Permalink
24d5473 Jan 19, 2018
3 contributors

Users who have contributed to this file

@samjabrahams @grassgit @pbabbicola
380 lines (270 sloc) 13.9 KB

Building TensorFlow for Raspberry Pi: a Step-By-Step Guide

Back to readme

What You Need

  • Raspberry Pi 2 or 3 Model B
  • An SD card running Raspbian with several GB of free space
    • An 8 GB card with a fresh install of Raspbian does not have enough space. A 16 GB SD card minimum is recommended.
    • These instructions may work on Linux distributions other than Raspbian
  • Internet connection to the Raspberry Pi
  • A USB memory drive that can be installed as swap memory (if it is a flash drive, make sure you don't care about the drive). Anything over 1 GB should be fine
  • A fair amount of time

Overview

These instructions were crafted for a Raspberry Pi 3 Model B running a vanilla copy of Raspbian 8.0 (jessie). It appears to work on Raspberry Pi 2, but there are some kinks that are being worked out. If these instructions work for different distributions, let me know! Updated (2017-09-11) to work with the latest (HEAD) version of tensorflow on Raspbian Strech (Vanilla version september 2017) and Python 3.5.

Here's the basic plan: build a RPi-friendly version of Bazel and use it to build TensorFlow.

Contents

  1. Install basic dependencies
  2. Install USB Memory as Swap
  3. Build Bazel
  4. Compiling TensorFlow
  5. Cleaning Up
  6. References

The Build

1. Install basic dependencies

First, update apt-get to make sure it knows where to download everything.

sudo apt-get update

Next, install some base dependencies and tools we'll need later.

For Bazel:

sudo apt-get install pkg-config zip g++ zlib1g-dev unzip

For TensorFlow:

# For Python 2.7
sudo apt-get install python-pip python-numpy swig python-dev
sudo pip install wheel

# For Python 3.3+
sudo apt-get install python3-pip python3-numpy swig python3-dev
sudo pip3 install wheel

To be able to take advantage of certain optimization flags:

sudo apt-get install gcc-4.8 g++-4.8
sudo update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-4.8 100
sudo update-alternatives --install /usr/bin/g++ g++ /usr/bin/g++-4.8 100

Finally, for cleanliness, make a directory that will hold the Protobuf, Bazel, and TensorFlow repositories.

mkdir tf
cd tf

2. Install a Memory Drive as Swap for Compiling

In order to succesfully build TensorFlow, your Raspberry Pi needs a little bit more memory to fall back on. Fortunately, this process is pretty straightforward. Grab a USB storage drive that has at least 1GB of memory. I used a flash drive I could live without that carried no important data. That said, we're only going to be using the drive as swap while we compile, so this process shouldn't do too much damage to a relatively new USB drive.

First, put insert your USB drive, and find the /dev/XXX path for the device.

sudo blkid

As an example, my drive's path was /dev/sda1

Once you've found your device, unmount it by using the umount command.

sudo umount /dev/XXX

Then format your device to be swap:

sudo mkswap /dev/XXX

If the previous command outputted an alphanumeric UUID, copy that now. Otherwise, find the UUID by running blkid again. Copy the UUID associated with /dev/XXX

sudo blkid

Now edit your /etc/fstab file to register your swap file. (I'm a Vim guy, but Nano is installed by default)

sudo nano /etc/fstab

On a separate line, enter the following information. Replace the X's with the UUID (without quotes)

UUID=XXXXXXXX-XXXX-XXXX-XXXX-XXXXXXXXXXXX none swap sw,pri=5 0 0

Save /etc/fstab, exit your text editor, and run the following command:

sudo swapon -a

If you get an error claiming it can't find your UUID, go back and edit /etc/fstab. Replace the UUID=XXX.. bit with the original /dev/XXX information.

sudo nano /etc/fstab
# Replace the UUID with /dev/XXX
/dev/XXX none swap sw,pri=5 0 0

Alright! You've got swap! Don't throw out the /dev/XXX information yet- you'll need it to remove the device safely later on.

