前言
在上一次的测试中,我们按照官方给的流程,使用EasyDL快速实现了一个具有性别检测功能的人脸识别系统,那么今天,我们将要试一下通过Paddlepaddle从零开始,训练一个自己的多分类模型,并进行嵌入式部署。 整个训练过程和模型在:https://aistudio.baidu.com/aistudio/projectDetail/61103 下面详细介绍模型训练的过程.
数据集准备
我们使用CIFAR10数据集。CIFAR10数据集包含60,000张32x32的彩色图片,10个类别,每个类包含6,000张。其中50,000张图片作为训练集,10000张作为验证集。
!mkdir -p /home/aistudio/.cache/paddle/dataset/cifar # wget将下载的文件存放到指定的文件夹下,同时重命名下载的文件,利用-O !wget "http://ai-atest.bj.bcebos.com/cifar-10-python.tar.gz" -O cifar-10-python.tar.gz !mv cifar-10-python.tar.gz /home/aistudio/.cache/paddle/dataset/cifar/
模型结构
我们选择了以三个卷积层串联一个全连接层的输出,作为猫狗分类的预测,采用固定维度输入,输出为分类数
def convolutional_neural_network(img): # 第一个卷积-池化层 conv_pool_1 = fluid.nets.simple_img_conv_pool( input=img, # 输入图像 filter_size=5, # 滤波器的大小 num_filters=20, # filter 的数量。它与输出的通道相同 pool_size=2, # 池化层大小2*2 pool_stride=2, # 池化层步长 act="relu") # 激活类型 # 第二个卷积-池化层 conv_pool_2 = fluid.nets.simple_img_conv_pool( input=conv_pool_1, filter_size=5, num_filters=50, pool_size=2, pool_stride=2, act="relu") # 第三个卷积-池化层 conv_pool_3 = fluid.nets.simple_img_conv_pool( input=conv_pool_2, filter_size=5, num_filters=50, pool_size=2, pool_stride=2, act="relu") # 以softmax为激活函数的全连接输出层,10类数据输出10个数字 prediction = fluid.layers.fc(input=conv_pool_3, size=10, act='softmax') return prediction
训练&验证
接下来在Paddlepaddle fluid上,进行训练。整个训练代码见附件train.py 模型验证,采用附件predict.py的代码进行验证与运行时间的测量,选取一张狗的图:dog.jpg (可以fork首页链接aistudio平台上的demo) 连续预测10000次,输出如下:
CPU 运行结果为:预处理时间为0.0006270000000085929,预测时间为:16.246494 Out: im_shape的维度: (1, 3, 32, 32) The run time of image process is 0.0006270000000085929 The run time of predict is 16.246494 results [array([[5.0159363e-04, 3.5942634e-05, 2.5955746e-02, 4.7745958e-02, 9.9251214e-03, 9.0146154e-01, 1.9564393e-03, 1.2230080e-02, 4.7619540e-08, 1.8753216e-04]], dtype=float32)] infer results: dog
GPU V100 运行结果为:预处理时间为0.0006390000000067175,预测时间为:15.903074000000018 Out: im_shape的维度: (1, 3, 32, 32) The run time of image process is 0.0006390000000067175 The run time of predict is 15.903074000000018 results [array([[5.0159392e-04, 3.5942641e-05, 2.5955772e-02, 4.7746032e-02, 9.9251205e-03, 9.0146142e-01, 1.9564414e-03, 1.2230078e-02, 4.7619821e-08, 1.8753250e-04]], dtype=float32)] infer results: dog
可以看到,模型可以正确的识别出图片中的动物为狗,接下来,我们就要尝试将这个模型部署到Edgeboard上面。
模型导出
我们需要将模型保存为模型文件model以及权重文件params,可以采用如下Paddle的API进行保存
fluid.io.save_inference_model(model_save_dir,['images'],[predict], exe,params_filename="mlp" + '-params',model_filename="mlp" + '-model',)
如图所示,在AiStudio的左侧打开模型文件所在的文件夹,下载mlp-model、mlp-params两个文件。
在Edgeboard上部署模型,完成预测
1、新建工程文件夹,目录结构如下(可以仿照sample里的resnet、inception例程):
-sample_image_catdog -build -image -include -paddlepaddle-mobile -... -lib -libpaddle-mobile.so -model -mlp -model -params -src -fpga_cv.cpp -main.cpp
2、将AiStudio上导出来的模型放置在model里的mlp文件夹,修改名字为model、params
3、新建 CMakeLists.txt
cmake_minimum_required(VERSION 3.5.1) project(paddle_edgeboard) set(CMAKE_CXX_STANDARD 14) set(CMAKE_EXE_LINKER_FLAGS "${CMAKE_EXE_LINKER_FLAGS} -pthread") add_definitions(-DPADDLE_MOBILE_FPGA_V1) add_definitions(-DPADDLE_MOBILE_FPGA) set(PADDLE_LIB_DIR "${PROJECT_SOURCE_DIR}/lib" ) set(EASYDL_INCLUDE_DIR "${PROJECT_SOURCE_DIR}/include" ) set(PADDLE_INCLUDE_DIR "${PROJECT_SOURCE_DIR}/include/paddle-mobile" ) set(APP_NAME "paddle_edgeboard" ) aux_source_directory(${CMAKE_CURRENT_SOURCE_DIR}/src SRC) find_package(OpenCV QUIET COMPONENTS core videoio highgui imgproc imgcodecs ml video) include_directories(SYSTEM ${OpenCV_INCLUDE_DIRS}) #list(APPEND Caffe_LINKER_LIBS ${OpenCV_LIBS}) message(STATUS "OpenCV found (${OpenCV_CONFIG_PATH}),${OpenCV_LIBS}") #add_definitions(-DUSE_OPENCV) include_directories(${EASYDL_INCLUDE_DIR}) include_directories(${PADDLE_INCLUDE_DIR}) LINK_DIRECTORIES(${PADDLE_LIB_DIR}) add_executable(${APP_NAME} ${SRC}) target_link_libraries(${APP_NAME} paddle-mobile) target_link_libraries(${APP_NAME} ${OpenCV_LIBS} )
