. Try this:

NOTE: You don't need Tensorflow installed on NCS, to run the below steps.

1. Install NCSDK on a laptop/desktop - https://developer.movidius.com/start
2. Clone ncappzoo on the same laptop/desktop - git clone https://github.com/movidius/ncappzoo
3. Build a tensorflow model on the same laptop/desktop - cd ncappzoo/tensorflow/mobilenets && make
4. Step 3 generates a Movidius graph file; copy ncappzoo/tensorflow/mobilenets/graph from the laptop onto a pendrive
5. Install NCSDK on your RPi - https://developer.movidius.com/start (or choose to install just the API framework - https://movidius.github.io/blog/ncs-apps-on-rpi/)
6. Clone ncappzoo on your RPi - git clone https://github.com/movidius/ncappzoo
7. Move the Movidius graph file from your pendrive to the RPi's ncappzoo/tensorflow/mobilenets folder
8. Open a terminal window on your RPi and run these 2 commands - cd ncappzoo/tensorflow/mobilenets and python3 run.py

[idata]

Hi thanks guys, I was able to get it all working following the method I describe in the tutorial, @Tome_at_Intel yes I did buddy it is on the tutorial below I made:

https://github.com/AdamMiltonBarker/IoT-JumpWay-Intel-Examples/tree/master/Intel-Movidius/TASS

@Tome_at_Intel I am at the final part of the project now, is there any docs anywhere regarding saving a custom inceptionv3 model (.pb) for use with Movidius.

[idata]

These are the steps I had taken trying to get tensorflow working, I made a new post about it:

1. Originally I was on Raspian Jessie, after getting most of the way through the installation it fails and says to upgrade to Stretch.
2. Next I installed on Stretch, it says Tensorflow is not supported in stretch and installs without Tensorflow.
3. Checking the (Movidius) website I saw Tensorflow is only supported in Ubuntu LTS, after installing on Ubuntu LTS it fails installing Tensorflow due to there not being a version of 1.4.0 for RPI, installing 1.3.0 does not work.
4. Next I tried installing Stretch by making the api only ~/workspace/ncsdk/api/src and installing https://github.com/lhelontra/tensorflow-on-arm/releases/tag/v1.3.1 this works but when I come to run make on the example I made it errors with issues about mvNCCompile
5. Next, having already got TF on the device, I just ran the normal install (This worked)

https://ncsforum.movidius.com/discussion/628/issues-with-raspberry-pi-tensorflow#latest

I now understand the steps that @AshwinVijayakumar gave would resolve issue 4 but after completing 5 there was no longer an issue.