• This is a wrapper of deep learning frameworks especially for inference
• This class provides a common interface to use various deep learnig frameworks, so that you can use the same application code

• TensorFlow Lite
• TensorFlow Lite with delegate (XNNPACK, GPU, EdgeTPU, NNAPI)
• TensorRT (GPU, DLA)
• OpenCV(dnn) (with GPU)
• OpenVINO with OpenCV (xml+bin)
• ncnn (with Vulkan)
• MNN (with Vulkan)
• SNPE (Snapdragon Neural Processing Engine SDK (Qualcomm Neural Processing SDK for AI v1.51.0))
• Arm NN
• NNabla (with CUDA)
• ONNX Runtime (with CUDA)
• LibTorch (with CUDA)
• TensorFlow (with GPU)

• Windows 10 (Visual Studio 2019 x64)
• Linux (x64, armv7, aarch64)
• Android (armeabi-v7a, arm64-v8a)

Framework	Windows (x64)	Linux (x64)	Linux (armv7)	Linux (aarch64)	Android (aarch64)
TensorFlow Lite	Build Test	Build Test	Build Test	Build Test	Build Test
TensorFlow Lite + XNNPACK	Build Test	Build Test	Unsupported	Build Test	Build Test
TensorFlow Lite + EdgeTPU	Build Test	Build Test	Build Test	Build Test	Unsupported
TensorFlow Lite + GPU	No library	No library	No library	No library	Build Test
TensorFlow Lite + NNAPI	Unsupported	Unsupported	Unsupported	Unsupported	Build Test
TensorRT	Build Test	Build Test	Build Test	Build Test	Unsupported
OpenCV(dnn)	Build Test	Build Test	Build Test	Build Test	Build Test
OpenVINO with OpenCV	Build Test	Build Test	Build Test	Build Test	Unsupported
ncnn	Build Test	Build Test	No library	No library	Build Test
MNN	Build Test	Build Test	No library	Build Test	Build Test
SNPE	Unsupported	Unsupported	Build Test	Build Test	Build Test
Arm NN	Unsupported	Build Test	Unsupported	Build Test	No library
NNabla	Build Test	No library	Unsupported	No library	No library
ONNX Runtime	Build Test	Build Test	Unsupported	Build Test	Build Test
LibTorch	Build Test	Build Test	No library	No library	No library
TensorFlow	Build Test	Build Test	No library	No library	No library

• Unchedked(blank) doesn't mean that the framework is unsupported. Unchecked just means that the framework is not tested in CI. For instance, TensorRT on Windows/Linux works and I confirmed it in my PC, but can't run it in CI.
• No Library means a pre-built library is not provided so that I cannot confirm it in CI. It may work if you build a library by yourself.

• https://github.com/iwatake2222/InferenceHelper_Sample
• https://github.com/iwatake2222/play_with_tflite
• https://github.com/iwatake2222/play_with_tensorrt
• https://github.com/iwatake2222/play_with_ncnn
• https://github.com/iwatake2222/play_with_mnn

Please refer to https://github.com/iwatake2222/InferenceHelper_Sample

• Add this repository into your project (Using git submodule is recommended)
• Download prebuilt libraries
  • sh third_party/download_prebuilt_libraries.sh

You need some additional steps if you use the frameworks listed below

• Install OpenCV or OpenVINO
  • You may need to set/modify OpenCV_DIR and PATH environment variable
  • To use OpenVINO, you may need to run C:\Program Files (x86)\Intel\openvino_2021\bin\setupvars.bat or source /opt/intel/openvino_2021/bin/setupvars.sh

• Install CUDA + cuDNN
• Install TensorRT 8.x

• Install the following library
  • Linux: https://github.com/google-coral/libedgetpu/releases/download/release-grouper/edgetpu_runtime_20210726.zip
  • Windows: https://github.com/google-coral/libedgetpu/releases/download/release-frogfish/edgetpu_runtime_20210119.zip
    • the latest version doesn't work
    • it may be better to delete C:\Windows\System32\edgetpu.dll to ensure the program uses our pre-built library

• Install Vulkan
  • You need Vulkan even if you don't use it because the pre-built libraries require it. Otherwise you need to build libraries by yourself disabling Vulkan
  • https://vulkan.lunarg.com/sdk/home
  • Windows
    • https://sdk.lunarg.com/sdk/download/latest/windows/vulkan-sdk.exe
    • It's better to check (Optional) Debuggable Shader API Libraries -64-bit , so that you can use Debug in Visual Studio
  • Linux (x64)

    wget https://sdk.lunarg.com/sdk/download/latest/linux/vulkan-sdk.tar.gz
    tar xzvf vulkan-sdk.tar.gz
    export VULKAN_SDK=$(pwd)/1.2.198.1/x86_64
    sudo apt install -y vulkan-utils libvulkan1 libvulkan-dev

• Download library from https://developer.qualcomm.com/software/qualcomm-neural-processing-sdk/tools
• Extract snpe-1.51.0.zip , then place lib and include folders to third_party/snpe_prebuilt

• Debug mode in Visual Studio doesn't work for ncnn, NNabla and LibTorch because debuggable libraries are not provided
  • Debug will cause unexpected bahavior, so use Release or RelWithDebInfo
• See third_party/download_prebuilt_libraries.sh and third_party/cmakes/* to check which libraries are being used. For instance, libraries without GPU(CUDA/Vulkan) are used to be safe. So, if you want to use GPU, modify these files.

