[Feature] Add onnxruntime inference

deploy/onnxruntime_cpp/CMakeLists.txt

@@ -0,0 +1,28 @@
+cmake_minimum_required(VERSION 3.0)
+project(onnxruntime_cpp_inference_demo CXX C)
+
+set(onnxruntime_INSTALL_PREFIX /usr/local)
+set(onnxruntime_INCLUDE_DIRS
+  ${onnxruntime_INSTALL_PREFIX}/include/onnxruntime
+  ${onnxruntime_INSTALL_PREFIX}/include/onnxruntime/core/session
+)
+
+find_library(onnxruntime_LIBS NAMES onnxruntime PATHS /usr/local/lib)
+
+find_package(OpenCV REQUIRED)
+
+add_executable(${PROJECT_NAME}_app
+  ${PROJECT_SOURCE_DIR}/src/test_seg.cpp
+  ${PROJECT_SOURCE_DIR}/src/ort_session_handler.cpp
+)
+
+target_link_libraries(${PROJECT_NAME}_app
+  ${OpenCV_LIBRARIES}
+  ${onnxruntime_LIBS}
+)
+
+target_include_directories(${PROJECT_NAME}_app
+  SYSTEM PUBLIC
+    ${OpenCV_INCLUDE_DIRS}
+    ${onnxruntime_INCLUDE_DIRS}
+)

deploy/onnxruntime_cpp/README.md

@@ -0,0 +1,56 @@
+# sample deploy app using [onnxruntime](https://github.com/microsoft/onnxruntime) #
+***
+
+This sample uses Cpp api of onnxruntime.
+
+## Dependencies ##
+***
+
+- [onnxruntime](https://github.com/microsoft/onnxruntime)
+  * tested with v1.10.0
+  * to build onnxruntime from source you need cmake>=3.18
+  * you can build from source with the [build script](https://github.com/microsoft/onnxruntime/blob/master/build.sh)
+  * here is the sample procedure to build cpu onnxruntime
+  ```bash
+    readonly ONNXRUNTIME_VERSION="v1.10.0"
+    git clone --recursive -b ${ONNXRUNTIME_VERSION} https://github.com/Microsoft/onnxruntime
+    ./build.sh --config RelWithDebInfo --build_shared_lib --skip_tests --parallel `nproc`
+    cd build/Linux/RelWithDebInfo
+    make install
+    ```
+
+- opencv
+```bash
+sudo apt-get install libopencv-dev
+```
+
+## How to Build ##
+***
+```bash
+mkdir build && cd build && cmake ../ && make -j`nproc`
+```
+
+## How to Run ##
+***
+
+- Download test images
+```bash
+wget https://paddleseg.bj.bcebos.com/dygraph/demo/cityscapes_demo.png
+```
+
+- Export PaddleSeg Model to onnx format
+  * [export PaddleSeg model](https://github.com/PaddlePaddle/PaddleSeg/blob/develop/docs/model_export.md)
+  * convert exported model to onnx format with [Paddle2ONNX](https://github.com/PaddlePaddle/Paddle2ONNX)
+
+- You can download the bisenetv2 onnx model for this demo directly from the following url:
+```bash
+wget https://github.com/PaddlePaddle/PaddleSeg/files/7918707/bisenetv2_cityscapes.zip && unzip bisenetv2_cityscapes.zip
+```
+
+- Run app
+
+```bash
+./build/onnxruntime_cpp_inference_demo_app /path/to/test/image /path/to/onnx/model
+```
+
+The result will be saved as *out_img.jpg*

