深度学习多框架多平台推理引擎工具

Posted 踟蹰横渡口,彳亍上滩舟。

tags:

篇首语：本文由小常识网(cha138.com)小编为大家整理，主要介绍了深度学习多框架多平台推理引擎工具相关的知识，希望对你有一定的参考价值。

一种深度学习推理引擎工具，支持多框架、支持多平台推理

项目下载地址：下载地址

支持的计算平台：

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

支持的模型框架：

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

下载相关库：
Download prebuilt libraries
- sh third_party/download_prebuilt_libraries.sh

配置编译参数：

  • 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

  • Enable/Disable preprocess using OpenCV:

    • By disabling this option, InferenceHelper is not dependent on OpenCV
    cmake .. -INFERENCE_HELPER_ENABLE_PRE_PROCESS_BY_OPENCV=off

APIs

InferenceHelper

Enumeration

typedef enum 
    kOpencv,
    kOpencvGpu,
    kTensorflowLite,
    kTensorflowLiteXnnpack,
    kTensorflowLiteGpu,
    kTensorflowLiteEdgetpu,
    kTensorflowLiteNnapi,
    kTensorrt,
    kNcnn,
    kNcnnVulkan,
    kMnn,
    kSnpe,
    kArmnn,
    kNnabla,
    kNnablaCuda,
 HelperType;

static InferenceHelper* Create(const HelperType helper_type)

  • Create InferenceHelper instance for the selected framework
std::unique_ptr<InferenceHelper> inference_helper(InferenceHelper::Create(InferenceHelper::kTensorflowLite));

static void PreProcessByOpenCV(const InputTensorInfo& input_tensor_info, bool is_nchw, cv::Mat& img_blob)

  • 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);

int32_t SetNumThreads(const int32_t num_threads)

  • Set the number of threads to be used
  • This function needs to be called before initialize
inference_helper->SetNumThreads(4);

int32_t SetCustomOps(const std::vector<std::pair<const char*, const void*>>& custom_ops)

  • 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);

int32_t Initialize(const std::string& model_filename, std::vector& input_tensor_info_list, std::vector& output_tensor_info_list)

  • 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);

int32_t Finalize(void)

  • Finalize inference helper
inference_helper->Finalize();

int32_t PreProcess(const std::vector& input_tensor_info_list)

  • 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);

int32_t Process(std::vector& output_tensor_info_list)

  • Run inference
inference_helper->Process(output_tensor_info_list)

TensorInfo (InputTensorInfo, OutputTensorInfo)

Enumeration

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

Properties

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

InputTensorInfo

Enumeration

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

Properties

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

OutputTensorInfo

Properties

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

float* GetDataAsFloat()

  • 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();

推理库引用：


- tensorflow
   - https://github.com/tensorflow/tensorflow
   - Copyright 2019 The TensorFlow Authors
   - Licensed under the Apache License, Version 2.0
   - Modification: no
   - Pre-built binary file is generated from this project

- libedgetpu
   - https://github.com/google-coral/libedgetpu
   - Copyright 2019 Google LLC
   - Licensed under the Apache License, Version 2.0
   - Modification: yes
   - Pre-built binary file is generated from this project

- TensorRT
   - https://github.com/nvidia/TensorRT
   - Copyright 2020 NVIDIA Corporation
   - Licensed under the Apache License, Version 2.0
   - Modification: yes
   - Some code are retrieved from this repository

- ncnn
   - https://github.com/Tencent/ncnn
   - Copyright (C) 2017 THL A29 Limited, a Tencent company.  All rights reserved.
   - Licensed under the BSD 3-Clause License
   - https://github.com/Tencent/ncnn/blob/master/LICENSE.txt
   - Modification: no
   - Pre-built binary file is generated from this project

- MNN
   - https://github.com/alibaba/MNN
   - Copyright (C) 2018 Alibaba Group Holding Limited
   - Licensed under the Apache License, Version 2.0
   - Modification: no
   - Pre-built binary file is generated from this project

- SNPE
   - https://developer.qualcomm.com/software/qualcomm-neural-processing-sdk
   - Copyright (c) 2017-2020 Qualcomm Technologies, Inc.

- Arm NN
   - https://github.com/Arm-software/armnn
   - Copyright (c) 2017 ARM Limited.

- NNabla
   - https://github.com/sony/nnabla
   - https://github.com/sony/nnabla-ext-cuda
   - Copyright 2018,2019,2020,2021 Sony Corporation.
   - Licensed under the Apache License, Version 2.0```

以上是关于深度学习多框架多平台推理引擎工具的主要内容，如果未能解决你的问题，请参考以下文章

百度飞桨重磅推出端侧推理引擎Paddle Lite 支持更多硬件平台

模型推理那些事

模型推理那些事

百度推出端侧推理引擎 Paddle Lite，支持华为 NPU 在线编译

清华深度学习框架 Jittor 开源，创新元算子和统一计算图，推理速度可提升 10%-50%

Adlik和Paddle达成合作意向，共同推进深度学习技术在通信领域落地