• Add InferenceHelper to your project

  set(INFERENCE_HELPER_DIR ${CMAKE_CURRENT_LIST_DIR}/../../InferenceHelper/)
  add_subdirectory(${INFERENCE_HELPER_DIR}/inference_helper inference_helper)
  target_include_directories(${LibraryName} PUBLIC ${INFERENCE_HELPER_DIR}/inference_helper)
  target_link_libraries(${LibraryName} InferenceHelper)

• Deep learning framework:

  • You can enable multiple options althoguh the following example enables just one option

  # OpenCV (dnn), OpenVINO
  cmake .. -DINFERENCE_HELPER_ENABLE_OPENCV=on
  # Tensorflow Lite
  cmake .. -DINFERENCE_HELPER_ENABLE_TFLITE=on
  # Tensorflow Lite (XNNPACK)
  cmake .. -DINFERENCE_HELPER_ENABLE_TFLITE_DELEGATE_XNNPACK=on
  # Tensorflow Lite (GPU)
  cmake .. -DINFERENCE_HELPER_ENABLE_TFLITE_DELEGATE_GPU=on
  # Tensorflow Lite (EdgeTPU)
  cmake .. -DINFERENCE_HELPER_ENABLE_TFLITE_DELEGATE_EDGETPU=on
  # Tensorflow Lite (NNAPI)
  cmake .. -DINFERENCE_HELPER_ENABLE_TFLITE_DELEGATE_NNAPI=on
  # TensorRT
  cmake .. -DINFERENCE_HELPER_ENABLE_TENSORRT=on
  # ncnn, ncnn + vulkan
  cmake .. -DINFERENCE_HELPER_ENABLE_NCNN=on
  # MNN (+ Vulkan)
  cmake .. -DINFERENCE_HELPER_ENABLE_MNN=on
  # SNPE
  cmake .. -DINFERENCE_HELPER_ENABLE_SNPE=on
  # Arm NN
  cmake .. -DINFERENCE_HELPER_ENABLE_ARMNN=on
  # NNabla
  cmake .. -DINFERENCE_HELPER_ENABLE_NNABLA=on
  # NNabla with CUDA
  cmake .. -DINFERENCE_HELPER_ENABLE_NNABLA_CUDA=on
  # ONNX Runtime
  cmake .. -DINFERENCE_HELPER_ENABLE_ONNX_RUNTIME=on
  # ONNX Runtime with CUDA
  cmake .. -DINFERENCE_HELPER_ENABLE_ONNX_RUNTIME_CUDA=on
  # LibTorch
  cmake .. -DINFERENCE_HELPER_ENABLE_LIBTORCH=on
  # LibTorch with CUDA
  cmake .. -DINFERENCE_HELPER_ENABLE_LIBTORCH_CUDA=on
  # TensorFlow
  cmake .. -DINFERENCE_HELPER_ENABLE_TENSORFLOW=on
  # TensorFlow with GPU
  cmake .. -DINFERENCE_HELPER_ENABLE_TENSORFLOW_GPU=on

• Enable/Disable preprocess using OpenCV:

  • By disabling this option, InferenceHelper is not dependent on OpenCV

  cmake .. -INFERENCE_HELPER_ENABLE_PRE_PROCESS_BY_OPENCV=off

typedef enum {
    kOpencv,
    kOpencvGpu,
    kTensorflowLite,
    kTensorflowLiteXnnpack,
    kTensorflowLiteGpu,
    kTensorflowLiteEdgetpu,
    kTensorflowLiteNnapi,
    kTensorrt,
    kNcnn,
    kNcnnVulkan,
    kMnn,
    kSnpe,
    kArmnn,
    kNnabla,
    kNnablaCuda,
    kOnnxRuntime,
    kOnnxRuntimeCuda,
    kLibtorch,
    kLibtorchCuda,
    kTensorflow,
    kTensorflowGpu,
} HelperType;

• Create InferenceHelper instance for the selected framework

std::unique_ptr<InferenceHelper> inference_helper(InferenceHelper::Create(InferenceHelper::kTensorflowLite));

• Run preprocess (convert image to blob(NCHW or NHWC))
• This is just a helper function. You may not use this function.
  • Available when INFERENCE_HELPER_ENABLE_PRE_PROCESS_BY_OPENCV=on

InferenceHelper::PreProcessByOpenCV(input_tensor_info, false, img_blob);

