在專欄之前的文章，我們介紹過ArmNN，詳情可參考被低估的ArmNN（一）如何編譯。這里，我們給大家介紹如何使用ArmNN在Android設(shè)備上進(jìn)行部署，部署的任務(wù)以Mobilenet分類器為例。關(guān)于Mobilenet回歸器的訓(xùn)練，大家可以參考如何DIY輕型的Mobilenet回歸器。我們今天的部署平臺仍然是基于RK3399嵌入式平臺，系統(tǒng)為Android-8.1。
作者：張新棟

我們知道ArmNN是一個(gè)非常高效的Inference框架，300x300的Mobilenet-SSD在depth_multiplier取1.0時(shí)inference最快可達(dá)90ms/幀。今天我們將使用ArmNN框架，用C++在RK-3399-Android-8.1中進(jìn)行Mobilenet回歸任務(wù)的部署。首先我們先進(jìn)行第一步，環(huán)境的配置。

環(huán)境配置

若想使用編譯好的ArmNN進(jìn)行inference，首先我們必須要先加載編譯好的ArmNN庫、頭文件及其他依賴文件。這里我們依舊為大家提供了Android.mk及Application.mk，

LOCAL_PATH := $(call my-dir)

include $(CLEAR_VARS)
LOCAL_MODULE := armnn
LOCAL_SRC_FILES := $(LOCAL_PATH)/../libarmnn.so
LOCAL_EXPORT_C_INCLUDES := $(LOCAL_PATH)/../../include/armnn
LOCAL_SHARED_LIBRARIES := c++_shared
include $(PREBUILT_SHARED_LIBRARY)


include $(CLEAR_VARS)
LOCAL_MODULE := tfliteParser
LOCAL_SRC_FILES := $(LOCAL_PATH)/../libarmnnTfLiteParser.so
LOCAL_EXPORT_C_INCLUDES := $(LOCAL_PATH)/../../include/libarmnnTfLiteParser
LOCAL_SHARED_LIBRARIES := c++_shared
include $(PREBUILT_SHARED_LIBRARY)

include $(CLEAR_VARS)
LOCAL_MODULE := armnnSerializer
LOCAL_SRC_FILES := $(LOCAL_PATH)/../libarmnnSerializer.so
LOCAL_EXPORT_C_INCLUDES := $(LOCAL_PATH)/../../include/armnn/armnnSerializer
LOCAL_SHARED_LIBRARIES := c++_shared
include $(PREBUILT_SHARED_LIBRARY)


include $(CLEAR_VARS)
OpenCV_INSTALL_MODULES := on
OPENCV_LIB_TYPE := STATIC
include /Users/xindongzhang/armnn-tflite/OpenCV-android-sdk/sdk/native/jni/OpenCV.mk
LOCAL_MODULE := face_detector

LOCAL_C_INCLUDES += $(OPENCV_INCLUDE_DIR)
LOCAL_C_INCLUDES += $(LOCAL_PATH)/../../include
LOCAL_C_INCLUDES += $(LOCAL_PATH)/../../../boost_1_64_0/
LOCAL_C_INCLUDES += $(LOCAL_PATH)/../../third-party/stb/
LOCAL_SRC_FILES := /
                face_detector.cpp


LOCAL_LDLIBS := -landroid -llog -ldl -lz 
LOCAL_CFLAGS   := -O2 -fvisibility=hidden -fomit-frame-pointer -fstrict-aliasing   /
                  -ffunction-sections -fdata-sections -ffast-math -ftree-vectorize / 
                  -fPIC -Ofast -ffast-math -w -std=c++14
LOCAL_CPPFLAGS := -O2 -fvisibility=hidden -fvisibility-inlines-hidden -fomit-frame-pointer /
                  -fstrict-aliasing -ffunction-sections -fdata-sections -ffast-math -fPIC  /
                  -Ofast -ffast-math -std=c++14
LOCAL_LDFLAGS  += -Wl,--gc-sections
LOCAL_CFLAGS   += -fopenmp
LOCAL_CPPFLAGS += -fopenmp
LOCAL_LDFLAGS  += -fopenmp
LOCAL_ARM_NEON := true

APP_ALLOW_MISSING_DEPS = true

LOCAL_SHARED_LIBRARIES :=                                   /
                        armnn                               /
            tfliteParser                        /
            armnnSerializer                     /
            android.hardware.neuralnetworks@1.0 /
            android.hidl.allocator@1.0          /
            android.hidl.memory@1.0             /
            libc++_shared

include $(BUILD_EXECUTABLE)

如下為Application.mk文件，

ANDROID_TOOLCHAIN=clang?
APP_ABI := arm64-v8a
APP_CPPFLAGS := -frtti -fexceptions -std=c++14
APP_PLATFORM := android-27
APP_STL := c++_shared

