今天，我结合代码来详细介绍如何使用 SciSharp STACK 的 TensorFlow.NET 来训练CNN模型，该模型主要实现 图像的分类 ，可以直接移植该代码在 CPU 或 GPU 下使用，并针对你们自己本地的图像数据集进行训练和推理。TensorFlow.NET是基于 .NET Standard 框架的完整实现的TensorFlow，可以支持 .NET Framework 或 .NET CORE , TensorFlow.NET 为广大.NET开发者提供了完美的机器学习框架选择。
SciSharp STACK：https://github.com/SciSharp
TensorFlow.NET 是 SciSharp STACK 开源社区团队的贡献，其使命是打造一个完全属于.NET开发者自己的机器学习平台，特别对于C#开发人员来说，是一个“0”学习成本的机器学习平台，该平台集成了大量API和底层封装，力图使TensorFlow的Python代码风格和编程习惯可以无缝移植到.NET平台，下图是同样TF任务的Python实现和C#实现的语法相似度对比，从中读者基本可以略窥一二。
由于TensorFlow.NET在.NET平台的优秀性能，同时搭配SciSharp的NumSharp、SharpCV、Pandas.NET、Keras.NET、Matplotlib.Net等模块，可以完全脱离Python环境使用，目前已经被微软ML.NET官方的底层算法集成，并被谷歌写入TensorFlow官网教程推荐给全球开发者。
SciSharp 产品结构
微软 ML.NET底层集成算法
谷歌官方推荐.NET开发者使用
URL: https://www.tensorflow.org/versions/r2.0/api_docs
本文利用TensorFlow.NET构建简单的图像分类模型，针对工业现场的印刷字符进行单字符OCR识别，从工业相机获取原始大尺寸的图像，前期使用OpenCV进行图像预处理和字符分割，提取出单个字符的小图，送入TF进行推理，推理的结果按照顺序组合成完整的字符串，返回至主程序逻辑进行后续的生产线工序。
实际使用中，如果你们需要训练自己的图像，只需要把训练的文件夹按照规定的顺序替换成你们自己的图片即可。支持GPU或CPU方式，该项目的完整代码在GitHub如下：
本项目的CNN模型主要由 2个卷积层&池化层 和 1个全连接层 组成，激活函数使用常见的Relu，是一个比较浅的卷积神经网络模型。其中超参数之一"学习率"，采用了自定义的动态下降的学习率，后面会有详细说明。具体每一层的Shape参考下图：
为了模型测试的训练速度考虑，图像数据集主要节选了一小部分的OCR字符（X、Y、Z），数据集的特征如下：
分类数量：3 classes 【X/Y/Z】
图像尺寸：Width 64 × Height 64
图像通道：1 channel（灰度图）
数据集数量：
train：X - 384pcs ； Y - 384pcs ； Z - 384pcs
validation：X - 96pcs ； Y - 96pcs ； Z - 96pcs
test：X - 96pcs ； Y - 96pcs ； Z - 96pcs
其它说明：数据集已经经过 随机 翻转/平移/缩放/镜像 等预处理进行增强
整体数据集情况如下图所示：
.NET 框架：使用.NET Framework 4.7.2及以上，或者使用.NET CORE 2.2及以上
CPU 配置： Any CPU 或 X64 皆可
GPU 配置：需要自行配置好CUDA和环境变量，建议 CUDA v10.1，Cudnn v7.5
从NuGet安装必要的依赖项，主要是SciSharp相关的类库，如下图所示：
注意事项：尽量安装最新版本的类库，CV须使用 SciSharp 的 SharpCV 方便内部变量传递
<PackageReference Include="Colorful.Console" Version="1.2.9" /><PackageReference Include="Newtonsoft.Json" Version="12.0.3" /><PackageReference Include="SciSharp.TensorFlow.Redist" Version="1.15.0" /><PackageReference Include="SciSharp.TensorFlowHub" Version="0.0.5" /><PackageReference Include="SharpCV" Version="0.2.0" /><PackageReference Include="SharpZipLib" Version="1.2.0" /><PackageReference Include="System.Drawing.Common" Version="4.7.0" /><PackageReference Include="TensorFlow.NET" Version="0.14.0" />
引用命名空间，包括 NumSharp、Tensorflow 和 SharpCV ;
using NumSharp;using NumSharp.Backends;using NumSharp.Backends.Unmanaged;using SharpCV;using System;using System.Collections;using System.Collections.Generic;using System.Diagnostics;using System.IO;using System.Linq;using System.Runtime.CompilerServices;using Tensorflow;using static Tensorflow.Binding;using static SharpCV.Binding;using System.Collections.Concurrent;using System.Threading.Tasks;
###
主逻辑：
准备数据
创建计算图
训练
预测
public bool Run(){ PrepareData(); BuildGraph(); using (var sess = tf.Session()) { Train(sess); Test(sess); } TestDataOutput(); return accuracy_test > 0.98;}
数据集地址：https://github.com/SciSharp/SciSharp-Stack-Examples/blob/master/data/data_CnnInYourOwnData.zip
数据集下载和解压代码 ( 部分封装的方法请参考 GitHub完整代码 )：
string url = "https://github.com/SciSharp/SciSharp-Stack-Examples/blob/master/data/data_CnnInYourOwnData.zip";Directory.CreateDirectory(Name);Utility.Web.Download(url, Name, "data_CnnInYourOwnData.zip");Utility.Compress.UnZip(Name + "\\data_CnnInYourOwnData.zip", Name);