deploy/onnxruntime_cpp/src/ort_session_handler.cpp

@@ -0,0 +1,54 @@
+#include "ort_session_handler.hpp"
+
+namespace deploy {
+OrtSessionHandler::OrtSessionHandler(const std::string &model_path,
+                                     const std::vector<std::vector<int64_t>> &input_tensor_shapes)
+    : _model_path(model_path),
+      _input_tensor_shapes(input_tensor_shapes),
+      _env(Ort::Env(ORT_LOGGING_LEVEL_WARNING, "ort session handler")),
+      _session(nullptr) {
+  Ort::SessionOptions session_options;
+
+  // if onnxruntime is built with cuda provider, the following function can be added to use cuda gpu
+  // Ort::ThrowOnError(OrtSessionOptionsAppendExecutionProvider_CUDA(session_options, gpu_index));
+
+  std::basic_string<ORTCHAR_T> ort_model_path;
+  std::copy(model_path.begin(), model_path.end(), std::back_inserter(ort_model_path));
+  _session.reset(new Ort::Experimental::Session(_env, ort_model_path, session_options));
+
+  if (_session->GetInputCount() != input_tensor_shapes.size()) {
+    throw std::runtime_error("invalid input size");
+  }
+}
+
+std::vector<float> OrtSessionHandler::preprocess(const cv::Mat &image, int target_height, int target_width,
+                                                 const std::vector<float> &mean_val,
+                                                 const std::vector<float> &std_val) const {
+  if (image.empty() || image.channels() != 3) {
+    throw std::runtime_error("invalid image");
+  }
+
+  if (target_height * target_width == 0) {
+    throw std::runtime_error("invalid target dimension");
+  }
+
+  cv::Mat processed = image.clone();
+
+  if (image.rows != target_height || image.cols != target_width) {
+    cv::resize(processed, processed, cv::Size(target_width, target_height), 0, 0, cv::INTER_CUBIC);
+  }
+  cv::cvtColor(processed, processed, cv::COLOR_BGR2RGB);
+  std::vector<float> data(3 * target_height * target_width);
+
+  for (int i = 0; i < target_height; ++i) {
+    for (int j = 0; j < target_width; ++j) {
+      for (int c = 0; c < 3; ++c) {
+        data[c * target_height * target_width + i * target_width + j] =
+            (processed.data[i * target_width * 3 + j * 3 + c] / 255.0 - mean_val[c]) / std_val[c];
+      }
+    }
+  }
+
+  return data;
+}
+}  // namespace deploy

deploy/onnxruntime_cpp/src/ort_session_handler.hpp

@@ -0,0 +1,59 @@
+#include <onnxruntime/core/session/experimental_onnxruntime_cxx_api.h>
+
+// if onnxruntime is built with cuda provider, the following header can be added to use cuda gpu
+// #include <onnxruntime/core/providers/cuda/cuda_provider_factory.h>
+
+#include <opencv2/opencv.hpp>
+
+namespace deploy {
+class OrtSessionHandler {
+ public:
+  /**
+   *  @param model_path path to onnx model
+   */
+  OrtSessionHandler(const std::string &model_path, const std::vector<std::vector<int64_t>> &input_tensor_shapes);
+
+  virtual std::vector<float> preprocess(const cv::Mat &image, int target_height, int target_width,
+                                        const std::vector<float> &mean_val = {0.5, 0.5, 0.5},
+                                        const std::vector<float> &std_val = {0.5, 0.5, 0.5}) const;
+
+  /**
+   *   @file function to get output tensors
+   *   @brief each std::pair<DataType *, std::vector<int64_t>> is a pair of output tensor's data and its dimension
+   *   most semantic segmentation networks will have only one output tensor
+   */
+  template <typename DataType = float>
+  std::vector<std::pair<DataType *, std::vector<int64_t>>> run(const std::vector<std::vector<float>> &input_data) const;
+
+ private:
+  std::string _model_path;
+  std::vector<std::vector<int64_t>> _input_tensor_shapes;
+  Ort::Env _env;
+  std::unique_ptr<Ort::Experimental::Session> _session;
+};
+
+template <typename DataType>
+std::vector<std::pair<DataType *, std::vector<int64_t>>> OrtSessionHandler::run(
+    const std::vector<std::vector<float>> &input_data) const {
+  if (_session->GetInputCount() != input_data.size()) {
+    throw std::runtime_error("invalid input size");
+  }
+
+  std::vector<Ort::Value> input_tensors;
+  for (int i = 0; i < _session->GetInputCount(); ++i) {
+    input_tensors.emplace_back(Ort::Experimental::Value::CreateTensor<float>(
+        const_cast<float *>(input_data[i].data()), input_data[i].size(), _input_tensor_shapes[i]));
+  }
+
+  std::vector<Ort::Value> output_tensors =
+      _session->Run(_session->GetInputNames(), input_tensors, _session->GetOutputNames());
+
+  std::vector<std::pair<DataType *, std::vector<int64_t>>> output(_session->GetOutputCount());
+  std::vector<std::vector<int64_t>> output_shapes = _session->GetOutputShapes();
+  for (int i = 0; i < _session->GetOutputCount(); ++i) {
+    output[i] = std::make_pair(std::move(output_tensors[i].GetTensorMutableData<DataType>()), output_shapes[i]);
+  }
+
+  return output;
+}
+}  // namespace deploy