這里需要注意的是Application.mk的APP_STL項(xiàng)，由于我們在編譯ArmNN時(shí)使用的STL為c++_shared，所以這里需要使用c++_shared，另外Android.mk文件中鏈接的OpenCV庫也需要使用c++_shared的stl進(jìn)行編譯（官網(wǎng)下載的即c++_shared編譯）。

編寫C++業(yè)務(wù)代碼

在配置好依賴項(xiàng)后，我們開始使用ArmNN提供的C++API進(jìn)行業(yè)務(wù)代碼的書寫。首先第一步我們需要加載模型，ArmNN提供了解析題 ITfLiteParserPtr，我們可以使用其進(jìn)行模型的加載。另外加載好的模型我們需要使用一個(gè)網(wǎng)絡(luò)結(jié)構(gòu)進(jìn)行存儲，ArmNN提供了INetworkPtr。為了在對應(yīng)的arm嵌入式平臺中高效的執(zhí)行，ArmNN還提供了IOptimizedNetworkPtr來對網(wǎng)絡(luò)的inference進(jìn)行優(yōu)化。更多的細(xì)節(jié)大家可參考如下的業(yè)務(wù)代碼。

armnnTfLiteParser::ITfLiteParserPtr parser = armnnTfLiteParser::ITfLiteParser::Create(); 
armnn::INetworkPtr pose_reg_network{nullptr, [](armnn::INetwork *){}};
armnn::IOptimizedNetworkPtr pose_reg_optNet{nullptr, [](armnn::IOptimizedNetwork *){}};
armnn::InputTensors pose_reg_in_tensors;
armnn::OutputTensors pose_reg_ou_tensors;
armnn::IRuntimePtr runtime{nullptr, [](armnn::IRuntime *){}};
float yaw[1];
float pose_reg_input[64*64*3];


// loading tflite model
std::string pose_reg_modelPath = "/sdcard/Algo/pose.tflite";
pose_reg_network = parser->CreateNetworkFromBinaryFile(pose_reg_modelPath.c_str());

// binding input and output
armnnTfLiteParser::BindingPointInfo pose_reg_input_bind  = 
                              parser->GetNetworkInputBindingInfo(0, "input/ImageInput");
armnnTfLiteParser::BindingPointInfo pose_reg_output_bind = 
                              parser->GetNetworkOutputBindingInfo(0, "yaw/yangle");

// wrapping pose reg input and output
armnn::Tensor pose_reg_input_tensor(pose_reg_input_bind.second, pose_reg_input);
pose_reg_in_tensors.push_back(std::make_pair(pose_reg_input_bind.first, pose_reg_input_tensor));

armnn::Tensor pose_reg_output_tensor(pose_reg_output_bind.second, yaw);
pose_reg_ou_tensors.push_back(std::make_pair(pose_reg_output_bind.first, pose_reg_output_tensor));

// config runtime, fp16 accuracy 
armnn::IRuntime::CreationOptions runtimeOptions;
runtime = armnn::IRuntime::Create(runtimeOptions);
armnn::OptimizerOptions OptimizerOptions;
OptimizerOptions.m_ReduceFp32ToFp16 = true;
this->pose_reg_optNet = 
armnn::Optimize(*pose_reg_network, {armnn::Compute::GpuAcc},runtime->GetDeviceSpec(), OptimizerOptions);
runtime->LoadNetwork(this->pose_reg_identifier, std::move(this->pose_reg_optNet));

// load image
cv::Mat rgb_image = cv::imread("face.jpg", 1);
cv::resize(rgb_image, rgb_image, cv::Size(pose_reg_input_size, pose_reg_input_size));
rgb_image.convertTo(rgb_image, CV_32FC3);
rgb_image = (rgb_image - 127.5f) * 0.017f;

// preprocess image
int TOTAL   = 64 * 64 * 3;
float* data = (float*) rgb_image.data;
for (int i = 0; i < TOTAL; ++i) {
    pose_reg_input[i] = data[i];
}

// invoke graph forward inference
armnn::Status ret = runtime->EnqueueWorkload(
    this->pose_reg_identifier,
    this->pose_reg_in_tensors,
    this->pose_reg_ou_tensors
);
float result = yaw[0] * 180 / 3.14; 

非常簡單易懂的業(yè)務(wù)代碼就可以完成ArmNN的一次inference，注意這里我們使用的是FP16來進(jìn)行inference，相比于FP32，F(xiàn)P16具有更高的加速比，且不會損失很多精度。后續(xù)我們會給出如何使用ArmNN來做INT8的inference例子。

最后

本文我們介紹了如何使用ArmNN來進(jìn)行Mobilenet的inference（其實(shí)很容易就可以改成分類任務(wù)），并使用FP16的精度進(jìn)行inference，該網(wǎng)絡(luò)在RK3399中執(zhí)行效率非常高（約10ms）。若你想在其他設(shè)備中使用FP16，首先你要保證設(shè)備中有GPU，且支持OpenCL。歡迎大家留言討論、關(guān)注專欄，謝謝大家！

審核編輯 黃昊宇