3. Build Bazel

To build Bazel, we're going to need to download a zip file containing a distribution archive. Let's do that now and extract it into a new directory called bazel:

wget https://github.com/bazelbuild/bazel/releases/download/0.4.5/bazel-0.4.5-dist.zip
unzip -d bazel bazel-0.4.5-dist.zip

Once it's done downloading and extracting, we can move into the directory to make a few changes:

cd bazel

Before building Bazel, we need to set the javac maximum heap size for this job, or else we'll get an OutOfMemoryError. To do this, we need to make a small addition to bazel/scripts/bootstrap/compile.sh. (Shout-out to @SangManLINUX for pointing this out..

nano scripts/bootstrap/compile.sh

Move down to line 117, where you'll see the following block of code:

run "${JAVAC}" -classpath "${classpath}" -sourcepath "${sourcepath}" \
      -d "${output}/classes" -source "$JAVA_VERSION" -target "$JAVA_VERSION" \
      -encoding UTF-8 "@${paramfile}"

At the end of this block, add in the -J-Xmx500M flag, which sets the maximum size of the Java heap to 500 MB:

run "${JAVAC}" -classpath "${classpath}" -sourcepath "${sourcepath}" \
      -d "${output}/classes" -source "$JAVA_VERSION" -target "$JAVA_VERSION" \
      -encoding UTF-8 "@${paramfile}" -J-Xmx500M

Finally, we have to add one thing to tools/cpp/cc_configure.bzl - open it up for editing:

nano tools/cpp/cc_configure.bzl

Place the line return "arm" around line 133 (at the beginning of the _get_cpu_value function):

...
"""Compute the cpu_value based on the OS name."""
return "arm"
...

In newer Bazel versions, the _get_cpu_value function will be found in the file tools/cpp/lib_cc_configure.bzl, and the same modification is required.

Now we can build Bazel! Note: this also takes some time.

sudo ./compile.sh

When the build finishes, you end up with a new binary, output/bazel. Copy that to your /usr/local/bin directory.

sudo cp output/bazel /usr/local/bin/bazel

To make sure it's working properly, run bazel on the command line and verify it prints help text. Note: this may take 15-30 seconds to run, so be patient!

bazel

Usage: bazel <command> <options> ...

Available commands:
  analyze-profile     Analyzes build profile data.
  build               Builds the specified targets.
  canonicalize-flags  Canonicalizes a list of bazel options.
  clean               Removes output files and optionally stops the server.
  dump                Dumps the internal state of the bazel server process.
  fetch               Fetches external repositories that are prerequisites to the targets.
  help                Prints help for commands, or the index.
  info                Displays runtime info about the bazel server.
  mobile-install      Installs targets to mobile devices.
  query               Executes a dependency graph query.
  run                 Runs the specified target.
  shutdown            Stops the bazel server.
  test                Builds and runs the specified test targets.
  version             Prints version information for bazel.

Getting more help:
  bazel help <command>
                   Prints help and options for <command>.
  bazel help startup_options
                   Options for the JVM hosting bazel.
  bazel help target-syntax
                   Explains the syntax for specifying targets.
  bazel help info-keys
                   Displays a list of keys used by the info command.

Move out of the bazel directory, and we'll move onto the next step.

cd ..

4. Compiling TensorFlow

First things first, clone the TensorFlow repository and move into the newly created directory.

git clone --recurse-submodules https://github.com/tensorflow/tensorflow.git
cd tensorflow

Note: if you're looking to build to a specific version or commit of TensorFlow (as opposed to the HEAD at master), you should git checkout it now.

Once in the directory, we have to write a nifty one-liner that is incredibly important. The next line goes through all files and changes references of 64-bit program implementations (which we don't have access to) to 32-bit implementations. Neat!

grep -Rl 'lib64' | xargs sed -i 's/lib64/lib/g'

Updating tensorflow/core/platform/platform.h and WORKSPACE as listed in the previous version is no longer needed with the latest version of tensorflow.