读取目录下的子文件夹名称，作为分类的字典，方便后面One-hot使用
private void FillDictionaryLabel(string DirPath) { string[] str_dir = Directory.GetDirectories(DirPath, "*", SearchOption.TopDirectoryOnly); int str_dir_num = str_dir.Length; if (str_dir_num > 0) { Dict_Label = new Dictionary<Int64, string>(); for (int i = 0; i < str_dir_num; i++) { string label = (str_dir[i].Replace(DirPath + "\\", "")).Split('\\').First(); Dict_Label.Add(i, label); print(i.ToString() + " : " + label); } n_classes = Dict_Label.Count; } }
从文件夹中读取train、validation、test的list，并随机打乱顺序。
读取目录
ArrayFileName_Train = Directory.GetFiles(Name + "\\train", "*.*", SearchOption.AllDirectories);ArrayLabel_Train = GetLabelArray(ArrayFileName_Train);ArrayFileName_Validation = Directory.GetFiles(Name + "\\validation", "*.*", SearchOption.AllDirectories);ArrayLabel_Validation = GetLabelArray(ArrayFileName_Validation);ArrayFileName_Test = Directory.GetFiles(Name + "\\test", "*.*", SearchOption.AllDirectories);ArrayLabel_Test = GetLabelArray(ArrayFileName_Test);
获得标签
private Int64[] GetLabelArray(string[] FilesArray){ Int64[] ArrayLabel = new Int64[FilesArray.Length]; for (int i = 0; i < ArrayLabel.Length; i++) { string[] labels = FilesArray[i].Split('\\'); string label = labels[labels.Length - 2]; ArrayLabel[i] = Dict_Label.Single(k => k.Value == label).Key; } return ArrayLabel;}
随机乱序
public (string[], Int64[]) ShuffleArray(int count, string[] images, Int64[] labels){ ArrayList mylist = new ArrayList(); string[] new_images = new string[count]; Int64[] new_labels = new Int64[count]; Random r = new Random(); for (int i = 0; i < count; i++) { mylist.Add(i); } for (int i = 0; i < count; i++) { int rand = r.Next(mylist.Count); new_images[i] = images[(int)(mylist[rand])]; new_labels[i] = labels[(int)(mylist[rand])]; mylist.RemoveAt(rand); } print("shuffle array list： " + count.ToString()); return (new_images, new_labels);}
Validation/Test数据集和标签一次性预先载入成NDArray格式。
private void LoadImagesToNDArray(){ //Load labels y_valid = np.eye(Dict_Label.Count)[new NDArray(ArrayLabel_Validation)]; y_test = np.eye(Dict_Label.Count)[new NDArray(ArrayLabel_Test)]; print("Load Labels To NDArray : OK!"); //Load Images x_valid = np.zeros(ArrayFileName_Validation.Length, img_h, img_w, n_channels); x_test = np.zeros(ArrayFileName_Test.Length, img_h, img_w, n_channels); LoadImage(ArrayFileName_Validation, x_valid, "validation"); LoadImage(ArrayFileName_Test, x_test, "test"); print("Load Images To NDArray : OK!");}private void LoadImage(string[] a, NDArray b, string c){ for (int i = 0; i < a.Length; i++) { b[i] = ReadTensorFromImageFile(a[i]); Console.Write("."); } Console.WriteLine(); Console.WriteLine("Load Images To NDArray: " + c);}private NDArray ReadTensorFromImageFile(string file_name){ using (var graph = tf.Graph().as_default()) { var file_reader = tf.read_file(file_name, "file_reader"); var decodeJpeg = tf.image.decode_jpeg(file_reader, channels: n_channels, name: "DecodeJpeg"); var cast = tf.cast(decodeJpeg, tf.float32); var dims_expander = tf.expand_dims(cast, 0); var resize = tf.constant(new int[] { img_h, img_w }); var bilinear = tf.image.resize_bilinear(dims_expander, resize); var sub = tf.subtract(bilinear, new float[] { img_mean }); var normalized = tf.divide(sub, new float[] { img_std }); using (var sess = tf.Session(graph)) { return sess.run(normalized); } }}
