TensorFlow C binding for Raspberry Pi

[marianopeck]

System information

• TensorFlow version (you are using): 1.13.1
• Are you willing to contribute it (Yes/No): yes, at least as a tester

Describe the feature and the current behavior/state.
I asked on StackOverflow if there was a TensorFlow C binding for ARM/RaspberryPi but I got no answer after many days.

I couldn't find anything on the website, so I suspect this is not yet available?

If it is just a matter that you don't distribute the binaries, I have no troubles to compile it myself if I have the instructions and if it is known to work. But I haven't found anything confirming that either.

Will this change the current api? How?
No

Who will benefit with this feature?
All language bindings that use the C library and want to run on any ARM board

Any Other info.

[freedomtan]

The last paragraph of the "TensorFlow for C" you cited points to a how-to file which does show you how to build TensorFlow C API from source code with bazel

[marianopeck]

Thanks @freedomtan. I didn't know if something special compilation was needed for ARM. Are you aware of people/projects compiling it and using successfully?

[freedomtan]

@marianopeck something like

CC=clang-6.0 CXX=clang++-6.0 \
bazel build --config opt --local_resources 1024.0,0.5,0.5 \
--copt=-mfpu=neon-vfpv4 \
--copt=-ftree-vectorize \
--copt=-funsafe-math-optimizations \
--copt=-ftree-loop-vectorize \
--copt=-fomit-frame-pointer \
--copt=-DRASPBERRY_PI \
--host_copt=-mfpu=neon-vfpv4 \
--host_copt=-ftree-vectorize \
--host_copt=-funsafe-math-optimizations \
--host_copt=-ftree-loop-vectorize \
--host_copt=-fomit-frame-pointer \
--host_copt=-DRASPBERRY_PI \
//tensorflow/tools/lib_package:libtensorflow

works for me.

[ravikyram]

@marianopeck Just to verify did you get a chance to follow instructions as suggested by freedomtan .Thanks!

[marianopeck]

Hi @ravikyram
Not yet. I will see if I can try it this week.
BTW, if there is a compilation instruction and people claim it works, any reason why the binaries are not officially distributed as the rest of the "TensorFlow for C"?

[marianopeck]

Hi @freedomtan
Which version did you use? It's 2.0? Or older too? I would like to try with 1.13.1 or similar.
Thanks

[freedomtan]

@marianopeck I use master branch most of the time. I think it works on TensorFlow versions before 1.10 to current master.

[marianopeck]

Hi @freedomtan
Thanks, I am trying now to compile it myself but I wasn't even able to compile bazel on the first place :( I am getting this issue: bazelbuild/bazel#8882
If I am able to pass that, I will continue..

[freedomtan]

@marianopeck I haven't tried newer bazel, general comment is to increase your swap space and maximum Java heap size (e.g., -J-Xmx1024m instead of -J-Xmx512m)

[marianopeck]

Hi, as I found no binary available, I had to manage myself.
I finally took the time to write down everything I found during my attempt to get TensorFlow C library compiled for Raspberry Pi. In that tutorial, I point back to this issue in case any progress is made.
Thank you

[freedomtan]

good to see another effort on Smalltalk + TensorFlow. Smalltalk is one of my favorite language. I tried Pharo bindings for TensorFlow before and was able to use both TensorFlow C API and TensorFlow Lite C API on x86 machines

[marianopeck]

Hi @freedomtan
We are in the same boat then :) That's great to hear. The Pharo bindings are actually a port made by @SergeStinckwich from the work done on Cuis Smalltalk. The latter was mostly done by @gerasdf and he is now doing most of the port to VASmalltalk.
I am interested in hearing about your experiments. To avoid making noise in this issue, I will send you an email.
Thanks!

[SergeStinckwich]

hi @marianopeck, @freedomtan already have done some contributions to the Pharo port.
Yes all on the same boat, let's try to go in the same direction :-) Let's discuss about that in ESUG 2019 ?

[marianopeck]

OK, I sent you both and the rest of the involved parts a private email to so that we can continue discussing :)

[PINTO0309]

Although I do not have C language implementation skills, I tried to generate a binary for RaspberryPi of Tensorflow v1.14.0. I have not confirmed the operation.
https://github.com/PINTO0309/Tensorflow-bin.git
https://github.com/PINTO0309/Tensorflow-bin/tree/master/C-library/1.14.0-armv7l

[marianopeck]

That worked perfectly!!