C++技術中的機器學習：使用C++訓練機器學習模型的最佳實踐

#include <tensorflow/core/framework/tensor.h>
#include <tensorflow/core/framework/tensor_shape.h>
#include <tensorflow/core/lib/io/path.h>
#include <tensorflow/core/public/session.h>

using namespace tensorflow;

int main() {
  // 创建会话
  Session* session = NewSession(SessionOptions());

  // 准备训练数据
  float training_data[6][2] = {
    {1, 1}, {2, 2}, {3, 3}, {4, 4}, {5, 5}, {6, 6}
  };
  float training_labels[6] = {2.0f, 4.0f, 6.0f, 8.0f, 10.0f, 12.0f};
  Tensor training_x(DT_FLOAT, TensorShape({6, 2}));
  Tensor training_y(DT_FLOAT, TensorShape({6}));
  memcpy(training_x.flat<float>().data(), training_data, sizeof(training_data));
  memcpy(training_y.flat<float>().data(), training_labels, sizeof(training_labels));

  // 构建模型
  GraphDef graph_def;
  auto status = ReadBinaryProto(Env::Default(), "model.pb", &graph_def);
  if (!status.ok()) throw std::runtime_error(status.message());
  status = session->Create(graph_def);
  if (!status.ok()) throw std::runtime_error(status.message());

  // 训练模型
  std::vector<std::pair<string, Tensor>> inputs = {
    {"x", training_x}, {"y", training_y}
  };
  std::vector<string> outputs = {"loss"};
  std::vector<Tensor> out;
  while (true) {
    session->Run(inputs, outputs, {}, &out);
    if (out[0].scalar<float>()() < 0.01) break;
  }

  // 保存模型
  string output_path = io::JoinPath("saved_model", "export");
  if (!io::gfile::Exists(output_path)) io::gfile::MakeDirectories(output_path);
  status = session->Run({}, {}, {"model"}, &out);
  if (!status.ok()) throw std::runtime_error(status.message());
  const Tensor& saved_model = out[0];
  io::gfile::DeleteRecursively(output_path, io::gfile::Recurse::kRecurse);
  string path = SavedModelUtil::WriteSavedModel(saved_model, output_path);
  if (!path.empty()) {
    std::cout << "模型已保存至 " << path << std::endl;
  }

  // 清理
  session->Close();
  delete session;
  return 0;
}