上一篇文章中讲了tflite模型的建立和Python端的导入，现在开始在ARM板上的导入。
为了不再重新生成输入数据，我将PC上python生成的数据保存到txt文件。所以稍微改了下上次那个代码：

import numpy as np
import time
import math
import tensorflow as tf

SIZE = 1000
X = np.random.rand(SIZE, 1)
X = X*(math.pi/2.0)
np.savetxt("/home/alcht0/share/project/tensorflow-v1.12.0/tmp.txt", X);

start = time.time()
x1 = tf.placeholder(tf.float32, [SIZE, 1], name='x1-input')
x2 = tf.placeholder(tf.float32, [SIZE, 1], name='x2-input')
y1 = tf.sin(x1)
y2 = tf.sin(x2)
y = y1*y2

with tf.Session() as sess:
    init_op = tf.global_variables_initializer()
    sess.run(init_op)
    converter = tf.lite.TFLiteConverter.from_session(sess, [x1, x2], [y])
    tflite_model = converter.convert()
    open("/home/alcht0/share/project/tensorflow-v1.12.0/converted_model.tflite", "wb").write(tflite_model)

end = time.time()
print("2nd ", str(end - start))

然后就是要用C++重写导入的代码。（网络上ARM版的Tensorflow Lite的资料真的不多。。搞底层的就是比较苦逼😭）
其实Tensorflow lite自己有例子，例子lable_image是读图的，和我做的不太合适，例子mnist里面指定输入的地方留白让我们自己加。。导致我在指定输入这里搞了好一会儿。。还是先上代码：

#include "tensorflow/contrib/lite/model.h"
#include "tensorflow/contrib/lite/mutable_op_resolver.h"
#include "tensorflow/contrib/lite/kernels/register.h"
#include <cstdlib>
#include <sys/time.h>   // NOLINT(build/include_order)
#include <iostream>
#include <cmath>

#define LOG(x) std::cerr

using namespace tflite;

float* arr;

double get_us(struct timeval t) { return (t.tv_sec * 1000000 + t.tv_usec); }

void loadtxt(const char* txt, float* arr)
{
    int r;
    int n;
    FILE *fpRead=fopen(txt,"r");
    if(fpRead==NULL)
    {
        printf("%s File Open Failed\n", txt);
        exit(-1);
    }
    n = 0;
    while(1)
    {
        r = fscanf(fpRead,"%f\n",&arr[n]);
        if(r!=1)
        {
            break;
        }
        n++;
    }
    printf("Read %d data from input file\n", n);
    fclose(fpRead);
}

void generateinputfromfile(int count)
{
    arr = (float*)malloc(count*sizeof(float));
    loadtxt("tmp.txt", arr);
}

int main(int argc, char** argv) {
    struct timeval start_time, stop_time;
    const char* filename = argv[1];
    int num_threads = 1;
    std::string input_layer_type = "float";
    int base_index = 0;
    int count = atoi(argv[2]);

    if(argc == 4)
    {
        num_threads = atoi(argv[3]);
    }
    printf("model size is %d\n", count);
    generateinputfromfile(count);

    gettimeofday(&start_time, nullptr);

    printf("Loading Model File ....\n");
    std::unique_ptr<tflite::FlatBufferModel> model;
    model = tflite::FlatBufferModel::BuildFromFile(filename);
    if (!model) 
    {
        LOG(FATAL) << "\nFailed to mmap model " << filename << "\n";
        exit(-1);
    }
    printf("Model Loading Complete\n");

    std::unique_ptr<tflite::Interpreter> interpreter;
    tflite::ops::builtin::BuiltinOpResolver resolver;
    tflite::InterpreterBuilder(*model, resolver)(&interpreter);
    if (!interpreter) 
    {
        LOG(FATAL) << "Failed to construct interpreter\n";
        exit(-1);
    }
    printf("Interpreter Construct Complete\n");

    if(num_threads != 1)
    {
        interpreter->SetNumThreads(num_threads);
    }

    if(interpreter->AllocateTensors() != kTfLiteOk)
    {
        printf("Failed to allocate tensors\n");
        exit(0);
    }

    for(unsigned i=0;i<count;i++)
    {
        interpreter->typed_input_tensor<float>(0)[i] = arr[i];
        interpreter->typed_input_tensor<float>(1)[i] = arr[i];
    }

    if(interpreter->Invoke() != kTfLiteOk)
    {
        std::printf("Failed to invoke!\n");
        exit(0);
    }

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

    gettimeofday(&stop_time, nullptr);
    printf("Tensorflow Complete time: %f ms\n", (get_us(stop_time) - get_us(start_time))/1000);

    free(arr);
    return 0;
}

首先是读入tflite模型文件。

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

然后，由于在ARM上是直接根据tflite文件来构造计算图得出结果，所以不需要建session，直接建interpreter。

    std::unique_ptr<tflite::Interpreter> interpreter;

这个据说是建立算子

    tflite::ops::builtin::BuiltinOpResolver resolver;

根据算子来build interpreter

    tflite::InterpreterBuilder(*model, resolver)(&interpreter);

然后，分配tensor内存

 if(interpreter->AllocateTensors() != kTfLiteOk)

接下来就是搞了好久的定义输入，其实很简单。。

    for(unsigned i=0;i<count;i++)
    {
        interpreter->typed_input_tensor<float>(0)[i] = arr[i];
        interpreter->typed_input_tensor<float>(1)[i] = arr[i];
    }

主要是之前没搞清楚这个怎么用。还需要注意的是不能直接把arr的地址指给interpreter->typed_input_tensor<float>(0)，会报错，报的好像是类型错误。不知道如果memcpy的话行不行，后面再试试。
后面就是运行和取输出了，没啥。

ARM上面也运行起来了。不过速度真的是不敢恭维。。。在PC的Python上做1000次运算的时候Tensorflow跑的比直接算慢，但是做100000次的时候就已经快很多了。然而在ARM上跑，Tensorflow Lite比直接算慢了太多太多。而且和输入定义时候的for循环无关，直接是invoke()就消耗了很多时间。不清楚是不是Tensorflow Lite的底层对指令优化的不好，又或者他的图计算优势实在有限，毕竟三角函数的运算C库可是用了很久了，优化什么的肯定是做到最好了。

root@imx6dl-XXX:~/TensorFlowLite/100000# ./test converted_model.tflite 100000
model size is 100000
Read 100000 data from input file
Loading Model File ....
Model Loading Complete
Tensorflow inport modle time: 1.047000 ms
Interpreter Construct Complete
Tensorflow build interpreter time: 3.944000 ms
Tensorflow alloc tensor time: 4.248000 ms
Tensorflow set input time: 9.393000 ms
Tensorflow Complete time: 40.937000 ms
C++ std Complete time: 0.001000 ms

后面再看看算子部分，看看能不能把cos()加进去。