deploy/onnxruntime_cpp/src/test_seg.cpp

@@ -0,0 +1,53 @@
+#include "ort_session_handler.hpp"
+
+namespace {
+constexpr int BISENETV2_CITYSCAPES_IMAGE_HEIGHT = 1024;
+constexpr int BISENETV2_CITYSCAPES_IMAGE_WIDTH = 1024;
+
+static const std::vector<std::vector<uint8_t>> CITYSCAPES_COLORS = {
+    {128, 64, 128}, {244, 35, 232}, {70, 70, 70},    {102, 102, 156}, {190, 153, 153}, {153, 153, 153}, {250, 170, 30},
+    {220, 220, 0},  {107, 142, 35}, {152, 251, 152}, {70, 130, 180},  {220, 20, 60},   {255, 0, 0},     {0, 0, 142},
+    {0, 0, 70},     {0, 60, 100},   {0, 80, 100},    {0, 0, 230},     {119, 11, 32}};
+}  // namespace
+
+int main(int argc, char *argv[]) {
+  if (argc != 3) {
+    std::cerr << "Usage: [app] [/path/to/image] [path/to/onnx/model]" << std::endl;
+    return EXIT_FAILURE;
+  }
+  const std::string image_path = argv[1];
+  cv::Mat image = cv::imread(image_path);
+
+  if (image.empty()) {
+    std::cerr << "failed to load " << image_path << std::endl;
+    return EXIT_FAILURE;
+  }
+
+  const std::string onnx_model_path = argv[2];
+
+  std::vector<std::vector<int64_t>> input_tensor_shapes{
+      {1, 3, BISENETV2_CITYSCAPES_IMAGE_HEIGHT, BISENETV2_CITYSCAPES_IMAGE_WIDTH}};
+  deploy::OrtSessionHandler ort_session_handler(onnx_model_path, input_tensor_shapes);
+  std::vector<float> input_data =
+      ort_session_handler.preprocess(image, BISENETV2_CITYSCAPES_IMAGE_HEIGHT, BISENETV2_CITYSCAPES_IMAGE_WIDTH);
+
+  // output data's type might change for each different model
+  auto output_data = ort_session_handler.run<int64_t>({input_data});
+
+  // postprocess
+  // this might change for each different model
+  cv::Mat segm(BISENETV2_CITYSCAPES_IMAGE_HEIGHT, BISENETV2_CITYSCAPES_IMAGE_WIDTH, CV_8UC(3));
+  for (int i = 0; i < BISENETV2_CITYSCAPES_IMAGE_HEIGHT; ++i) {
+    cv::Vec3b *ptr_segm = segm.ptr<cv::Vec3b>(i);
+    for (int j = 0; j < BISENETV2_CITYSCAPES_IMAGE_WIDTH; ++j) {
+      const auto &color = CITYSCAPES_COLORS[output_data[0].first[i * BISENETV2_CITYSCAPES_IMAGE_WIDTH + j]];
+      ptr_segm[j] = cv::Vec3b(color[0], color[1], color[2]);
+    }
+  }
+  cv::resize(segm, segm, image.size(), 0, 0, cv::INTER_NEAREST);
+  float blended_alpha = 0.4;
+  segm = (1 - blended_alpha) * image + blended_alpha * segm;
+  cv::imwrite("out_img.jpg", segm);
+
+  return EXIT_SUCCESS;
+}