1.确定要提供服务的inference的input,output,以及exporter的signature；（这里用classify的signature做例子，input为byte数组，output为float数组）
2.编写proto文件
syntax = "proto3";option java_package = "com.lenovo.tensorflow";option java_outer_classname = "ServingApiProtocol";option java_generic_services = true;option java_generate_equals_and_hash = true;package com.lenovo;message ClassificationRequest { bytes imageData = 1 ;}message ClassificationResponse { repeated float score = 1;}service ClassificationService { rpc classify(ClassificationRequest) returns (ClassificationResponse) {}} |
2.生成相关代码（这里client为java,server为c++）
具体操作参考grpc.io的文档
3.编写inference C++代码
/* Copyright 2016 Google Inc. All Rights Reserved.Licensed under the Apache License, Version 2.0 (the "License");you may not use this file except in compliance with the License.You may obtain a copy of the License atUnless required by applicable law or agreed to in writing, softwaredistributed under the License is distributed on an "AS IS" BASIS,WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.See the License for the specific language governing permissions andlimitations under the License.==============================================================================*/// A gRPC server that classifies images into digit 0-9.// Given each request with an image pixels encoded as floats, the server// responds with 10 float values as probabilities for digit 0-9 respectively.// The classification is done by running image data through a convolutional// network trained and exported by mnist_model.py.// The server constantly monitors a file system storage path for models.// Whenever a new version of model is available, it eagerly unloads older// version before loading the new one. The server also batches multiple// requests together and does batched inference for efficiency.// The intention of this example to demonstrate usage of DymanicManager,// VersionPolicy and BasicBatchScheduler.#include <stddef.h>#include <algorithm>#include <memory>#include <string>#include <vector>#include "grpc++/completion_queue.h"#include "grpc++/security/server_credentials.h"#include "grpc++/server.h"#include "grpc++/server_builder.h"#include "grpc++/server_context.h"#include "grpc++/support/async_unary_call.h"#include "grpc++/support/status.h"#include "grpc++/support/status_code_enum.h"#include "grpc/grpc.h"#include "tensorflow/core/framework/tensor.h"#include "tensorflow/core/framework/tensor_types.h"#include "tensorflow/core/framework/types.pb.h"#include "tensorflow/core/lib/core/status.h"#include "tensorflow/core/lib/strings/strcat.h"#include "tensorflow/core/platform/env.h"#include "tensorflow/core/platform/init_main.h"#include "tensorflow/core/platform/types.h"#include "tensorflow/core/util/command_line_flags.h"#include "tensorflow_serving/batching/basic_batch_scheduler.h"#include "tensorflow_serving/batching/batch_scheduler.h"#include "tensorflow_serving/core/manager.h"#include "tensorflow_serving/core/servable_handle.h"#include "tensorflow_serving/core/servable_id.h"#include "tensorflow_serving/lenovo/serving-api.pb.h"#include "tensorflow_serving/lenovo/serving-api.grpc.pb.h"#include "tensorflow_serving/servables/tensorflow/simple_servers.h"#include "tensorflow_serving/session_bundle/manifest.pb.h"#include "tensorflow_serving/session_bundle/session_bundle.h"#include "tensorflow_serving/session_bundle/signature.h"using grpc::InsecureServerCredentials;using grpc::Server;using grpc::ServerAsyncResponseWriter;using grpc::ServerBuilder;using grpc::ServerContext;using grpc::ServerCompletionQueue;using grpc::Status;using grpc::StatusCode;using tensorflow::string;using tensorflow::Tensor;using tensorflow::serving::ClassificationSignature;using com::lenovo::ClassificationRequest;using com::lenovo::ClassificationResponse;using com::lenovo::ClassificationService;namespace {const int kImageSize = 107;const int kNumChannels = 3;const int kImageDataSize = kImageSize * kImageSize * kNumChannels;const int kNumLabels = 2;class LenovoServiceImpl;// Class encompassing the state and logic needed to serve a request.class CallData { public: CallData(LenovoServiceImpl* service_impl, ClassificationService::AsyncService* service, ServerCompletionQueue* cq); void Proceed(); void Finish(Status status); const ClassificationRequest& request() { return