Skip to content

Latest commit

 

History

History
353 lines (290 loc) · 23.4 KB

ort_useful_api.zh.md

File metadata and controls

353 lines (290 loc) · 23.4 KB

🔥🔥🔥 全网最详细 ONNXRuntime C++/Java/Python 使用教程 ！

整理不易，持续更新，如果对你有用的话，❤️那就给个⭐️🌟支持一下吧~ 🙃🤪🍀

前言

这份文档主要记录了onnxruntime的一些参考资料，包含C++/Java/Python，以及本人在使用过程中的一些经验。

1. onnxruntime官方资料

2.onnxruntime-C++参考

3. onnxruntime-java参考

4. onnxruntime的Docker镜像

5. onnxruntime源码编译

6. 开源项目

7. 与ONNX的Opset兼容性

与各个转换工具的兼容性

8. 获取Ort::Value的值

8.1 通过At<>获取

TEST(CApiTest, access_tensor_data_elements) {
  /**
   * Create a 2x3 data blob that looks like:
   *
   *  0 1 2
   *  3 4 5
   */
  std::vector<int64_t> shape = {2, 3};
  int element_count = 6;  // 2*3
  std::vector<float> values(element_count);
  for (int i = 0; i < element_count; i++)
    values[i] = static_cast<float>(i);

  Ort::MemoryInfo info("Cpu", OrtDeviceAllocator, 0, OrtMemTypeDefault);

  Ort::Value tensor = Ort::Value::CreateTensor<float>(info, values.data(), values.size(), shape.data(), shape.size());

  float expected_value = 0;
  for (int64_t row = 0; row < shape[0]; row++) {
    for (int64_t col = 0; col < shape[1]; col++) {
      ASSERT_EQ(expected_value++, tensor.At<float>({row, col}));
    }
  }
}

8.2 通过裸指针获取

  const float *var_angles = output_var_tensors.front().GetTensorMutableData<float>();
  const float *conv_angles = output_conv_tensors.front().GetTensorMutableData<float>();
  const float mean_yaw = (var_angles[0] + conv_angles[0]) / 2.0f;
  const float mean_pitch = (var_angles[1] + conv_angles[1]) / 2.0f;
  const float mean_roll = (var_angles[2] + conv_angles[2]) / 2.0f;

8.3 通过引用&和At获取

Ort::Value不允许复制拷贝，因此无法通过=去使用输出的结果，但是可以结合引用&+At<>来获取，使用方式如下:

  ort::Value &var_angles_tensor = output_var_tensors.at(0);
  ort::Value &conv_angles_tensor = output_conv_tensors.at(0);
  const float mean_ref_yaw = 
      (var_angles_tensor.At<float>({0}) + conv_angles_tensor.At<float>({0})) / 2.0f;
  const float mean_ref_pitch = 
      (var_angles_tensor.At<float>({1}) + conv_angles_tensor.At<float>({1})) / 2.0f;
  const float mean_ref_roll =
      (var_angles_tensor.At<float>({2}) + conv_angles_tensor.At<float>({2})) / 2.0f;

可以看到At<>的源码为：

inline T& Value::At(const std::vector<int64_t>& location) {
  static_assert(!std::is_same<T, std::string>::value, "this api does not support std::string");
  T* out;
  ThrowOnError(GetApi().TensorAt(p_, location.data(), location.size(), (void**)&out));
  return *out;
}

实际上也是利用指针p_获取数据，只不过At替我们处理了数据位置的计算。这个方法返回的实际上是一个非const引用，这意味着你可以通过这个引用直接修改内存中的值。

8.4 其他用法参考

std::vector<OrtSessionHandler::DataOutputType> OrtSessionHandler::OrtSessionHandlerIml::
operator()(const std::vector<float*>& inputData)
{
    if (m_numInputs != inputData.size()) {
        throw std::runtime_error("Mismatch size of input data\n");
    }