构建CNN静态计算图，其中学习率每n轮Epoch进行1次递减。
#region BuildGraphpublic Graph BuildGraph(){ var graph = new Graph().as_default(); tf_with(tf.name_scope("Input"), delegate { x = tf.placeholder(tf.float32, shape: (-1, img_h, img_w, n_channels), name: "X"); y = tf.placeholder(tf.float32, shape: (-1, n_classes), name: "Y"); }); var conv1 = conv_layer(x, filter_size1, num_filters1, stride1, name: "conv1"); var pool1 = max_pool(conv1, ksize: 2, stride: 2, name: "pool1"); var conv2 = conv_layer(pool1, filter_size2, num_filters2, stride2, name: "conv2"); var pool2 = max_pool(conv2, ksize: 2, stride: 2, name: "pool2"); var layer_flat = flatten_layer(pool2); var fc1 = fc_layer(layer_flat, h1, "FC1", use_relu: true); var output_logits = fc_layer(fc1, n_classes, "OUT", use_relu: false); //Some important parameter saved with graph , easy to load later var img_h_t = tf.constant(img_h, name: "img_h"); var img_w_t = tf.constant(img_w, name: "img_w"); var img_mean_t = tf.constant(img_mean, name: "img_mean"); var img_std_t = tf.constant(img_std, name: "img_std"); var channels_t = tf.constant(n_channels, name: "img_channels"); //learning rate decay gloabl_steps = tf.Variable(0, trainable: false); learning_rate = tf.Variable(learning_rate_base); //create train images graph tf_with(tf.variable_scope("LoadImage"), delegate { decodeJpeg = tf.placeholder(tf.@byte, name: "DecodeJpeg"); var cast = tf.cast(decodeJpeg, tf.float32); var dims_expander = tf.expand_dims(cast, 0); var resize = tf.constant(new int[] { img_h, img_w }); var bilinear = tf.image.resize_bilinear(dims_expander, resize); var sub = tf.subtract(bilinear, new float[] { img_mean }); normalized = tf.divide(sub, new float[] { img_std }, name: "normalized"); }); tf_with(tf.variable_scope("Train"), delegate { tf_with(tf.variable_scope("Loss"), delegate { loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels: y, logits: output_logits), name: "loss"); }); tf_with(tf.variable_scope("Optimizer"), delegate { optimizer = tf.train.AdamOptimizer(learning_rate: learning_rate, name: "Adam-op").minimize(loss, global_step: gloabl_steps); }); tf_with(tf.variable_scope("Accuracy"), delegate { var correct_prediction = tf.equal(tf.argmax(output_logits, 1), tf.argmax(y, 1), name: "correct_pred"); accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32), name: "accuracy"); }); tf_with(tf.variable_scope("Prediction"), delegate { cls_prediction = tf.argmax(output_logits, axis: 1, name: "predictions"); prob = tf.nn.softmax(output_logits, axis: 1, name: "prob"); }); }); return graph;}/// <summary>/// Create a 2D convolution layer/// </summary>/// <param name="x">input from previous layer</param>/// <param name="filter_size">size of each filter</param>/// <param name="num_filters">number of filters(or output feature