yolov5 模型部署NCNN（详细过程）

yolov5 模型部署NCNN

一、编译ncnn

1、在github官网输入 ncnn ,找到Tencen\ncnn 下载ncnn

2.打开vs2017 工具 x64 Native Tools Command Prompt for VS 2017

3.下载protobuf-3.4.0到本地

4.编译protobuf的动态库，输入以下命令：

cd <protobuf-root-dir>
mkdir build
cd build
cmake -G"NMake Makefiles" -DCMAKE_BUILD_TYPE=Release -DCMAKE_INSTALL_PREFIX=%cd%/install -Dprotobuf_BUILD_TESTS=OFF -Dprotobuf_MSVC_STATIC_RUNTIME=OFF ../cmake
nmake
nmake install

5.编译结果如下，成功编译。

6.编译ncnn的动态库，输入命令：

cd <ncnn-root-dir>
mkdir -p build
cd build
cmake -G"NMake Makefiles" -DCMAKE_BUILD_TYPE=Release -DCMAKE_INSTALL_PREFIX=%cd%/install -DProtobuf_INCLUDE_DIR=<protobuf-root-dir>/build/install/include -DProtobuf_LIBRARIES=<protobuf-root-dir>/build/install/lib/libprotobuf.lib -DProtobuf_PROTOC_EXECUTABLE=<protobuf-root-dir>/build/install/bin/protoc.exe -DNCNN_VULKAN=ON ..  //需要一次性全部输入
nmake
nmake install

7.编译过程出现报错，查找原因。出现有的子模块没有下载下来，对应参照CmakeLists.txt 244行发现glslang文件夹为空。

CMake Error at CMakeLists.txt:244 (message):
  The submodules were not downloaded! Please update submodules with "git
  submodule update --init" and try again.

8.检查发现是命令中-DNCNN_VULKAN=ON，由于没有安装VULKAN ，所以改成-DNCNN_VULKAN=OFF。编译成功

二、模型转换

1.yolov5模型转换onnx

python models/export.py --weights ./runs/train/exp22/weights/yolov5.pt         //使用yolov5自带的export 将pt模型转换成onnx                                                                                                                  
pip install onnx-simplifier //安装onnx-simplifier
python -m onnxsim ./yolov5.onnx ./yolo5-sim.onnx //简化转换得到的onnx

2.使用onnx2ncnn.exe 转换模型

cd tools/onnx
onnx2ncnn.exe  <your-model-path>/model-sim.onnx  <save-model-path>/model-sim.param <save-model-path>/model-sim.bin

3.出现以下错误，因为ncnn目前没有定义yolov5的切片操作。

Unsupported slice step !
Unsupported slice step !
Unsupported slice step !
Unsupported slice step !
Unsupported slice step !
Unsupported slice step !
Unsupported slice step !
Unsupported slice step !

4.打开生成的.param文件,去除不支持的网络层。

5.除去split 和crop 及concat层，一共十层，换成YoloV5Focus，总层数309变成300，如图所示。

6.将Permute的前一层Reshape 中0=xxx 改成 0 = -1。修改结果如下：

7.将修改完毕的.param 和.bin文件加载，测试完整代码如下：

#include "layer.h"
#include "net.h"

#if defined(USE_NCNN_SIMPLEOCV)
#include "simpleocv.h"
#else
#include 
#include 
#include 
#endif
#include 
#include 
#include 

class YoloV5Focus : public ncnn::Layer
{
public:
    YoloV5Focus()
    {
        one_blob_only = true;
    }

    virtual int forward(const ncnn::Mat& bottom_blob, ncnn::Mat& top_blob, const ncnn::Option& opt) const
    {
        int w = bottom_blob.w;
        int h = bottom_blob.h;
        int channels = bottom_blob.c;

        int outw = w / 2;
        int outh = h / 2;
        int outc = channels * 4;

        top_blob.create(outw, outh, outc, 4u, 1, opt.blob_allocator);
        if (top_blob.empty())
            return -100;

        #pragma omp parallel for num_threads(opt.num_threads)
        for (int p = 0; p < outc; p++)
        {
            const float* ptr = bottom_blob.channel(p % channels).row((p / channels) % 2) + ((p / channels) / 2);
            float* outptr = top_blob.channel(p);

            for (int i = 0; i < outh; i++)
            {
                for (int j = 0; j < outw; j++)
                {
                    *outptr = *ptr;

                    outptr += 1;
                    ptr += 2;
                }

                ptr += w;
            }
        }

        return 0;
    }
};

