C# 编写 TensorFlow 人工智能应用

TensorFlowSharp入门使用C#编写TensorFlow人工智能应用学习。

 

TensorFlow简单介绍

 

TensorFlow 是谷歌的第二代机器学习系统，按照谷歌所说，在某些基准测试中，TensorFlow的表现比第一代的DistBelief快了2倍。

 

TensorFlow 内建深度学习的扩展支持，任何可以用计算流图形来表达的计算，均可以使用TensorFlow。

 

任何基于梯度的机器学习算法都可以受益于TensorFlow的自动分化（auto-differentiation）。经过灵活的Python接口，要在TensorFlow中表达想法也会很容易。

 

TensorFlow 对于实际的产品也是颇有意义的。将思路从桌面GPU训练无缝搬迁到手机中运行。

 

示例Python代码：

 

import tensorflow as tf

import numpy as np

 

# Create 100 phony x, y data points in NumPy, y = x * 0.1 + 0.3

x_data = np.random.rand(100).astype(np.float32)

y_data = x_data * 0.1 + 0.3

 

# Try to find values for W and b that compute y_data = W * x_data + b

# (We know that W should be 0.1 and b 0.3, but TensorFlow will

# figure that out for us.)

W = tf.Variable(tf.random_uniform([1], -1.0, 1.0))

b = tf.Variable(tf.zeros([1]))

y = W * x_data + b

 

# Minimize the mean squared errors.

loss = tf.reduce_mean(tf.square(y - y_data))

optimizer = tf.train.GradientDescentOptimizer(0.5)

train = optimizer.minimize(loss)

 

# Before starting, initialize the variables.  We will 'run' this first.

init = tf.global_variables_initializer()

 

# Launch the graph.

sess = tf.Session()

sess.run(init)

 

# Fit the line.

for step in range(201):

    sess.run(train)

    if step % 20 == 0:

        print(step, sess.run(W), sess.run(b))

 

# Learns best fit is W: [0.1], b: [0.3]

 

使用TensorFlowSharp 

 

GitHub:https://github.com/migueldeicaza/TensorFlowSharp

 

官方源码库，该项目支持跨平台，使用Mono。

 

能够使用NuGet 安装TensorFlowSharp，以下：

 

Install-Package TensorFlowSharp

 

编写简单应用

 

使用VS2017新建一个.NET Framework 控制台应用 tensorflowdemo，接着添加TensorFlowSharp 引用。

 

TensorFlowSharp 包比较大，须要耐心等待。

 

而后在项目属性中生成->平台目标 改成 x64。

 

打开Program.cs 写入以下代码：

 

static void Main(string[] args)

{

    using (var session = new TFSession())

    {

        var graph = session.Graph;

        Console.WriteLine(TFCore.Version);

        var a = graph.Const(2);

        var b = graph.Const(3);

        Console.WriteLine("a=2 b=3");

 

        // 两常量加

        var addingResults = session.GetRunner().Run(graph.Add(a, b));

        var addingResultValue = addingResults[0].GetValue();

        Console.WriteLine("a+b={0}", addingResultValue);

 

        // 两常量乘

        var multiplyResults = session.GetRunner().Run(graph.Mul(a, b));

        var multiplyResultValue = multiplyResults[0].GetValue();

        Console.WriteLine("a*b={0}", multiplyResultValue);

        var tft = new TFTensor(Encoding.UTF8.GetBytes($"Hello TensorFlow Version {TFCore.Version}! LineZero"));

        var hello = graph.Const(tft);

        var helloResults = session.GetRunner().Run(hello);

        Console.WriteLine(Encoding.UTF8.GetString((byte[])helloResults[0].GetValue()));

    }

    Console.ReadKey();

}        

 

运行程序结果以下：

 

 

TensorFlow C# image recognition

 

图像识别示例体验

 

https://github.com/migueldeicaza/TensorFlowSharp/tree/master/Examples/ExampleInceptionInference

 

下面学习一个实际的人工智能应用，是很是简单的一个示例，图像识别。

 

新建一个 imagerecognition .NET Framework 控制台应用项目，接着添加TensorFlowSharp 引用。

 

而后在项目属性中生成->平台目标 改成 x64。

 

接着编写以下代码：

 

class Program

{

    static string dir, modelFile, labelsFile;

    public static void Main(string[] args)

    {

        dir = "tmp";

        List<string> files = Directory.GetFiles("img").ToList();

        ModelFiles(dir);

        var graph = new TFGraph();

        // 从文件加载序列化的GraphDef

        var model = File.ReadAllBytes(modelFile);

        //导入GraphDef

        graph.Import(model, "");

        using (var session = new TFSession(graph))

        {

            var labels = File.ReadAllLines(labelsFile);

            Console.WriteLine("TensorFlow图像识别 LineZero");

            foreach (var file in files)

            {

                // Run inference on the image files

                // For multiple images, session.Run() can be called in a loop (and

                // concurrently). Alternatively, images can be batched since the model

                // accepts batches of image data as input.

                var tensor = CreateTensorFromImageFile(file);

 

                var runner = session.GetRunner();

                runner.AddInput(graph["input"][0], tensor).Fetch(graph["output"][0]);

                var output = runner.Run();

                // output[0].Value() is a vector containing probabilities of

                // labels for each image in the "batch". The batch size was 1.

                // Find the most probably label index.

