How to add an NDK-based test from a TFLite model and example

tflite_example_converter is a small tool that converts a TFLite model and examples to a form that only depends on NNAPI. This README shows how we could use the tool to turn a TFLite test into an NNAPI test.

Build tflite_example_converter in TFLite

First, copy google3/tflite_example_converter.cc to a CitC client, say at $G3:

cp google3/tflite_example_converter.cc \
$G3/third_party/tensorflow/contrib/lite/testing

Then, edit BUILD of the same directory so blaze can build tflite_example_converter, add a target like:

cc_binary(
  name = "tflite_example_converter",
  srcs = [
      "tflite_example_converter.cc",
  ],
  deps = [
      ":parse_testdata_lib",
      "//third_party/tensorflow/contrib/lite:builtin_ops",
      "//third_party/tensorflow/contrib/lite:framework",
  ],
)

Convert a TFLite model and examples

Let's take a single-op model like “add” as an example.

cd $G3/third_party/tensorflow/contrib/lite/testing/generated_data/simple

run:

  $G3/blaze-bin/third_party/tensorflow/contrib/lite/testing/\
  tflite_example_converter add.bin add_tests.txt

This step creates two files, one in C++, representing the test inputs, and another in python, for the test generator.

The output would look like:

Model file generated at: add.bin.model.py
Example file generated at: add_tests.txt.example.cc

run test_generator.py on the generated model file, i.e.:

./test_generator.py $G3/google3/third_party/tensorflow/contrib/lite/testing/generated_data/simple/add.bin.model.py

and it would print out a C++ function like:

void CreateModel(Model *model) {
  OperandType type0(Type::TENSOR_FLOAT32, {1, 8, 8, 3});
  // Phase 1, operands
  auto op1 = model->addOperand(&type0);
  auto op2 = model->addOperand(&type0);
  auto op0 = model->addOperand(&type0);
  // Phase 2, operations
  model->addOperation(ANEURALNETWORKS_ADD, {op1, op1}, {op0});
  model->addOperation(ANEURALNETWORKS_ADD, {op0, op1}, {op2});
  // Phase 3, inputs and outputs
  model->setInputsAndOutputs(
    {op1},
    {op2});
  assert(model->isValid());
}

Save the output as a C++ source file to frameworks/ml/nn/runtime/test/generated/model, and copy the example to frameworks/ml/nn/runtime/test/generated/examples.

Then edit frameworks/ml/nn/runtime/test/Conv.cpp to include the generated files similar to how other tests are included. Add a new Google test TEST_F as needed.