4、main.cpp
#include #include "io/paddle_inference_api.h" #include "math.h" #include #include #include #include #include #include #include #include #include #include #include "fpga/KD/float16.hpp" #include "fpga/KD/llapi/zynqmp_api.h" using namespace paddle_mobile; #include #include using namespace cv; cv::Mat sample_float; static std::vector label_list(10); void readImage(std::string filename, float* buffer) { Mat img = imread(filename); if (img.empty()) { std::cerr << "Can't read image from the file: " << filename << std::endl; exit(-1); } Mat img2; resize(img, img2, Size(32,32)); img2.convertTo(sample_float, CV_32FC3); int index = 0; for (int row = 0; row < sample_float.rows; ++row) { float* ptr = (float*)sample_float.ptr(row); for (int col = 0; col < sample_float.cols; col++) { float* uc_pixel = ptr; // uc_pixel[0] -= 102; // uc_pixel[1] -= 117; // uc_pixel[1] -= 124; float r = uc_pixel[0]; float g = uc_pixel[1]; float b = uc_pixel[2]; buffer[index] = b / 255.0; buffer[index + 1] = g / 255.0; buffer[index + 2] = r / 255.0; // sum += a + b + c; ptr += 3; // DLOG << "r:" << r << " g:" << g << " b:" << b; index += 3; } } // return sample_float; } PaddleMobileConfig GetConfig() { PaddleMobileConfig config; config.precision = PaddleMobileConfig::FP32; config.device = PaddleMobileConfig::kFPGA; // config.model_dir = "../models/mobilenet/"; config.prog_file = "../model/mlp/model"; config.param_file = "../model/mlp/params"; config.thread_num = 4; return config; } int main() { clock_t startTime,endTime; zynqmp::open_device(); std::cout << " open_device success " << std::endl; PaddleMobileConfig config = GetConfig(); std::cout << " GetConfig success " << std::endl; auto predictor = CreatePaddlePredictor(config); std::cout << " predictor success " << std::endl; startTime = clock();//计时开始 float data[1 * 3 * 32 * 32] = {1.0f}; readImage("../image/cat.jpg", data); endTime = clock();//计时结束 std::cout << "The run time of image process is: " <<(double)(endTime - startTime) / CLOCKS_PER_SEC << "s" << std::endl; PaddleTensor tensor; tensor.shape = std::vector({1, 3, 32, 32}); tensor.data = PaddleBuf(data, sizeof(data)); tensor.dtype = PaddleDType::FLOAT32; std::vector paddle_tensor_feeds(1, tensor); PaddleTensor tensor_out; tensor_out.shape = std::vector({}); tensor_out.data = PaddleBuf(); tensor_out.dtype = PaddleDType::FLOAT32; std::vector outputs(1, tensor_out); std::cout << " before predict " << std::endl; predictor->Run(paddle_tensor_feeds, &outputs); std::cout << " after predict " << std::endl; // assert(); endTime = clock();//计时结束 std::cout << "The run time of predict is: " <<(double)(endTime - startTime) / CLOCKS_PER_SEC << "s" << std::endl; float* data_o = static_cast(outputs[0].data.data()); for (size_t j = 0; j < outputs[0].data.length() / sizeof(float); ++j) { std::cout << "output[" << j << "]: " << data_o[j] << std::endl; } int index = 0; float max = 0.0; for (int i = 0;i < 10; i++) { float val = data_o[i]; if (val > max) { max = val > max ? val : max; index = i; } } label_list = {"airplane", "automobile", "bird", "cat", "deer", "dog", "frog", "horse", "ship", "truck" }; std::cout << "Result" << " is " << label_list[index] << std::endl; return 0; }
5、编译运行
insmod /home/root/workspace/driver/fpgadrv.ko cd /home/root/workspace/sample/sample_image_catdog mkdir build cd build rm -rf * cmake .. make ./paddle_edgeboard
修改main文件要预测的图像:
6、修改main文件后重复执行预测,可得结果如下:图像处理时间大概为:0.006秒,预测时间大概为:0.008秒
总结
优点:
1、EdgeBoard内置的Paddle-Mobile,可以与Paddle训练出来的模型进行较好的对接。
2、预测速度上: Edge在预测小模型的时候,能与双核CPU和GPU在一个数量级,估计是模型较小,batch size也为1,gpu,cpu的性能优势抵不过通信的开销,后续将进行大模型、高batch size的测试。
3、提供的demo也足够简单,修改起来难度很低。
不足:
Paddle-Mobile相关文档具有一定门槛,且较为分散。初次使用的时候会走一些弯路出现问题的时候往往是个黑盒,不易于定位。在这次进行模型训练的尝试中,出现过一次op不支持的情况,我们在官网上甚至没有找到支持的op列表,这个在开发哥们的支持下升级版本后解决。如果后续能在稳定的固件版本下使用,并有比较易用的sdk,开发门槛可能会进一步降低。
作者:Litchll