Tensorflow模型部署

参考:  Tensorflow 模型线上部署
    构建 TensorFlow Serving Java 客户端

• docker安装及部署

  • windows下docker安装

  • tf-serving

      下载tensorflow服务并使用docker部署，这一步如果占用C盘空间太大的话，可以使用Hyper-v工具将下载的镜像转到其他盘

    # 在 cmd 中执行以下命令
    docker pull tensorflow/serving   # 下载镜像
    docker run -itd -p 5000:5000 --name tfserving tensorflow/serving   # 运行镜像并指定镜像名
    docker ps  # 查看镜像id  dockerID
    docker cp ./mnist dockerID:/models  # 将 pb 文件夹拷贝到容器中，模型训练见下面
    
    docker exec -it dockerID /bin/bash  # 进入到镜像里面
    tensorflow_model_server --port=5000 --model_name=mnist --model_base_path=/models/mnist  # 容器内运行服务
    

• 训练模型

    使用官方给出的mnist样例进行训练，改下代码路径就可以，训练得到pb文件如下，并使用 saved_model_cli show --dir ./mnist/1 --all 命令查看节点名称(后面客户端使用)，并将模型复制到docker里面docker cp ./mnist dockerID:/models，此处注意文件夹层级
  

  mnist-pb.png
  sigdef.png
  models.png

• python端

    仿照官方代码 mnist_clien.py编写预测代码

  import grpc
  import tensorflow as tf
  from tensorflow_serving.apis import predict_pb2
  from tensorflow_serving.apis import prediction_service_pb2_grpc
  
  server = 'localhost:5000'
  
  channel = grpc.insecure_channel(server)
  stub = prediction_service_pb2_grpc.PredictionServiceStub(channel)
  request = predict_pb2.PredictRequest()
  request.model_spec.name = 'mnist'
  request.model_spec.signature_name = 'predict_images'
  
  test_data_set = mnist_input_data.read_data_sets('./data').test
  image, label = test_data_set.next_batch(1)
  request.inputs['images'].CopyFrom(tf.make_tensor_proto(image[0], shape=[1, image[0].size]))
  pred = stub.Predict(request, 5.0)
  score = pred.outputs['scores'].float_val
  print(score)
  # [1.6178478001727115e-10, 1.6928293322847278e-15, 1.6151154341059737e-05, 0.000658366538118571, 8.010060947860609e-10, 2.2359495588375466e-08, 3.5608297452131843e-13, 0.9993133544921875, 5.620326870570125e-09, 1.1990837265329901e-05]
  

• Java端

    Java端流程差不多，主要是编译proto麻烦一些

  • proto安装

      windows下proto的安装参考windows之google protobuf安装与使用，下载proto-3.4.0并解压，注意目录不要有空格，否则后面编译会报错，找到protoc.exe所在路径，我的是D:\protoc-3.4.0-win32\bin

  • pom配置编译proto

      此处主要参考构建 TensorFlow Serving Java 客户端，给出的那个proto文件列表太棒了(未理解为什么是这些文件，对java-grpc不熟悉)，仿照其流程，下载tensorflow及tensorflow-serving两个项目，复制相应的proto文件出来

    src/main/proto
    ├── tensorflow
    │   └── core
    │       ├── example
    │       │   ├── example.proto
    │       │   └── feature.proto
    │       ├── framework
    │       │   ├── attr_value.proto
    │       │   ├── function.proto
    │       │   ├── graph.proto
    │       │   ├── node_def.proto
    │       │   ├── op_def.proto
    │       │   ├── resource_handle.proto
    │       │   ├── tensor.proto
    │       │   ├── tensor_shape.proto
    │       │   ├── types.proto
    │       │   └── versions.proto
    │       └── protobuf
    │           ├── meta_graph.proto
    │           └── saver.proto
    └── tensorflow_serving
        └── apis
            ├── classification.proto
            ├── get_model_metadata.proto
            ├── inference.proto
            ├── input.proto
            ├── model.proto
            ├── predict.proto
            ├── prediction_service.proto
            └── regression.proto
    