request_; } ClassificationResponse* mutable_response() { return &response_; } private: // Service implementation. LenovoServiceImpl* service_impl_; // The means of communication with the gRPC runtime for an asynchronous // server. ClassificationService::AsyncService* service_; // The producer-consumer queue where for asynchronous server notifications. ServerCompletionQueue* cq_; // Context for the rpc, allowing to tweak aspects of it such as the use // of compression, authentication, as well as to send metadata back to the // client. ServerContext ctx_; // What we get from the client. ClassificationRequest request_; // What we send back to the client. ClassificationResponse response_; // The means to get back to the client. ServerAsyncResponseWriter<ClassificationResponse> responder_; // Let's implement a tiny state machine with the following states. enum CallStatus { CREATE, PROCESS, FINISH }; CallStatus status_; // The current serving state.};// A Task holds all of the information for a single inference request.struct Task : public tensorflow::serving::BatchTask { ~Task() override = default; size_t size() const override { return 1; } Task(CallData* calldata_arg) : calldata(calldata_arg) {} CallData* calldata;};class LenovoServiceImpl final { public: LenovoServiceImpl(const string& servable_name, std::unique_ptr<tensorflow::serving::Manager> manager); void Classify(CallData* call_data); // Produces classifications for a batch of requests and associated responses. void DoClassifyInBatch( std::unique_ptr<tensorflow::serving::Batch<Task>> batch); // Name of the servable to use for inference. const string servable_name_; // Manager in charge of loading and unloading servables. std::unique_ptr<tensorflow::serving::Manager> manager_; // A scheduler for batching multiple request calls into single calls to // Session->Run(). std::unique_ptr<tensorflow::serving::BasicBatchScheduler<Task>> batch_scheduler_;};// Take in the "service" instance (in this case representing an asynchronous// server) and the completion queue "cq" used for asynchronous communication// with the gRPC runtime.CallData::CallData(LenovoServiceImpl* service_impl, ClassificationService::AsyncService* service, ServerCompletionQueue* cq) : service_impl_(service_impl), service_(service), cq_(cq), responder_(&ctx_), status_(CREATE) { // Invoke the serving logic right away. Proceed();}void CallData::Proceed() { if (status_ == CREATE) { // As part of the initial CREATE state, we *request* that the system // start processing Classify requests. In this request, "this" acts are // the tag uniquely identifying the request (so that different CallData // instances can serve different requests concurrently), in this case // the memory address of this CallData instance. service_->Requestclassify(&ctx_, &request_, &responder_, cq_, cq_, this); // Make this instance progress to the PROCESS state. status_ = PROCESS; } else if (status_ == PROCESS) { // Spawn a new CallData instance to serve new clients while we process // the one for this CallData. The instance will deallocate itself as // part of its FINISH state. new CallData(service_impl_, service_, cq_); // Start processing. service_impl_->Classify(this); } else { GPR_ASSERT(status_ == FINISH); // Once in the FINISH state, deallocate ourselves (CallData). delete this; }}void CallData::Finish(Status status) { status_ = FINISH; responder_.Finish(response_, status, this);}LenovoServiceImpl::LenovoServiceImpl( const string& servable_name, std::unique_ptr<tensorflow::serving::Manager> manager) : servable_name_(servable_name), manager_(std::move(manager)) { // Setup a batcher used to combine multiple requests (tasks) into a single // graph run for efficiency. // The batcher queues tasks until, // (a) the next task would cause the batch to exceed the size target; // (b) waiting for more tasks to be added would exceed the timeout. // at which point it processes the entire batch. // // Use the default batch-size, timeout and thread options. In general // the numbers are extremely performance critical and should be tuned based // specific graph structure and usage. tensorflow::serving::BasicBatchScheduler<Task>::Options scheduler_options; scheduler_options.thread_pool_name = "mnist_service_batch_threads"; // Use a very large queue, to avoid rejecting requests. (Note: