TensorFlow Lite C++ API example for inference

Question

I am trying to get a TensorFlow Lite example to run on a machine with an ARM Cortex-A72 processor. Unfortunately, I wasn't able to deploy a test model due to the lack of examples on how to use the C++ API. I will try to explain what I have achieved so far.

Create the tflite model

I have created a simple linear regression model and converted it, which should approximate the function f(x) = 2x - 1. I got this code snippet from some tutorial, but I am unable to find it anymore.

import tensorflow as tf
import numpy as np
from tensorflow import keras
from tensorflow.contrib import lite

model = keras.Sequential([keras.layers.Dense(units=1, input_shape=[1])])
model.compile(optimizer='sgd', loss='mean_squared_error')

xs = np.array([ -1.0, 0.0, 1.0, 2.0, 3.0, 4.0], dtype=float)
ys = np.array([ -3.0, -1.0, 1.0, 3.0, 5.0, 7.0], dtype=float)

model.fit(xs, ys, epochs=500)

print(model.predict([10.0]))

keras_file = 'linear.h5'
keras.models.save_model(model, keras_file)

converter = lite.TocoConverter.from_keras_model_file(keras_file)
tflite_model = converter.convert()
open('linear.tflite', 'wb').write(tflite_model)

This creates a binary called linear.tflite, which I should be able to load.

Compile TensorFlow Lite for my machine

TensorFlow Lite comes with a script for the compilation on machines with the aarch64 architecture. I followed the guide here to do this, even though I had to modify the Makefile slightly. Note that I compiled this natively on my target system. This created a static library called libtensorflow-lite.a.

Problem: Inference

I tried to follow the tutorial on the site here, and simply pasted the the code snippets from loading and running the model together, e.g.

class FlatBufferModel {
  // Build a model based on a file. Return a nullptr in case of failure.
  static std::unique_ptr<FlatBufferModel> BuildFromFile(
      const char* filename,
      ErrorReporter* error_reporter);

  // Build a model based on a pre-loaded flatbuffer. The caller retains
  // ownership of the buffer and should keep it alive until the returned object
  // is destroyed. Return a nullptr in case of failure.
  static std::unique_ptr<FlatBufferModel> BuildFromBuffer(
      const char* buffer,
      size_t buffer_size,
      ErrorReporter* error_reporter);
};

tflite::FlatBufferModel model("./linear.tflite");

tflite::ops::builtin::BuiltinOpResolver resolver;
std::unique_ptr<tflite::Interpreter> interpreter;
tflite::InterpreterBuilder(*model, resolver)(&interpreter);

// Resize input tensors, if desired.
interpreter->AllocateTensors();

float* input = interpreter->typed_input_tensor<float>(0);
// Fill `input`.

interpreter->Invoke();

float* output = interpreter->typed_output_tensor<float>(0);

When trying to compile this via

g++ demo.cpp libtensorflow-lite.a

I get a load of errors. Log:

root@localhost:/inference# g++ demo.cpp libtensorflow-lite.a 
demo.cpp:3:15: error: ‘unique_ptr’ in namespace ‘std’ does not name a template type
   static std::unique_ptr<FlatBufferModel> BuildFromFile(
               ^~~~~~~~~~
demo.cpp:10:15: error: ‘unique_ptr’ in namespace ‘std’ does not name a template type
   static std::unique_ptr<FlatBufferModel> BuildFromBuffer(
               ^~~~~~~~~~
demo.cpp:16:1: error: ‘tflite’ does not name a type
 tflite::FlatBufferModel model("./linear.tflite");
 ^~~~~~
demo.cpp:18:1: error: ‘tflite’ does not name a type
 tflite::ops::builtin::BuiltinOpResolver resolver;
 ^~~~~~
demo.cpp:19:6: error: ‘unique_ptr’ in namespace ‘std’ does not name a template type
 std::unique_ptr<tflite::Interpreter> interpreter;
      ^~~~~~~~~~
demo.cpp:20:1: error: ‘tflite’ does not name a type
 tflite::InterpreterBuilder(*model, resolver)(&interpreter);
 ^~~~~~
demo.cpp:23:1: error: ‘interpreter’ does not name a type
 interpreter->AllocateTensors();
 ^~~~~~~~~~~
demo.cpp:25:16: error: ‘interpreter’ was not declared in this scope
 float* input = interpreter->typed_input_tensor<float>(0);
                ^~~~~~~~~~~
demo.cpp:25:48: error: expected primary-expression before ‘float’
 float* input = interpreter->typed_input_tensor<float>(0);
                                                ^~~~~
demo.cpp:28:1: error: ‘interpreter’ does not name a type
 interpreter->Invoke();
 ^~~~~~~~~~~
demo.cpp:30:17: error: ‘interpreter’ was not declared in this scope
 float* output = interpreter->typed_output_tensor<float>(0);
                 ^~~~~~~~~~~
demo.cpp:30:50: error: expected primary-expression before ‘float’
 float* output = interpreter->typed_output_tensor<float>(0);

I am relatively new to C++, so I may be missing something obvious here. It seems, however, that other people have trouble with the C++ API as well (look at this GitHub issue). Has anybody also stumbled across this and got it to run?

