Introduction
-----------
TensorFlow is an open source software library for high performance numerical
computation primarily used in machine learning. Its flexible architecture
allows easy deployment of computation across a variety of types of platforms
(CPUs, GPUs, TPUs), and a range of systems from single desktops to clusters
of servers to mobile and edge devices.
(https://www.tensorflow.org/)

The build system of TensorFlow is Bazel (https://bazel.build/).

This layer integrates TensorFlow to OE/Yocto platform
- Integrate Google's bazel to Yocto
- Add Yocto toolchain for bazel to support cross compiling.
- Replace python package system(pip/wheel) with Yocto package system(rpm/deb/ipk).

Dependencies
------------
URI: git://github.com/openembedded/openembedded-core.git
branch: master
revision: HEAD

URI: git://github.com/openembedded/bitbake.git
branch: master
revision: HEAD

URI: git://github.com/openembedded/meta-openembedded.git
layers: meta-python, meta-oe
branch: master
revision: HEAD

URI: git://git.yoctoproject.org/meta-java
branch: master
revision: HEAD

Source code
-----------
git://git.yoctoproject.org/meta-tensorflow (TODO, github first?)

Maintenance
-----------
Maintainers: Hongxu Jia <[email protected]> | <[email protected]>

Contributing
-----------
Contributions and patches can be sent to the Yocto Project mailing
list: [email protected]"

When sending patches please take a look at the contribution guide available
here: https://wiki.yoctoproject.org/wiki/Contribution_Guidelines

example:
git send-email -1 -M --to [email protected]  --subject-prefix=meta-tensorflow][PATCH

Limitation
-----------
- Bazel build takes lots of time, since it like bitbake which has own rules and builds
  everything from scratch. Currently bazel could not reuse Yocto DEPENDS/RDEPENDS.

- Do not support offline build since bazel build system require fetches archive
  tarballs through network.

- In order to run tensorflow cases in a reasonable time, although it builds
  successfully on qemuarm, qemuarm64, qemumips, qemumips64, qemux86 and qemux86-64,
  only qemux86-64 with kvm for runtime test.

- Do not support 32-bit powerpc (qemuppc) since BoringSSL does not support it.
  (BoringSSL is a fork of OpenSSL used to implement cryptography and TLS across
  most of Google's products)

Future plan
-----------
- Support offline build which bazel build system fetches archive tarballs
  from Yocto download mirror.

- Support more BSP, such as atom, beagleboard, raspberrypi.

- Introduce more machine learning cases to meta-tensorflow.

- Recipe maintenance and upgrade

Build and run
-----------
1. Clone away
$ mkdir <ts-project>
$ cd <ts-project>
$ git clone git://git.yoctoproject.org/meta-tensorflow
$ git clone git://git.yoctoproject.org/meta-java
$ git clone git://git.openembedded.org/meta-openembedded
$ git clone git://git.openembedded.org/openembedded-core oe-core
$ cd oe-core
$ git clone git://git.openembedded.org/bitbake

2. Prepare build
$ . <ts-project>/oe-core/oe-init-build-env <build>

# Build qemux86-64 which runqemu supports kvm.
$ echo 'MACHINE = "qemux86-64"' >> conf/local.conf

$ echo 'IMAGE_INSTALL_append = " tensorflow"' >> conf/local.conf

Edit conf/bblayers.conf to include other layers
BBLAYERS ?= " \
    <ts-project>/oe-core/meta \
    <ts-project>/meta-openembedded/meta-python \
    <ts-project>/meta-openembedded/meta-oe \
    <ts-project>/meta-java \
    <ts-project>/meta-tensorflow \
"


3. Build image in <build>.
$ bitbake core-image-minimal

4. Start qemu with slrip + kvm + 5GB memory:
$ runqemu qemux86-64 core-image-minimal slirp kvm qemuparams="-m 5120"

5. Verify the install
root@qemux86-64:~# python3 -c "import tensorflow as tf; tf.enable_eager_execution(); print(tf.reduce_sum(tf.random_normal([1000, 1000])))"
tf.Tensor(-604.65454, shape=(), dtype=float32)

6. Run tutorial case
https://www.tensorflow.org/tutorials

root@qemux86-64:~# cat >code.py <<ENDOF
import tensorflow as tf
mnist = tf.keras.datasets.mnist

(x_train, y_train),(x_test, y_test) = mnist.load_data()
x_train, x_test = x_train / 255.0, x_test / 255.0

model = tf.keras.models.Sequential([
  tf.keras.layers.Flatten(input_shape=(28, 28)),
  tf.keras.layers.Dense(512, activation=tf.nn.relu),
  tf.keras.layers.Dropout(0.2),
  tf.keras.layers.Dense(10, activation=tf.nn.softmax)
])
model.compile(optimizer='adam',
              loss='sparse_categorical_crossentropy',
              metrics=['accuracy'])

model.fit(x_train, y_train, epochs=5)
model.evaluate(x_test, y_test)
ENDOF

root@qemux86-64:~# python3 ./code.py
Downloading data from https://storage.googleapis.com/tensorflow/tf-keras-datasets/mnist.npz
11493376/11490434 [==============================] - 7s 1us/step
Instructions for updating:
Colocations handled automatically by placer.
Instructions for updating:
Please use `rate` instead of `keep_prob`. Rate should be set to `rate = 1 - keep_prob`.
Epoch 1/5
60000/60000 [==============================] - 27s 449us/sample - loss: 0.2211 - acc: 0.9346
Epoch 2/5
60000/60000 [==============================] - 24s 408us/sample - loss: 0.0969 - acc: 0.9702
Epoch 3/5
60000/60000 [==============================] - 26s 439us/sample - loss: 0.0694 - acc: 0.9780
Epoch 4/5
60000/60000 [==============================] - 23s 390us/sample - loss: 0.0540 - acc: 0.9832
Epoch 5/5
60000/60000 [==============================] - 24s 399us/sample - loss: 0.0447 - acc: 0.9851
10000/10000 [==============================] - 1s 91us/sample - loss: 0.0700 - acc: 0.9782

License
-------

All metadata is MIT licensed unless otherwise stated. Source code included
in tree for individual recipes is under the LICENSE stated in each recipe
(.bb file) unless otherwise stated.