a production // server with load balancing may want to use the default, much smaller, // value.) scheduler_options.max_enqueued_batches = 1; TF_CHECK_OK(tensorflow::serving::BasicBatchScheduler<Task>::Create( scheduler_options, [this](std::unique_ptr<tensorflow::serving::Batch<Task>> batch) { this->DoClassifyInBatch(std::move(batch)); }, &batch_scheduler_));}// Creates a gRPC Status from a TensorFlow Status.Status ToGRPCStatus(const tensorflow::Status& status) { return Status(static_cast<grpc::StatusCode>(status.code()), status.error_message());}// WARNING(break-tutorial-inline-code): The following code snippet is// in-lined in tutorials, please update tutorial documents accordingly// whenever code changes.void LenovoServiceImpl::Classify(CallData* calldata) { // Verify input. /** FIXME if (calldata->request().image_data_size() != kImageDataSize) { calldata->Finish( Status(StatusCode::INVALID_ARGUMENT, tensorflow::strings::StrCat( "expected image_data of size ", kImageDataSize, ", got ", calldata->request().image_data_size()))); return; } */ // Create and submit a task to the batch scheduler. std::unique_ptr<Task> task(new Task(calldata)); tensorflow::Status status = batch_scheduler_->Schedule(&task); if (!status.ok()) { calldata->Finish(ToGRPCStatus(status)); return; }}// Produces classifications for a batch of requests and associated responses.void LenovoServiceImpl::DoClassifyInBatch( std::unique_ptr<tensorflow::serving::Batch<Task>> batch) { batch->WaitUntilClosed(); if (batch->empty()) { return; } const int batch_size = batch->num_tasks(); // Replies to each task with the given error status. auto complete_with_error = [&batch](StatusCode code, const string& msg) { Status status(code, msg); for (int i = 0; i < batch->num_tasks(); i++) { Task* task = batch->mutable_task(i); task->calldata->Finish(status); } }; // Get a handle to the SessionBundle. The handle ensures the Manager does // not reload this while it is in use. // WARNING(break-tutorial-inline-code): The following code snippet is // in-lined in tutorials, please update tutorial documents accordingly // whenever code changes. auto handle_request = tensorflow::serving::ServableRequest::Latest(servable_name_); tensorflow::serving::ServableHandle<tensorflow::serving::SessionBundle> bundle; const tensorflow::Status lookup_status = manager_->GetServableHandle(handle_request, &bundle); if (!lookup_status.ok()) { complete_with_error(StatusCode::INTERNAL, lookup_status.error_message()); return; } // Get the default signature of the graph. Expected to be a // classification signature. tensorflow::serving::ClassificationSignature signature; const tensorflow::Status signature_status = GetClassificationSignature(bundle->meta_graph_def, &signature); if (!signature_status.ok()) { complete_with_error(StatusCode::INTERNAL, signature_status.error_message()); return; } // Transform protobuf input to inference input tensor. // See mnist_model.py for details. // WARNING(break-tutorial-inline-code): The following code snippet is // in-lined in tutorials, please update tutorial documents accordingly // whenever code changes. Tensor input(tensorflow::DT_STRING, {batch_size, kImageDataSize}); auto dst = input.flat_outer_dims<float>().data(); for (int i = 0; i < batch_size; ++i) { std::copy_n( batch->mutable_task(i)->calldata->request().imagedata().begin(), kImageDataSize, dst); dst += kImageDataSize; } // Run classification. tensorflow::Tensor scores; const tensorflow::Status run_status = RunClassification(signature, input, bundle->session.get(), nullptr /* classes */, &scores); if (!run_status.ok()) { complete_with_error(StatusCode::INTERNAL, run_status.error_message()); return; } if (scores.dtype() != tensorflow::DT_FLOAT) { complete_with_error( StatusCode::INTERNAL, tensorflow::strings::StrCat( "Expected output Tensor of DT_FLOAT. Got: ", tensorflow::DataType_Name(scores.dtype()))); return; } if (scores.dim_size(1) != kNumLabels) { complete_with_error( StatusCode::INTERNAL, tensorflow::strings::StrCat( "Expected ", kNumLabels, " labels in each output. Got: ", scores.dim_size(1))); return; } // Transform inference output tensor to protobuf output. // See mnist_model.py for details. const auto& scores_mat = scores.matrix<float>(); for (int i = 0; i < batch_size; ++i) { auto calldata = batch->mutable_task(i)->calldata; for (int c = 0; c < scores.dim_size(1); ++c) { calldata->mutable_response()->add_score(scores_mat(i, c)); } calldata->Finish(Status::OK); }}void HandleRpcs(LenovoServiceImpl* service_impl, ClassificationService::AsyncService* service, ServerCompletionQueue* cq) { // Spawn a new CallData instance to serve new clients. new CallData(service_impl, service, cq); void* tag; // uniquely identifies a request. bool ok; while (true) { // Block waiting to read the next event from the completion queue. The // event is uniquely identified by its tag, which in this case is the // memory address of a CallData instance. cq->Next(&tag, &ok); GPR_ASSERT(ok); static_cast<CallData*>(tag)->Proceed(); }}// Runs MnistService server until shutdown.void RunServer(const int port, const string& servable_name, std::unique_ptr<tensorflow::serving::Manager> manager) { // "0.0.0.0" is the way to listen on localhost in gRPC. const string server_address = "0.0.0.0:" + std::to_string(port); ClassificationService::AsyncService service; ServerBuilder builder; std::shared_ptr<grpc::ServerCredentials> creds = InsecureServerCredentials(); builder.AddListeningPort(server_address, creds); builder.RegisterService(&service); std::unique_ptr<ServerCompletionQueue> cq = builder.AddCompletionQueue(); std::unique_ptr<Server> server(builder.BuildAndStart()); LOG(INFO) << "Running..."; LenovoServiceImpl service_impl(servable_name, std::move(manager)); HandleRpcs(&service_impl, &service, cq.get());}} // namespaceint main(int argc, char** argv) { // Parse command-line options. tensorflow::int32 port = 0; const bool parse_result = tensorflow::ParseFlags(&argc, argv, {tensorflow::Flag("port", &port)}); if (!parse_result) { LOG(FATAL) << "Error parsing command line flags."; } if (argc != 2) { LOG(FATAL) << "Usage: lenovo_inference --port=9000 /path/to/exports"; } const string export_base_path(argv[1]); tensorflow::port::InitMain(argv[0], &argc, &argv); // WARNING(break-tutorial-inline-code): The following code snippet is // in-lined in tutorials, please update tutorial documents accordingly // whenever code changes. std::unique_ptr<tensorflow::serving::Manager> manager; tensorflow::Status status = tensorflow::serving::simple_servers:: CreateSingleTFModelManagerFromBasePath(export_base_path, &manager); TF_CHECK_OK(status) << "Error creating manager"; // Wait until at least one model is loaded. std::vector<tensorflow::serving::ServableId> ready_ids; // TODO(b/25545573): Create a more streamlined startup mechanism than polling. do { LOG(INFO) << "Waiting for models to be loaded..."; tensorflow::Env::Default()->SleepForMicroseconds(1 * 1000 * 1000 /*1 sec*/); ready_ids = manager->ListAvailableServableIds(); } while (ready_ids.empty()); // Run the service. RunServer(port, ready_ids[0].name, std::move(manager)); return 0;} |
4.在源代码目录下编写bzael的BUILD文件
# Description: Tensorflow Serving examples.package( default_visibility = ["//tensorflow_serving:internal"], features = [ "-parse_headers", "no_layering_check", ],)licenses(["notice"]) # Apache 2.0exports_files(["LICENSE"])load("//tensorflow_serving:serving.bzl", "serving_proto_library")filegroup( name = "all_files", srcs = glob( ["**/*"], exclude = [ "**/METADATA", "**/OWNERS", ], ),)serving_proto_library( name = "lenovo_inference_proto", srcs = ["serving-api.proto"], has_services = 1, cc_api_version = 2, cc_grpc_version = 1,)cc_binary( name = "lenovo_inference", srcs = [ "lenovo_inference.cc", ], linkopts = ["-lm"], deps = [ ":lenovo_inference_proto", "//tensorflow_serving/batching:basic_batch_scheduler", "//tensorflow_serving/batching:batch_scheduler", "//tensorflow_serving/core:manager", "//tensorflow_serving/core:servable_handle", "//tensorflow_serving/core:servable_id", "//tensorflow_serving/servables/tensorflow:simple_servers", "//tensorflow_serving/session_bundle", "//tensorflow_serving/session_bundle:manifest_proto", "//tensorflow_serving/session_bundle:signature", "@grpc//:grpc++", "@org_tensorflow//tensorflow/core:framework", "@org_tensorflow//tensorflow/core:lib", "@org_tensorflow//tensorflow/core:protos_all_cc", "@org_tensorflow//tensorflow/core:tensorflow", ],) |
5.编译代码
bazel build tensorflow_serving/lenovo/... |
6.运行server
./bazel-bin/tensorflow_serving/lenovo/lenovo_inference --port=<server port> <export model path> |