  • the IS_MOBILE_PLATFORM check now includes a specific check for the Raspberry
  • numeric/1.2.6 is no longer listed WORKSPACE

Finally, we have to replace the eigen version dependency. The version included in the current tensorflow version may result in an error (near the end of the build):

ERROR: /mnt/tensorflow/tensorflow/core/kernels/BUILD:2128:1: C++ compilation of rule '//tensorflow/core/kernels:svd_op' failed: gcc failed: error executing command 
...com.google.devtools.build.lib.shell.BadExitStatusException: Process exited with status 1.
...
external/eigen_archive/Eigen/src/Jacobi/Jacobi.h:359:55: error: 'struct Eigen::internal::conj_helper<__vector(4) __builtin_neon_sf, Eigen::internal::Packet2cf, false, false>' has no member named 'pmul'

to resolve this

sudo nano tensorflow/workspace.bzl

Replace the following

  native.new_http_archive(
      name = "eigen_archive",
      urls = [
          "http://mirror.bazel.build/bitbucket.org/eigen/eigen/get/f3a22f35b044.tar.gz",
          "https://bitbucket.org/eigen/eigen/get/f3a22f35b044.tar.gz",
      ],
      sha256 = "ca7beac153d4059c02c8fc59816c82d54ea47fe58365e8aded4082ded0b820c4",
      strip_prefix = "eigen-eigen-f3a22f35b044",
      build_file = str(Label("//third_party:eigen.BUILD")),
  )

with

  native.new_http_archive(
      name = "eigen_archive",
      urls = [
          "http://mirror.bazel.build/bitbucket.org/eigen/eigen/get/d781c1de9834.tar.gz",
          "https://bitbucket.org/eigen/eigen/get/d781c1de9834.tar.gz",
      ],
      sha256 = "a34b208da6ec18fa8da963369e166e4a368612c14d956dd2f9d7072904675d9b",
      strip_prefix = "eigen-eigen-d781c1de9834",
      build_file = str(Label("//third_party:eigen.BUILD")),
  )

Reference: https://stackoverflow.com/questions/44418657/how-to-build-eigen-with-arm-neon-compile-error-for-tensorflow

These options have changed with exception of jemalloc use No for all Now let's configure the build:

./configure

Please specify the location of python. [Default is /usr/bin/python]: /usr/bin/python
Please specify optimization flags to use during compilation when bazel option "--config=opt" is specified [Default is -march=native]: 
Do you wish to use jemalloc as the malloc implementation? [Y/n] Y
Do you wish to build TensorFlow with Google Cloud Platform support? [y/N] N
Do you wish to build TensorFlow with Hadoop File System support? [y/N] N
Do you wish to build TensorFlow with the XLA just-in-time compiler (experimental)? [y/N] N
Please input the desired Python library path to use. Default is [/usr/local/lib/python2.7/dist-packages]
Do you wish to build TensorFlow with OpenCL support? [y/N] N
Do you wish to build TensorFlow with CUDA support? [y/N] N

Note: if you want to build for Python 3, specify /usr/bin/python3 for Python's location and /usr/local/lib/python3.5/dist-packages for the Python library path.

Bazel will now attempt to clean. This takes a really long time (and often ends up erroring out anyway), so you can send some keyboard interrupts (CTRL-C) to skip this and save some time.

Now we can use it to build TensorFlow! Warning: This takes a really, really long time. Several hours.

bazel build -c opt --copt="-mfpu=neon-vfpv4" --copt="-funsafe-math-optimizations" --copt="-ftree-vectorize" --copt="-fomit-frame-pointer" --local_resources 1024,1.0,1.0 --verbose_failures tensorflow/tools/pip_package:build_pip_package

Note: I toyed around with telling Bazel to use all four cores in the Raspberry Pi, but that seemed to make compiling more prone to completely locking up. This process takes a long time regardless, so I'm sticking with the more reliable options here. If you want to be bold, try using --local_resources 1024,2.0,1.0 or --local_resources 1024,4.0,1.0

When you wake up the next morning and it's finished compiling, you're in the home stretch! Use the built binary file to create a Python wheel.

bazel-bin/tensorflow/tools/pip_package/build_pip_package /tmp/tensorflow_pkg

And then install it!

sudo pip install /tmp/tensorflow_pkg/tensorflow-1.1.0-cp27-none-linux_armv7l.whl

5. Cleaning Up

There's one last bit of house-cleaning we need to do before we're done: remove the USB drive that we've been using as swap.

First, turn off your drive as swap:

sudo swapoff /dev/XXX

Finally, remove the line you wrote in /etc/fstab referencing the device

sudo nano /etc/fstab

Then reboot your Raspberry Pi.

And you're done! You deserve a break.

References

  1. Building TensorFlow for Jetson TK1 (an update to this post is available here)
  2. Turning a USB Drive into swap
  3. Safely removing USB swap space

Back to top