Tensorflow Lite Develop Environment

Prepare Docker

$ cd docker
$ ./build-docker.sh
# This steps build a docker image from tensorflow:latest-devel
# and add required android sdk and ndk
# and with a normal user tflite (995) for Arch uid starts below 1000

Run Docker

$ ./scripts/run-docker.sh
# This will also git clone a tensorflow if there is not exist

Prepare WORKSPACE

android_sdk_repository(
    name = "androidsdk",
    api_level = 23,
    build_tools_version = "26.0.1",
    path = "/home/tflite/lib/android-sdk",
)

android_ndk_repository(
    name="androidndk",
    path="/home/tflite/lib/android-ndk",
    api_level=14)

bazelrc

$ cat ~/.bazelrc
startup --max_idle_secs=100000000

Build TFLite Demo App

$ bazel build --cxxopt='--std=c++11' //tensorflow/contrib/lite/java/demo/app/src/main:TfLiteCameraDemo

Build toco

$ bazel build //tensorflow/contrib/lite/toco:toco

Build freeze_graph

$ bazel build tensorflow/python/tools:freeze_graph

Build cc_test, py_test

$ bazel build //tensorflow/contrib/lite:model_test
Target //tensorflow/contrib/lite:model_test up-to-date:
  bazel-bin/tensorflow/contrib/lite/model_test
$ bazel-bin/tensorflow/contrib/lite/model_test

$ bazel build //tensorflow/contrib/quantize:input_to_ops_test
Target //tensorflow/contrib/quantize:input_to_ops_test up-to-date:
  bazel-bin/tensorflow/contrib/quantize/input_to_ops_test
$ bazel-bin/tensorflow/contrib/quantize/input_to_ops_test

Build pip Package

# CPU
$ bazel build --config=opt //tensorflow/tools/pip_package:build_pip_package
# GPU
$ bazel build --config=opt --config=cuda //tensorflow/tools/pip_package:build_pip_package

$ bazel-bin/tensorflow/tools/pip_package/build_pip_package /tmp/tensorflow_pkg
$ ls /tmp/tensorflow_pkg/
tensorflow-1.4.0-cp27-cp27mu-linux_x86_64.whl

$ sudo pip install -U /tmp/tensorflow_pkg/tensorflow-1.4.0-cp27-cp27mu-linux_x86_64.whl

Build schema_generated.h from schema.fbs

$ bazel build @flatbuffers//:flatc
$ bazel-bin/external/flatbuffers/flatc --cpp --gen-object-api tensorflow/contrib/lite/schema/schema.fbs
# add explicit to OperatorOptionBuilder
$ sed -e 's/\([a-zA-Z0-9]\+Builder(\)/explicit \1/' schema_generated.h > schema_generated_e.h
# $ clang-format schema_generated_e.h --style=google > schema_generated_f.h
$ git checkout --theirs tensorflow/contrib/lite/schema/schema_generated.h
# $ vimdiff schema_generated_f.h tensorflow/contrib/lite/schema/schema_generated.h
$ vimdiff schema_generated_e.h tensorflow/contrib/lite/schema/schema_generated.h
# need to install clang-format to make diff easier
$ sudo apt-get install clang-format

MNIST

MNIST apks

# build
$ bazel build --cxxopt='--std=c++11' //tensorflow/contrib/lite/java/demo/app/src/main:TfLiteCameraDemo
$ rm -f ../sandbox/TfLiteCameraDemo.apk
$ cp bazel-bin/tensorflow/contrib/lite/java/demo/app/src/main/TfLiteCameraDemo.apk ../sandbox/
# install
$ adb shell pm uninstall -k com.example.android.tflitecamerademo && adb install -f TfLiteCameraDemo.apk

nvidia-docker

tensorflow lastest-gpu docker image

$ nvidia-docker run -it tensorflow/tensorflow:latest-devel-gpu bash

tensorflow lastest-gpu docker image

$ DEVICE=GPU scripts/run-docker.sh
# This will also git clone a tensorflow if there is not exist

Notes for Tensorflow Lite

mobilenet

change directory to tensorflow/tensorflow/contrib/lite

in java/demo/app/src/main/java/com/example/android/tflitecamerademo/ImageClassifier.java

/** Name of the model file stored in Assets. */
  private static final String MODEL_PATH = "mobilenet_quant_v1_224.tflite";

in java/demo/app/src/main/BUILD, there is a section

android_binary(
    name = "TfLiteCameraDemo",
    srcs = glob(["java/**/*.java"]),
    assets = [
        "@tflite_mobilenet//:labels.txt",
        "@tflite_mobilenet//:mobilenet_quant_v1_224.tflite",
    ],

change directory to tensorflow

in tensorflow/workspace.bzl

native.new_http_archive(
      name = "tflite_mobilenet",
      build_file = str(Label("//third_party:tflite_mobilenet.BUILD")),
      sha256 = "23f814d1c076bdf03715dfb6cab3713aa4fbdf040fd5448c43196bd2e97a4c1b",
      urls = [
          "https://mirror.bazel.build/storage.googleapis.com/download.tensorflow.org/models/tflite/mobilenet_v1_224_android_quant_2017_11_08.zip",
          "https://storage.googleapis.com/download.tensorflow.org/models/tflite/mobilenet_v1_224_android_quant_2017_11_08.zip",
      ],

in /home/tflite/.cache/bazel/_bazel_tflite/f82e7d13eaeac899986b03b38680d292/external/tflite_mobilenet here is where @tflite_mobilenet stores

print out exception backtrace

try {
tflite.run(imgData, labelProbArray);
} catch (Exception e) {
Log.e(TAG, "YMK Exception " + e);
}

dummy-quantization with graph visualize

$ curl https://storage.googleapis.com/download.tensorflow.org/models/mobilenet_v1_0.50_128_frozen.tgz \
  | tar xzv -C /tmp
$ bazel run --config=opt \
  //tensorflow/contrib/lite/toco:toco -- \
  --input_file=/tmp/mobilenet_v1_0.50_128/frozen_graph.pb \
  --output_file=/tmp/foo.cc \
  --input_format=TENSORFLOW_GRAPHDEF \
  --output_format=TFLITE \
  --inference_type=QUANTIZED_UINT8 \
  --input_shape=1,128,128,3 \
  --input_array=input \
  --output_array=MobilenetV1/Predictions/Reshape_1 \
  --default_ranges_min=0 \
  --default_ranges_max=6 \
  --mean_value=127.5 \
  --std_value=127.5 \
  --dump_graphviz=/home/tflite/sandbox/dots

$ cd /home/tflite/sandbox/dots
$ dot -Tpdf -O ./toco_*.dot