DEFINE_LAYER_CREATOR(YoloV5Focus)

struct Object
{
    cv::Rect_<float> rect;
    int label;
    float prob;
};

static inline float intersection_area(const Object& a, const Object& b)
{
    cv::Rect_<float> inter = a.rect & b.rect;
    return inter.area();
}

static void qsort_descent_inplace(std::vector<Object>& faceobjects, int left, int right)
{
    int i = left;
    int j = right;
    float p = faceobjects[(left + right) / 2].prob;

    while (i <= j)
    {
        while (faceobjects[i].prob > p)
            i++;

        while (faceobjects[j].prob < p)
            j--;

        if (i <= j)
        {
            // swap
            std::swap(faceobjects[i], faceobjects[j]);

            i++;
            j--;
        }
    }

    #pragma omp parallel sections
    {
        #pragma omp section
        {
            if (left < j) qsort_descent_inplace(faceobjects, left, j);
        }
        #pragma omp section
        {
            if (i < right) qsort_descent_inplace(faceobjects, i, right);
        }
    }
}

static void qsort_descent_inplace(std::vector<Object>& faceobjects)
{
    if (faceobjects.empty())
        return;

    qsort_descent_inplace(faceobjects, 0, faceobjects.size() - 1);
}

static void nms_sorted_bboxes(const std::vector<Object>& faceobjects, std::vector<int>& picked, float nms_threshold)
{
    picked.clear();

    const int n = faceobjects.size();

    std::vector<float> areas(n);
    for (int i = 0; i < n; i++)
    {
        areas[i] = faceobjects[i].rect.area();
    }

    for (int i = 0; i < n; i++)
    {
        const Object& a = faceobjects[i];

        int keep = 1;
        for (int j = 0; j < (int)picked.size(); j++)
        {
            const Object& b = faceobjects[picked[j]];

            // intersection over union
            float inter_area = intersection_area(a, b);
            float union_area = areas[i] + areas[picked[j]] - inter_area;
            // float IoU = inter_area / union_area
            if (inter_area / union_area > nms_threshold)
                keep = 0;
        }

        if (keep)
            picked.push_back(i);
    }
}

static inline float sigmoid(float x)
{
    return static_cast<float>(1.f / (1.f + exp(-x)));
}

static void generate_proposals(const ncnn::Mat& anchors, int stride, const ncnn::Mat& in_pad, const ncnn::Mat& feat_blob, float prob_threshold, std::vector<Object>& objects)
{
    const int num_grid = feat_blob.h;

    int num_grid_x;
    int num_grid_y;
    if (in_pad.w > in_pad.h)
    {
        num_grid_x = in_pad.w / stride;
        num_grid_y = num_grid / num_grid_x;
    }
    else
    {
        num_grid_y = in_pad.h / stride;
        num_grid_x = num_grid / num_grid_y;
    }

    const int num_class = feat_blob.w - 5;

    const int num_anchors = anchors.w / 2;

    for (int q = 0; q < num_anchors; q++)
    {
        const float anchor_w = anchors[q * 2];
        const float anchor_h = anchors[q * 2 + 1];

        const ncnn::Mat feat = feat_blob.channel(q);

        for (int i = 0; i < num_grid_y; i++)
        {
            for (int j = 0; j < num_grid_x; j++)
            {
                const float* featptr = feat.row(i * num_grid_x + j);

                // find class index with max class score
                int class_index = 0;
                float class_score = -FLT_MAX;
                for (int k = 0; k < num_class; k++)
                {
                    float score = featptr[5 + k];
                    if (score > class_score)
                    {
                        class_index = k;
                        class_score = score;
                    }
                }

                float box_score = featptr[4];

                float confidence = sigmoid(box_score) * sigmoid(class_score);

                if (confidence >= prob_threshold)
                {
                    // yolov5/models/yolo.py Detect forward
                    // y = x[i].sigmoid()
                    // y[..., 0:2] = (y[..., 0:2] * 2. - 0.5 + self.grid[i].to(x[i].device)) * self.stride[i]  # xy
                    // y[..., 2:4] = (y[..., 2:4] * 2) ** 2 * self.anchor_grid[i]  # wh

                    float dx = sigmoid(featptr[0]);
                    float dy = sigmoid(featptr[1]);
                    float dw = sigmoid(featptr[2]);
                    float dh = sigmoid(featptr[3]);

                    float pb_cx = (dx * 2.f - 0.5f + j) * stride;
                    float pb_cy = (dy * 2.f - 0.5f + i) * stride;

                    float pb_w = pow(dw * 2.f, 2) * anchor_w;
                    float pb_h = pow(dh * 2.f, 2) * anchor_h;

                    float x0 = pb_cx - pb_w * 0.5f;
                    float y0 = pb_cy - pb_h * 0.5f;
                    float x1 = pb_cx + pb_w * 0.5f;
                    float y1 = pb_cy + pb_h * 0.5f;

