TensorFlow Lite model 测试：正确的代码，类的概率

Question

I'm trying to test TensorFlow lite c++ code with TensorflowLite model. Model gets 256*256 array of floats (spectrogram or image) and do some inference on this data. The TF Lite model is designed to solve the problem of classification into 5 classes. It was derived from a conventional TF model by conversion. I use TF Lite 2.0.

Test model

This is my code:

#include <iostream>
#include <cstdio>
#include "../tensorflow/tensorflow/lite/interpreter.h"
#include "../tensorflow/tensorflow/lite/model.h"
#include "../tensorflow/tensorflow/lite/kernels/register.h"
#include "../tensorflow/tensorflow/lite/op_resolver.h"
#include <cstdlib>

int main(int argc, char** argv)
{

    const char* filename = argv[1];

    std::unique_ptr<tflite::FlatBufferModel> model = tflite::FlatBufferModel::BuildFromFile(filename);

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

    interpreter->SetNumThreads(4);
    interpreter->AllocateTensors();

    for(int i = 0; i < 256*256; i++){
    float input = rand() % 10 + rand() % 10;
            interpreter->typed_input_tensor<float>(0)[i] = input;
            //printf("%f ", input);
    }
    //printf("\n");

    interpreter->Invoke();
    int output = interpreter->outputs()[0];

    printf("%d ",  output);

    for(int i = 0; i < 5; i++)
    {
        float output  = interpreter->typed_output_tensor<float>(0)[i];
        printf("%f ", (output));
    }
    printf("\n");
} 

I have some questions:

• how to organize the input data (how to apply a two-dimensional spectrogram to the input of the model)?

• how to get the output probability of classes in right way?

• did I write the right code to test the model?

Answer 1

Your code looks about right. And since Tensorflow Lite looks at tensors in row-major format, your way of assigning inputs seems reasonable.

You probably don't need this:

int output = interpreter->outputs()[0];

printf("%d ",  output);

Otherwise, things look okay. If you pre-process the input image/spectogram the same way you did during training, you should obtain the outputs you expect.