The most important aspects for me to cover would be:

1.) Where and how do I define the signature, so that the model knows what to treat as inputs and outputs?

2.) Which headers do I have to include?

Thanks!

EDIT

Thanks to @Alex Cohn, the linker was able to find the correct headers. I also realized that I probably do not need to redefine the flatbuffers class, so I ended up with this code (minor change is marked):

#include "tensorflow/lite/interpreter.h"
#include "tensorflow/lite/kernels/register.h"
#include "tensorflow/lite/model.h"
#include "tensorflow/lite/tools/gen_op_registration.h"

auto model = tflite::FlatBufferModel::BuildFromFile("linear.tflite");   //CHANGED

tflite::ops::builtin::BuiltinOpResolver resolver;
std::unique_ptr<tflite::Interpreter> interpreter;
tflite::InterpreterBuilder(*model, resolver)(&interpreter);

// Resize input tensors, if desired.
interpreter->AllocateTensors();

float* input = interpreter->typed_input_tensor<float>(0);
// Fill `input`.

interpreter->Invoke();

float* output = interpreter->typed_output_tensor<float>(0);

This reduces the number of errors greatly, but I am not sure how to resolve the rest:

root@localhost:/inference# g++ demo.cpp -I/tensorflow
demo.cpp:10:34: error: expected ‘)’ before ‘,’ token
 tflite::InterpreterBuilder(*model, resolver)(&interpreter);
                                  ^
demo.cpp:10:44: error: expected initializer before ‘)’ token
 tflite::InterpreterBuilder(*model, resolver)(&interpreter);
                                            ^
demo.cpp:13:1: error: ‘interpreter’ does not name a type
 interpreter->AllocateTensors();
 ^~~~~~~~~~~
demo.cpp:18:1: error: ‘interpreter’ does not name a type
 interpreter->Invoke();
 ^~~~~~~~~~~

How do I have to tackle these? It seems that I have to define my own resolver, but I have no clue on how to do that.

Answer 1

I finally got it to run. Considering my directory structure looks like this:

/(root)
    /tensorflow
        # whole tf repo
    /demo
        demo.cpp
        linear.tflite
        libtensorflow-lite.a

I changed demo.cpp to

#include <stdio.h>
#include "tensorflow/lite/interpreter.h"
#include "tensorflow/lite/kernels/register.h"
#include "tensorflow/lite/model.h"
#include "tensorflow/lite/tools/gen_op_registration.h"

int main(){

    std::unique_ptr<tflite::FlatBufferModel> model = tflite::FlatBufferModel::BuildFromFile("linear.tflite");

    if(!model){
        printf("Failed to mmap model\n");
        exit(0);
    }

    tflite::ops::builtin::BuiltinOpResolver resolver;
    std::unique_ptr<tflite::Interpreter> interpreter;
    tflite::InterpreterBuilder(*model.get(), resolver)(&interpreter);

    // Resize input tensors, if desired.
    interpreter->AllocateTensors();

    float* input = interpreter->typed_input_tensor<float>(0);
    // Dummy input for testing
    *input = 2.0;

    interpreter->Invoke();

    float* output = interpreter->typed_output_tensor<float>(0);

    printf("Result is: %f\n", *output);

    return 0;
}

Also, I had to adapt my compile command (I had to install flatbuffers manually to make it work). What worked for me was:

g++ demo.cpp -I/tensorflow -L/demo -ltensorflow-lite -lrt -ldl -pthread -lflatbuffers -o demo

Thanks to @AlexCohn for getting me on the right track!

Answer 2

Here is the minimal set of includes:

#include "tensorflow/lite/interpreter.h"
#include "tensorflow/lite/kernels/register.h"
#include "tensorflow/lite/model.h"
#include "tensorflow/lite/tools/gen_op_registration.h"

These will include other headers, e.g. <memory> which defines std::unique_ptr.