      创建Maven工程，将上面的proto文件放在src/main下面，在pom中添加以下信息，此处额外添加了编译文件的输入及输出目录，否则会报错 protoc did not exit cleanly

    
        
            
                org.xolstice.maven.plugins
                protobuf-maven-plugin
                0.5.0
                
                    D:\protoc-3.4.0-win32\bin\protoc.exe
                    ${project.basedir}/src/main/proto/
                    ${project.basedir}/src/main/resources/
                
                
                    
                        
                            compile
                            compile-custom
                        
                    
                
            
        
    
    
    
        
            com.google.protobuf
            protobuf-java
            3.11.4
        
        
            io.grpc
            grpc-protobuf
            1.28.0
        
        
            io.grpc
            grpc-stub
            1.28.0
        
        
            io.grpc
            grpc-netty-shaded
            1.28.0
        
    
    

      配置完后，执行maven -> protobuf:compile编译，在resources目录下会生成org及tensorflow两个文件夹，将这两个文件夹复制到src/main/java目录下
    

    proto.png

  • 预测

      编写java程序进行预测，过程中发现没有tensorflow/serving/PredictionServiceGrpc.java这个文件，试了很多方法都没有编译出来，最后是直接把别人的给复制过来了，PredictionServiceGrpc，拷过来后发现报了@java.lang.Override这几行代码提示有问题，直接将override注释掉

    在src/main/java下建表及类，编写预测代码，完整代码如下,运行得预测结果

    package SimpleAdd;
    
    import io.grpc.ManagedChannel;
    import io.grpc.ManagedChannelBuilder;
    import tensorflow.serving.Model;
    import org.tensorflow.framework.DataType;
    import org.tensorflow.framework.TensorProto;
    import org.tensorflow.framework.TensorShapeProto;
    
    import tensorflow.serving.Predict;
    import tensorflow.serving.PredictionServiceGrpc;
    
    
    public class MnistPredict {
        public static void main(String[] args) throws Exception {
            // create a channel for gRPC
            ManagedChannel channel = ManagedChannelBuilder.forAddress("localhost", 5000).usePlaintext().build();
            PredictionServiceGrpc.PredictionServiceBlockingStub stub = PredictionServiceGrpc.newBlockingStub(channel);
    
            // create a modelspec
            Model.ModelSpec.Builder modelSpec = Model.ModelSpec.newBuilder();
            modelSpec.setName("mnist");
            modelSpec.setSignatureName("predict_images");
            Predict.PredictRequest.Builder request = Predict.PredictRequest.newBuilder();
            request.setModelSpec(modelSpec);
    
            // data shape & load data
            TensorShapeProto.Builder shape = TensorShapeProto.newBuilder();
            shape.addDim(TensorShapeProto.Dim.newBuilder().setSize(1));
            shape.addDim(TensorShapeProto.Dim.newBuilder().setSize(784));
            TensorProto.Builder tensor = TensorProto.newBuilder();
            tensor.setTensorShape(shape);
            tensor.setDtype(DataType.DT_FLOAT);
            for(int i=0; i<784; i++){
                tensor.addFloatVal(0);
            }
            request.putInputs("images", tensor.build());
            tensor.clear();
    
            // Predict 
            Predict.PredictResponse response = stub.predict(request.build());
            System.out.println(response);
            TensorProto result = response.toBuilder().getOutputsOrThrow("scores");
            System.out.println("predict: " + result.getFloatValList());
            System.out.println("predict: " + response.getOutputsMap().get("scores").getFloatValList());
            // predict: [0.032191742, 0.09621494, 0.06525445, 0.039610844, 0.05699038, 0.46822935, 0.040578533, 0.1338098, 0.009549928, 0.057570033]
        }
    }