We do not use tensorflow-build anymore at snips, we make our models run on small devices using tract. The scripts here work to compile tf 1.3.1 and have not be tested on later versions
A set of scripts to (cross-)build the Tensorflow C lib for various architectures / OS
The goal of this repo is to provide packages containing the libtensorflow_c.so file as well as a corresponding pc file so that is .so can be used with pkg-config (one use of this is linking tensorflow with rust programs, see the bonus at the end)
We're using these scripts on Ubuntu and Archlinux, the commands given below assume that you're on a fairly recent Ubuntu box.
Theses script are a WIP but a good starting point on how to cross compile tensorflow. Expect to have to edit and tweak them to fit your needs.
Supported OS / Arch:
| OS | Available archs |
|---|---|
| Debian Like | armhf [deb] (raspbian, crosscompiled, tested on Raspberry Pi 0, 2 & 3) amd64 [deb] |
| Android | armeabi-v7a [sysroot overlay] arm64-v8a [sysroot overlay] x86 [sysroot overlay] x86_64 [sysroot overlay] |
| iOS | [sysroot overlay] |
| Archlinux | i686 / x86_64 use this aur [PKGBUILD] |
| macOS | x86_64 brew install libtensorflow |
Before runing these scripts, you need to have bazel
$ echo "deb [arch=amd64] http://storage.googleapis.com/bazel-apt stable jdk1.8" | sudo tee /etc/apt/sources.list.d/bazel.list
$ curl https://bazel.build/bazel-release.pub.gpg | sudo apt-key add -
$ sudo apt-get update
$ sudo apt-get install bazel
You'll also need numpy
$ sudo apt-get install python3-numpy
Here is a small description of interesting scripts in this repo
| Script | Description |
|---|---|
compile.sh |
Clones and build tensorflow for current machine. Takes the tensorflow version as a parameter |
cross-compile.sh |
Generic script for cloning and building tensorflow using a cross toolchain. Launch it without args for usage |
compile-arm.sh |
Clones the raspberry/tools repository an launches a cross compilation using the toolchain in it. Takes the tensorflow version as a parameter |
compile-android.sh |
Download Android NDK and launch a crosscompilation for android using it. Takes the tensorflow version, android arch and package version as parameters |
compile-ios.sh |
Clone and build tensorflow for iOS with the C API. |
create-pkgconfig.sh |
Generic script for creating a pc. Launch it without args for usage |
$ cd debian
$ ./create-deb-native.sh # or ./create-deb-armhf.sh then go grab a cup of coffee
$ sudo dpkg -i target/libtensorflow_1.0.0-snips-5_amd64.deb
You can use tensorflow with your rust projects using the tensorflow crate, its dependency tensorflow-sys will seek the libtensorflow_c.so using pkgconfig and build it if it doesn't find it, which can take some time...
Install the package then cargo build should find the lib \o/
For cross compilation, you need a few more steps, the examples below are for a building for a Raspberry Pi 2/3 from a linux box
First, you need the rust target for the Pi
$ rustup target install armv7-unknown-linux-gnueabihf
Then you need a toolchain for the pi, let's use the one provided by RaspberryPi
$ git clone https://github.com/raspberrypi/tools
Configure cargo to use the toochain by adding this snippet to you ~/.cargo/config
[target.armv7-unknown-linux-gnueabihf]
linker = "/path/to/raspberrypi/tools/arm-bcm2708/gcc-linaro-arm-linux-gnueabihf-raspbian-x64/bin/arm-linux-gnueabihf-gcc"
We now need the .so file of tensorflow for the target arch, let's build the 1.0.0
$ ./compile-arm.sh v1.0.0
The .so is located in target/tensorflow/bazel-bin/tensorflow/libtensorflow.so you may want to move it somewhere else
Let's then generate the .pc file
$ ./create-pkgconfig.sh tensorflow_c /folder/where/the/so/is 1.0.0 "Tensorflow C Library" > /path/to/pc/file
Now that all is done, we can use cargo to crossbuild the app
$ PKG_CONFIG_LIBDIR=/folder/where/the/pcfile/is PKG_CONFIG_ALLOW_CROSS=1 cargo build
Enjoy!