    Ort::MemoryInfo memoryInfo = Ort::MemoryInfo::CreateCpu(OrtArenaAllocator, OrtMemTypeDefault);

    std::vector<Ort::Value> inputTensors;
    inputTensors.reserve(m_numInputs);

    for (int i = 0; i < m_numInputs; ++i) {
        inputTensors.emplace_back(std::move(
            Ort::Value::CreateTensor<float>(memoryInfo, const_cast<float*>(inputData[i]), m_inputTensorSizes[i],
                                            m_inputShapes[i].data(), m_inputShapes[i].size())));
    }

    auto outputTensors = m_session.Run(Ort::RunOptions{nullptr}, m_inputNodeNames.data(), inputTensors.data(),
                                       m_numInputs, m_outputNodeNames.data(), m_numOutputs);

    assert(outputTensors.size() == m_numOutputs);
    std::vector<DataOutputType> outputData;
    outputData.reserve(m_numOutputs);

    int count = 1;
    for (auto& elem : outputTensors) {
        DEBUG_LOG("type of input %d: %s", count++, toString(elem.GetTensorTypeAndShapeInfo().GetElementType()).c_str());
        outputData.emplace_back(
            std::make_pair(std::move(elem.GetTensorMutableData<float>()), elem.GetTensorTypeAndShapeInfo().GetShape()));
    }

    return outputData;
}

9. 源码应用案例

#include <assert.h>
#include <vector>
#include <onnxruntime_cxx_api.h>

int main(int argc, char* argv[]) {
  Ort::Env env(ORT_LOGGING_LEVEL_WARNING, "test");
  Ort::SessionOptions session_options;
  session_options.SetIntraOpNumThreads(1);
  session_options.SetGraphOptimizationLevel(GraphOptimizationLevel::ORT_ENABLE_EXTENDED);

#ifdef _WIN32
  const wchar_t* model_path = L"model.onnx";
#else
  const char* model_path = "model.onnx";
#endif

  Ort::Session session(env, model_path, session_options);
  // print model input layer (node names, types, shape etc.)
  Ort::AllocatorWithDefaultOptions allocator;

  // print number of model input nodes
  size_t num_input_nodes = session.GetInputCount();
  std::vector<const char*> input_node_names = {"input","input_mask"};
  std::vector<const char*> output_node_names = {"output","output_mask"};
    
  std::vector<int64_t> input_node_dims = {10, 20};
  size_t input_tensor_size = 10 * 20; 
  std::vector<float> input_tensor_values(input_tensor_size);
  for (unsigned int i = 0; i < input_tensor_size; i++)
    input_tensor_values[i] = (float)i / (input_tensor_size + 1);
  // create input tensor object from data values
  auto memory_info = Ort::MemoryInfo::CreateCpu(OrtArenaAllocator, OrtMemTypeDefault);
  Ort::Value input_tensor = Ort::Value::CreateTensor<float>(memory_info, input_tensor_values.data(), input_tensor_size, input_node_dims.data(), 2);
  assert(input_tensor.IsTensor());

  std::vector<int64_t> input_mask_node_dims = {1, 20, 4};
  size_t input_mask_tensor_size = 1 * 20 * 4; 
  std::vector<float> input_mask_tensor_values(input_mask_tensor_size);
  for (unsigned int i = 0; i < input_mask_tensor_size; i++)
    input_mask_tensor_values[i] = (float)i / (input_mask_tensor_size + 1);
  // create input tensor object from data values
  auto mask_memory_info = Ort::MemoryInfo::CreateCpu(OrtArenaAllocator, OrtMemTypeDefault);
  Ort::Value input_mask_tensor = Ort::Value::CreateTensor<float>(mask_memory_info, input_mask_tensor_values.data(), input_mask_tensor_size, input_mask_node_dims.data(), 3);
  assert(input_mask_tensor.IsTensor());
    
  std::vector<Ort::Value> ort_inputs;
  ort_inputs.push_back(std::move(input_tesor));
  ort_inputs.push_back(std::move(input_mask_tensor));
  // score model & input tensor, get back output tensor
  auto output_tensors = session.Run(Ort::RunOptions{nullptr}, input_node_names.data(), ort_inputs.data(), ort_inputs.size(), output_node_names.data(), 2);
  
  // Get pointer to output tensor float values
  float* floatarr = output_tensors[0].GetTensorMutableData<float>();
  float* floatarr_mask = output_tensors[1].GetTensorMutableData<float>();
  
  printf("Done!\n");
  return 0;
}

10. onnxruntime动态维度推理

11. onnxruntime源码学习

12. C++ API使用案例

12.1 目标检测

12.2 人脸识别

12.3 抠图

12.4 人脸检测

12.5 人脸关键点检测

12.6 头部姿态估计

12.7 人脸属性识别

12.8 图像分类

12.9 语义分割

12.10 风格迁移

12.11 图片着色

12.12 超分辨率

【新的模型案例将会不断加入进来！！！欢迎关注收藏~】