• Architecture
• C++ API
• Optimization
• Performance

• Architecture Introduction

  See the architecture graph: https://github.com/tensorflow/tensorflow/tree/master/tensorflow/contrib/lite

  • Lite Converter Also called freeze Graph, it will merge the checkpoint values with the graph structure.
  • Android APP
    • Jave API
    • C++ API
    • Interpreter: The main executive engines
    • Android Neural Network API.

• What is the relationship between TensorFlow and TfLite?

  There is no relationship between TensorFlow and TfLite. TfLite is another lightweight inference framework.

The simple usage is as follows:

// 1. Load Model
tflite::FlatBufferModel model(path_to_model);

// 2. Init and Build Interpreter
tflite::ops::builtin::BuiltinOpResolver resolver;
std::unique_ptr<tflite::Interpreter> interpreter;
tflite::InterpreterBuilder(*model, resolver)(&interpreter);

// 3. Resize input tensors, if desired.
// Allocate Tensors and fill `input`.
interpreter->AllocateTensors();
float* input = interpreter->typed_input_tensor<float>(0);

// 4. Inference
interpreter->Invoke();

// 5. Read the output
float* output = interpreter->type_output_tensor<float>(0);

• BuiltinOpResolver
  • The regular usage will require the developer to use the BuiltinOpResolver, which has many operators.
• Operator pruning
  • Users can rewrite other OpResolver to prune the operators.