maps)</param>/// <param name="stride">filter stride</param>/// <param name="name">layer name</param>/// <returns>The output array</returns>private Tensor conv_layer(Tensor x, int filter_size, int num_filters, int stride, string name){ return tf_with(tf.variable_scope(name), delegate { var num_in_channel = x.shape[x.NDims - 1]; var shape = new[] { filter_size, filter_size, num_in_channel, num_filters }; var W = weight_variable("W", shape); // var tf.summary.histogram("weight", W); var b = bias_variable("b", new[] { num_filters }); // tf.summary.histogram("bias", b); var layer = tf.nn.conv2d(x, W, strides: new[] { 1, stride, stride, 1 }, padding: "SAME"); layer += b; return tf.nn.relu(layer); });}/// <summary>/// Create a max pooling layer/// </summary>/// <param name="x">input to max-pooling layer</param>/// <param name="ksize">size of the max-pooling filter</param>/// <param name="stride">stride of the max-pooling filter</param>/// <param name="name">layer name</param>/// <returns>The output array</returns>private Tensor max_pool(Tensor x, int ksize, int stride, string name){ return tf.nn.max_pool(x, ksize: new[] { 1, ksize, ksize, 1 }, strides: new[] { 1, stride, stride, 1 }, padding: "SAME", name: name);}/// <summary>/// Flattens the output of the convolutional layer to be fed into fully-connected layer/// </summary>/// <param name="layer">input array</param>/// <returns>flattened array</returns>private Tensor flatten_layer(Tensor layer){ return tf_with(tf.variable_scope("Flatten_layer"), delegate { var layer_shape = layer.TensorShape; var num_features = layer_shape[new Slice(1, 4)].size; var layer_flat = tf.reshape(layer, new[] { -1, num_features }); return layer_flat; });}/// <summary>/// Create a weight variable with appropriate initialization/// </summary>/// <param name="name"></param>/// <param name="shape"></param>/// <returns></returns>private RefVariable weight_variable(string name, int[] shape){ var initer = tf.truncated_normal_initializer(stddev: 0.01f); return tf.get_variable(name, dtype: tf.float32, shape: shape, initializer: initer);}/// <summary>/// Create a bias variable with appropriate initialization/// </summary>/// <param name="name"></param>/// <param name="shape"></param>/// <returns></returns>private RefVariable bias_variable(string name, int[] shape){ var initial = tf.constant(0f, shape: shape, dtype: tf.float32); return tf.get_variable(name, dtype: tf.float32, initializer: initial);}/// <summary>/// Create a fully-connected layer/// </summary>/// <param name="x">input from previous layer</param>/// <param name="num_units">number of hidden units in the fully-connected layer</param>/// <param name="name">layer name</param>/// <param name="use_relu">boolean to add ReLU non-linearity (or not)</param>/// <returns>The output array</returns>private Tensor fc_layer(Tensor x, int num_units, string name, bool use_relu = true){ return tf_with(tf.variable_scope(name), delegate { var in_dim = x.shape[1]; var W = weight_variable("W_" + name, shape: new[] { in_dim, num_units }); var b = bias_variable("b_" + name, new[] { num_units }); var layer = tf.matmul(x, W) + b; if (use_relu) layer = tf.nn.relu(layer); return layer; });}#endregion