• Set the number of threads to be used
• This function needs to be called before initialize

inference_helper->SetNumThreads(4);

• Set custom ops
• This function needs to be called before initialize

std::vector<std::pair<const char*, const void*>> custom_ops;
custom_ops.push_back(std::pair<const char*, const void*>("Convolution2DTransposeBias", (const void*)mediapipe::tflite_operations::RegisterConvolution2DTransposeBias()));
inference_helper->SetCustomOps(custom_ops);

• Initialize inference helper
  • Load model
  • Set tensor information

std::vector<InputTensorInfo> input_tensor_list;
InputTensorInfo input_tensor_info("input", TensorInfo::TENSOR_TYPE_FP32, false);    /* name, data_type, NCHW or NHWC */
input_tensor_info.tensor_dims = { 1, 224, 224, 3 };
input_tensor_info.data_type = InputTensorInfo::kDataTypeImage;
input_tensor_info.data = img_src.data;
input_tensor_info.image_info.width = img_src.cols;
input_tensor_info.image_info.height = img_src.rows;
input_tensor_info.image_info.channel = img_src.channels();
input_tensor_info.image_info.crop_x = 0;
input_tensor_info.image_info.crop_y = 0;
input_tensor_info.image_info.crop_width = img_src.cols;
input_tensor_info.image_info.crop_height = img_src.rows;
input_tensor_info.image_info.is_bgr = false;
input_tensor_info.image_info.swap_color = false;
input_tensor_info.normalize.mean[0] = 0.485f;   /* https://github.com/onnx/models/tree/master/vision/classification/mobilenet#preprocessing */
input_tensor_info.normalize.mean[1] = 0.456f;
input_tensor_info.normalize.mean[2] = 0.406f;
input_tensor_info.normalize.norm[0] = 0.229f;
input_tensor_info.normalize.norm[1] = 0.224f;
input_tensor_info.normalize.norm[2] = 0.225f;
input_tensor_list.push_back(input_tensor_info);

std::vector<OutputTensorInfo> output_tensor_list;
output_tensor_list.push_back(OutputTensorInfo("MobilenetV2/Predictions/Reshape_1", TensorInfo::TENSOR_TYPE_FP32));

inference_helper->initialize("mobilenet_v2_1.0_224.tflite", input_tensor_list, output_tensor_list);

• Finalize inference helper

inference_helper->Finalize();

• Run preprocess
• Call this function before invoke
• Call this function even if the input data is already pre-processed in order to copy data to memory
• Note : Some frameworks don't support crop, resize. So, it's better to resize image before calling preProcess.

inference_helper->PreProcess(input_tensor_list);

• Run inference

inference_helper->Process(output_tensor_info_list)

enum {
    kTensorTypeNone,
    kTensorTypeUint8,
    kTensorTypeInt8,
    kTensorTypeFp32,
    kTensorTypeInt32,
    kTensorTypeInt64,
};

std::string name;           // [In] Set the name_ of tensor
int32_t     id;             // [Out] Do not modify (Used in InferenceHelper)
int32_t     tensor_type;    // [In] The type of tensor (e.g. kTensorTypeFp32)
std::vector<int32_t> tensor_dims;    // InputTensorInfo:   [In] The dimentions of tensor. (If empty at initialize, the size is updated from model info.)
                                     // OutputTensorInfo: [Out] The dimentions of tensor is set from model information
bool        is_nchw;        // [IN] NCHW or NHWC

enum {
    kDataTypeImage,
    kDataTypeBlobNhwc,  // data_ which already finished preprocess(color conversion, resize, normalize_, etc.)
    kDataTypeBlobNchw,
};

void*   data;      // [In] Set the pointer to image/blob
int32_t data_type; // [In] Set the type of data_ (e.g. kDataTypeImage)

struct {
    int32_t width;
    int32_t height;
    int32_t channel;
    int32_t crop_x;
    int32_t crop_y;
    int32_t crop_width;
    int32_t crop_height;
    bool    is_bgr;        // used when channel == 3 (true: BGR, false: RGB)
    bool    swap_color;
} image_info;              // [In] used when data_type_ == kDataTypeImage

struct {
    float mean[3];
    float norm[3];
} normalize;              // [In] used when data_type_ == kDataTypeImage

void* data;     // [Out] Pointer to the output data_
struct {
    float   scale;
    uint8_t zero_point;
} quant;        // [Out] Parameters for dequantization (convert uint8 to float)

• Get output data in the form of FP32
• When tensor type is INT8 (quantized), the data is converted to FP32 (dequantized)

const float* val_float = output_tensor_list[0].GetDataAsFloat();

• InferenceHelper
• https://github.com/iwatake2222/InferenceHelper
• Copyright 2020 iwatake2222
• Licensed under the Apache License, Version 2.0

• This project utilizes OSS (Open Source Software)
  • NOTICE.md