 

                var result = output[0];

                var rshape = result.Shape;

                if (result.NumDims != 2 || rshape[0] != 1)

                {

                    var shape = "";

                    foreach (var d in rshape)

                    {

                        shape += $"{d} ";

                    }

                    shape = shape.Trim();

                    Console.WriteLine($"Error: expected to produce a [1 N] shaped tensor where N is the number of labels, instead it produced one with shape [{shape}]");

                    Environment.Exit(1);

                }

 

                // You can get the data in two ways, as a multi-dimensional array, or arrays of arrays, 

                // code can be nicer to read with one or the other, pick it based on how you want to process

                // it

                bool jagged = true;

 

                var bestIdx = 0;

                float p = 0, best = 0;

 

                if (jagged)

                {

                    var probabilities = ((float[][])result.GetValue(jagged: true))[0];

                    for (int i = 0; i < probabilities.Length; i++)

                    {

                        if (probabilities[i] > best)

                        {

                            bestIdx = i;

                            best = probabilities[i];

                        }

                    }

 

                }

                else

                {

                    var val = (float[,])result.GetValue(jagged: false);

 

                    // Result is [1,N], flatten array

                    for (int i = 0; i < val.GetLength(1); i++)

                    {

                        if (val[0, i] > best)

                        {

                            bestIdx = i;

                            best = val[0, i];

                        }

                    }

                }

 

                Console.WriteLine($"{Path.GetFileName(file)} 最佳匹配: [{bestIdx}] {best * 100.0}% 标识为：{labels[bestIdx]}");

            }

        }

        Console.ReadKey();

    }

 

    // Convert the image in filename to a Tensor suitable as input to the Inception model.

    static TFTensor CreateTensorFromImageFile(string file)

    {

        var contents = File.ReadAllBytes(file);

 

        // DecodeJpeg uses a scalar String-valued tensor as input.

        var tensor = TFTensor.CreateString(contents);

 

        TFGraph graph;

        TFOutput input, output;

 

        // Construct a graph to normalize the image

        ConstructGraphToNormalizeImage(out graph, out input, out output);

 

        // Execute that graph to normalize this one image

        using (var session = new TFSession(graph))

        {

            var normalized = session.Run(

                     inputs: new[] { input },

                     inputValues: new[] { tensor },

                     outputs: new[] { output });

 

            return normalized[0];

        }

    }

 

    // The inception model takes as input the image described by a Tensor in a very

    // specific normalized format (a particular image size, shape of the input tensor,

    // normalized pixel values etc.).

    //

    // This function constructs a graph of TensorFlow operations which takes as

    // input a JPEG-encoded string and returns a tensor suitable as input to the

    // inception model.

    static void ConstructGraphToNormalizeImage(out TFGraph graph, out TFOutput input, out TFOutput output)

    {

        // Some constants specific to the pre-trained model at:

        // https://storage.googleapis.com/download.tensorflow.org/models/inception5h.zip

        //

        // - The model was trained after with images scaled to 224x224 pixels.

        // - The colors, represented as R, G, B in 1-byte each were converted to

        //   float using (value - Mean)/Scale.

 

        const int W = 224;

        const int H = 224;

        const float Mean = 117;

        const float Scale = 1;

 

        graph = new TFGraph();

        input = graph.Placeholder(TFDataType.String);

 

        output = graph.Div(

            x: graph.Sub(

                x: graph.ResizeBilinear(

                    images: graph.ExpandDims(

                        input: graph.Cast(

                            graph.DecodeJpeg(contents: input, channels: 3), DstT: TFDataType.Float),

                        dim: graph.Const(0, "make_batch")),

                    size: graph.Const(new int[] { W, H }, "size")),

                y: graph.Const(Mean, "mean")),

            y: graph.Const(Scale, "scale"));

    }

 

    /// <summary>

    /// 下载初始Graph和标签

    /// </summary>

    /// <param name="dir"></param>

    static void ModelFiles(string dir)

    {

        string url = "https://storage.googleapis.com/download.tensorflow.org/models/inception5h.zip";

 

        modelFile = Path.Combine(dir, "tensorflow_inception_graph.pb");

        labelsFile = Path.Combine(dir, "imagenet_comp_graph_label_strings.txt");

        var zipfile = Path.Combine(dir, "inception5h.zip");

 

        if (File.Exists(modelFile) && File.Exists(labelsFile))

            return;

 

        Directory.CreateDirectory(dir);

        var wc = new WebClient();

        wc.DownloadFile(url, zipfile);

        ZipFile.ExtractToDirectory(zipfile, dir);

        File.Delete(zipfile);

    }

}

 

这里须要注意的是因为须要下载初始Graph和标签，并且是google的站点，因此得使用一些特殊手段。

 

最终我随便下载了几张图放到bin\Debug\img

 

 

而后运行程序，首先确保bin\Debug\tmp文件夹下有tensorflow_inception_graph.pb及imagenet_comp_graph_label_strings.txt。

 

 

人工智能的魅力很是大，本文只是一个入门，复制上面的代码，你无法训练模型等等操做。因此道路仍是很远，需一步一步来。

 

更多能够查看 https://github.com/migueldeicaza/TensorFlowSharp 及 https://github.com/tensorflow/models

 

参考文档：

 

TensorFlow 官网：https://www.tensorflow.org/get_started/

 

TensorFlow 中文社区：http://www.tensorfly.cn/

 

TensorFlow 官方文档中文版：http://wiki.jikexueyuan.com/project/tensorflow-zh/