Raspberry Pi 3 - TensorFlow 1.1.0

Updates to this Project

• No longer use OpenJDK (instead using the default Oracle JDK included with Raspbian)
• Updated ./configure script to show all options used.
• Added note that it's alright to skip the automatic bazel clean after running ./configure

TensorFlow Updates

See the official release notes for details on latest supported features, updates, and changes to TensorFlow itself.

Version of Rasbpian used in the binaries.:

Raspbian 8.0 "Jessie"
Release: March 2, 2017
Installed via NOOBS 2.3

Raspbian release notes

Raspberry Pi 3 - TensorFlow 1.0.1

Updates to this Project

• Users no longer have to build protoc from source in order to build Bazel/TensorFlow
• Instructions using sudo pip install... have been changed to use pip install --user... instead (#72)
  • This caused issues, so it has been reverted for now.
• More information about the version of Raspbian used has been added to the README, as well as the release notes (#73).

TensorFlow Updates

See the official release notes for details on latest supported features, updates, and changes to TensorFlow itself.

Version of Rasbpian used in the binaries.:

Raspbian 8.0 "Jessie"
Release: March 2, 2017
Installed via NOOBS 2.3

Raspbian release notes

Raspberry Pi 3 - TensorFlow 1.0.0

Updates to this Project

• When building from scratch, users have to switch the Numeric JS library protocol from https to http
• SHA-256 checksums for binaries will be included in this and future releases.

TensorFlow Updates

See the official release notes for details on latest supported features, updates, and changes to TensorFlow itself.

Raspberry Pi 3 - TensorFlow 0.12.1

Updates to this Project

• No longer have to build protobuf Java plugin
• protoc is installed to /usr/local/bin/ instead of /usr/bin/
• No longer have to compile gRPC manually
• Bazel source files are downloaded as a distribution zip file instead of cloning from GitHub
• Most of the finagling to compile Bazel has been removed
• Added a donation link to the README

TensorFlow Updates

See the official release notes for details on latest supported features, updates, and changes to TensorFlow itself.

Raspberry Pi 3 - TensorFlow 0.11.0

Updates to this Project

• Removed archive and bin directories from repository, as the large files were causing slow pushes/pulls
  • Previous versions are still available from the releases page
• Build tested with fresh install using NOOBS 2.1

TensorFlow Updates

See the official release notes for details on latest supported features, updates, and changes to TensorFlow itself.

Raspberry Pi 3 - TensorFlow 0.10.0

Updates to this Project

• Updated GUIDE.md to compile gRPC from scratch, as well as update Bazel build instructions in order to work for newer versions of Bazel.
• Final TensorFlow Bazel build now uses the flags --copt="-mfpu=neon-vfpv4", --copt="-funsafe-math-optimizations", and --copt="-ftree-vectorize"

TensorFlow Updates

See the official release notes for details on latest supported features, updates, and changes to TensorFlow itself.

Raspberry Pi 3 - TensorFlow 0.9.0

Updates to this Project

• In order to build from source, this line must be deleted from tensorflow/core/platform/platform.h. Otherwise there will be errors on compilation
• The Bazel build option --copt="-mfpu=neon" is added back in, as new compatibility was introduced into Eigen. Without this option, the compiler will throw errors complaining about not knowing how to handle certain variables.

TensorFlow Updates

See the official release notes for details on latest supported features, updates, and changes to TensorFlow itself.

Raspberry Pi 3 - TensorFlow v0.9.0 RC0

Updates to this Project

• The NEON flag to GCC does not work due to changes related to use of the Eigen::half (16-bit floating point) type inside of TensorFlow. Because of this, the --copt="-mfpu=neon" flag used when building TensorFlow has been removed from GUIDE.md. Hopefully we'll find new compiler instructions that can help optimize TensorFlow.

TensorFlow Updates

See the official pre-release notes for details on latest supported features, updates, and changes to TensorFlow itself.

Raspberry Pi 3 - TensorFlow v0.8.0

TensorFlow Updates

See the official release notes for details on latest supported features, updates, and changes to TensorFlow itself.

---

Fixes on Top of RC0

tf.train.ClusterSpec no longer throws an AttributeError when used as a parameter to tf.train.Server

You should be able to use the distributed capabilities as you would on any other system!

Raspberry Pi 3 - TensorFlow v0.8.0 RC0

Updates to this Project

• Latest binaries are now available in the bin directory
• Old binaries will now be kept in the archive directory
• Figured out how to build the latest version of Bazel natively on RPi3- the Bazel portion of GUIDE.md has been updated
• We should be able to release new binaries to coincide with the official release now that we have Bazel version>0.2 running on RPi3.
  • In the future, the goal will be to release within a week of the official release binaries.
  • Future binaries will always be built at the exact commit of the official release. You probably didn't notice, but the 0.7.1 release was built at HEAD somewhere in the middle :)

TensorFlow Updates

See the official pre-release notes for details on latest supported features, updates, and changes to TensorFlow itself.

---

Notable quirks:

tf.train.ClusterSpec throws AttributeError when used as a parameter to tf.train.Server

For whatever reason, it seems that the internal _cluster_spec parameter wants to hide itself when used as an input to the constructor of tf.train.Server. Hopefully we'll be able to figure out what the deal is, but in the meantime, it is possible to start distributed servers by just passing in the ClusterSpec dictionary directly. For example, instead of declaring a server like this:

cluster = tf.train.ClusterSpec({"local": ["localhost:2222", "localhost:2223"]})
server = tf.train.Server(cluster, job_name="local", task_index=0)

You can declare it directly like this:

server = tf.train.Server({"local": ["localhost:2222", "localhost:2223"]}, job_name="local", task_index=0)