Batch数据集的读取，采用了 SharpCV 的cv2.imread，可以直接读取本地图像文件至NDArray，实现CV和Numpy的无缝对接；
使用.NET的异步线程安全队列BlockingCollection<T>，实现TensorFlow原生的队列管理器FIFOQueue；
在训练模型的时候，我们需要将样本从硬盘读取到内存之后，才能进行训练。我们在会话中运行多个线程，并加入队列管理器进行线程间的文件入队出队操作，并限制队列容量，主线程可以利用队列中的数据进行训练，另一个线程进行本地文件的IO读取，这样可以实现数据的读取和模型的训练是异步的，降低训练时间。
模型的保存，可以选择每轮训练都保存，或最佳训练模型保存
#region Trainpublic void Train(Session sess){ // Number of training iterations in each epoch var num_tr_iter = (ArrayLabel_Train.Length) / batch_size; var init = tf.global_variables_initializer(); sess.run(init); var saver = tf.train.Saver(tf.global_variables(), max_to_keep: 10); path_model = Name + "\\MODEL"; Directory.CreateDirectory(path_model); float loss_val = 100.0f; float accuracy_val = 0f; var sw = new Stopwatch(); sw.Start(); foreach (var epoch in range(epochs)) { print($"Training epoch: {epoch + 1}"); // Randomly shuffle the training data at the beginning of each epoch (ArrayFileName_Train, ArrayLabel_Train) = ShuffleArray(ArrayLabel_Train.Length, ArrayFileName_Train, ArrayLabel_Train); y_train = np.eye(Dict_Label.Count)[new NDArray(ArrayLabel_Train)]; //decay learning rate if (learning_rate_step != 0) { if ((epoch != 0) && (epoch % learning_rate_step == 0)) { learning_rate_base = learning_rate_base * learning_rate_decay; if (learning_rate_base <= learning_rate_min) { learning_rate_base = learning_rate_min; } sess.run(tf.assign(learning_rate, learning_rate_base)); } } //Load local images asynchronously,use queue,improve train efficiency BlockingCollection<(NDArray c_x, NDArray c_y, int iter)> BlockC = new BlockingCollection<(NDArray C1, NDArray C2, int iter)>(TrainQueueCapa); Task.Run(() => { foreach (var iteration in range(num_tr_iter)) { var start = iteration * batch_size; var end = (iteration + 1) * batch_size; (NDArray x_batch, NDArray y_batch) = GetNextBatch(sess, ArrayFileName_Train, y_train, start, end); BlockC.Add((x_batch, y_batch, iteration)); } BlockC.CompleteAdding(); }); foreach (var item in BlockC.GetConsumingEnumerable()) { sess.run(optimizer, (x, item.c_x), (y, item.c_y)); if (item.iter % display_freq == 0) { // Calculate and display the batch loss and accuracy var result = sess.run(new[] { loss, accuracy }, new FeedItem(x, item.c_x), new FeedItem(y, item.c_y)); loss_val = result[0]; accuracy_val = result[1]; print("CNN：" + ($"iter {item.iter.ToString("000")}: Loss={loss_val.ToString("0.0000")}, Training Accuracy={accuracy_val.ToString("P")} {sw.ElapsedMilliseconds}ms")); sw.Restart(); } } // Run validation after every epoch (loss_val, accuracy_val) = sess.run((loss, accuracy), (x, x_valid), (y, y_valid)); print("CNN：" + "---------------------------------------------------------"); print("CNN：" + $"gloabl steps: {sess.run(gloabl_steps) },learning rate: {sess.run(learning_rate)}, validation loss: {loss_val.ToString("0.0000")}, validation accuracy: {accuracy_val.ToString("P")}"); print("CNN：" + "---------------------------------------------------------"); if (SaverBest) { if (accuracy_val > max_accuracy) { max_accuracy = accuracy_val; saver.save(sess, path_model + "\\CNN_Best"); print("CKPT Model is save."); } } else { saver.save(sess, path_model + string.Format("\\CNN_Epoch_{0}_Loss_{1}_Acc_{2}", epoch, loss_val, accuracy_val)); print("CKPT Model is save."); } } Write_Dictionary(path_model + "\\dic.txt", Dict_Label);}private void Write_Dictionary(string path, Dictionary<Int64, string> mydic){ FileStream fs = new FileStream(path, FileMode.Create); StreamWriter sw = new StreamWriter(fs); foreach (var d in mydic) { sw.Write(d.Key + "," + d.Value + "\r\n"); } sw.Flush(); sw.Close(); fs.Close(); print("Write_Dictionary");}private (NDArray, NDArray) Randomize(NDArray x, NDArray y){ var perm = np.random.permutation(y.shape[0]); np.random.shuffle(perm); return (x[perm], y[perm]);}private (NDArray, NDArray) GetNextBatch(NDArray x, NDArray y, int start, int end){ var slice = new Slice(start, end); var x_batch = x[slice]; var y_batch = y[slice]; return (x_batch, y_batch);}private unsafe (NDArray, NDArray) GetNextBatch(Session sess, string[] x, NDArray y, int start, int end){ NDArray x_batch = np.zeros(end - start, img_h, img_w, n_channels); int n = 0; for (int i = start; i < end; i++) { NDArray img4 = cv2.imread(x[i], IMREAD_COLOR.IMREAD_GRAYSCALE); x_batch[n] = sess.run(normalized, (decodeJpeg, img4)); n++; } var slice = new Slice(start, end); var y_batch = y[slice]; return (x_batch, y_batch);}#endregion
训练完成的模型对test数据集进行预测，并统计准确率
计算图中增加了一个提取预测结果Top-1的概率的节点，最后测试集预测的时候可以把详细的预测数据进行输出，方便实际工程中进行调试和优化。
public void Test(Session sess){ (loss_test, accuracy_test) = sess.run((loss, accuracy), (x, x_test), (y, y_test)); print("CNN：" + "---------------------------------------------------------"); print("CNN：" + $"Test loss: {loss_test.ToString("0.0000")}, test accuracy: {accuracy_test.ToString("P")}"); print("CNN：" + "---------------------------------------------------------"); (Test_Cls, Test_Data) = sess.run((cls_prediction, prob), (x, x_test));}private void TestDataOutput(){ for (int i = 0; i < ArrayLabel_Test.Length; i++) { Int64 real = ArrayLabel_Test[i]; int predict = (int)(Test_Cls[i]); var probability = Test_Data[i, predict]; string result = (real == predict) ? "OK" : "NG"; string fileName = ArrayFileName_Test[i]; string real_str = Dict_Label[real]; string predict_str = Dict_Label[predict]; print((i + 1).ToString() + "|" + "result:" + result + "|" + "real_str:" + real_str + "|" + "predict_str:" + predict_str + "|" + "probability:" + probability.GetSingle().ToString() + "|" + "fileName:" + fileName); }}
本文主要是.NET下的TensorFlow在实际工业现场视觉检测项目中的应用，使用SciSharp的TensorFlow.NET构建了简单的CNN图像分类模型，该模型包含输入层、卷积与池化层、扁平化层、全连接层和输出层，这些层都是CNN分类模型的必要的层，针对工业现场的实际图像进行了分类，分类准确性较高。
完整代码可以直接用于大家自己的数据集进行训练，已经在工业现场经过大量测试，可以在GPU或CPU环境下运行，只需要更换tensorflow.dll文件即可实现训练环境的切换。
同时，训练完成的模型文件，可以使用 “CKPT+Meta” 或 冻结成“PB” 2种方式，进行现场的部署，模型部署和现场应用推理可以全部在.NET平台下进行，实现工业现场程序的无缝对接。摆脱了以往Python下 需要通过Flask搭建服务器进行数据通讯交互 的方式，现场部署应用时无需配置Python和TensorFlow的环境【无需对工业现场的原有PC升级安装一大堆环境】，整个过程全部使用传统的.NET的DLL引用的方式。
欢迎广大.NET开发者们加入TensorFlow.NET社区，SciSharp STACK QQ群：461855582 ，或有任何问题可以直接联系我的个人QQ：50705111 。
SciSharp STACK QQ群：
我的个人QQ：
联系客服