                    Object obj;
                    obj.rect.x = x0;
                    obj.rect.y = y0;
                    obj.rect.width = x1 - x0;
                    obj.rect.height = y1 - y0;
                    obj.label = class_index;
                    obj.prob = confidence;

                    objects.push_back(obj);
                }
            }
        }
    }
}

static int detect_yolov5(const cv::Mat& bgr, std::vector<Object>& objects)
{
    ncnn::Net yolov5;

    yolov5.opt.use_vulkan_compute = true;
    // yolov5.opt.use_bf16_storage = true;

    yolov5.register_custom_layer("YoloV5Focus", YoloV5Focus_layer_creator);

    // original pretrained model from https://github.com/ultralytics/yolov5
    // the ncnn model https://github.com/nihui/ncnn-assets/tree/master/models
    yolov5.load_param("yolov5.param");
    yolov5.load_model("yolov5.bin");

    const int target_size = 640;
    const float prob_threshold = 0.25f;
    const float nms_threshold = 0.45f;

    int img_w = bgr.cols;
    int img_h = bgr.rows;

    // letterbox pad to multiple of 32
    int w = img_w;
    int h = img_h;
    float scale = 1.f;
    if (w > h)
    {
        scale = (float)target_size / w;
        w = target_size;
        h = h * scale;
    }
    else
    {
        scale = (float)target_size / h;
        h = target_size;
        w = w * scale;
    }

    ncnn::Mat in = ncnn::Mat::from_pixels_resize(bgr.data, ncnn::Mat::PIXEL_BGR2RGB, img_w, img_h, w, h);

    // pad to target_size rectangle
    // yolov5/utils/datasets.py letterbox
    int wpad = (w + 31) / 32 * 32 - w;
    int hpad = (h + 31) / 32 * 32 - h;
    ncnn::Mat in_pad;
    ncnn::copy_make_border(in, in_pad, hpad / 2, hpad - hpad / 2, wpad / 2, wpad - wpad / 2, ncnn::BORDER_CONSTANT, 114.f);

    const float norm_vals[3] = {1 / 255.f, 1 / 255.f, 1 / 255.f};
    in_pad.substract_mean_normalize(0, norm_vals);

    ncnn::Extractor ex = yolov5.create_extractor();

    ex.input("images", in_pad);

    std::vector<Object> proposals;

    // anchor setting from yolov5/models/yolov5s.yaml

    // stride 8
    {
        ncnn::Mat out;
        ex.extract("output", out);

        ncnn::Mat anchors(6);
        anchors[0] = 10.f;
        anchors[1] = 13.f;
        anchors[2] = 16.f;
        anchors[3] = 30.f;
        anchors[4] = 33.f;
        anchors[5] = 23.f;

        std::vector<Object> objects8;
        generate_proposals(anchors, 8, in_pad, out, prob_threshold, objects8);

        proposals.insert(proposals.end(), objects8.begin(), objects8.end());
    }

    // stride 16
    {
        ncnn::Mat out;
        ex.extract("781", out);       //对应网络修改

        ncnn::Mat anchors(6);
        anchors[0] = 30.f;
        anchors[1] = 61.f;
        anchors[2] = 62.f;
        anchors[3] = 45.f;
        anchors[4] = 59.f;
        anchors[5] = 119.f;

        std::vector<Object> objects16;
        generate_proposals(anchors, 16, in_pad, out, prob_threshold, objects16);

        proposals.insert(proposals.end(), objects16.begin(), objects16.end());
    }

    // stride 32
    {
        ncnn::Mat out;
        ex.extract("801", out);      //对应网络修改

        ncnn::Mat anchors(6);
        anchors[0] = 116.f;
        anchors[1] = 90.f;
        anchors[2] = 156.f;
        anchors[3] = 198.f;
        anchors[4] = 373.f;
        anchors[5] = 326.f;

        std::vector<Object> objects32;
        generate_proposals(anchors, 32, in_pad, out, prob_threshold, objects32);

        proposals.insert(proposals.end(), objects32.begin(), objects32.end());
    }

    // sort all proposals by score from highest to lowest
    qsort_descent_inplace(proposals);

    // apply nms with nms_threshold
    std::vector<int> picked;
    nms_sorted_bboxes(proposals, picked, nms_threshold);

    int count = picked.size();

    objects.resize(count);
    for (int i = 0; i < count; i++)
    {
        objects[i] = proposals[picked[i]];

        // adjust offset to original unpadded
        float x0 = (objects[i].rect.x - (wpad / 2)) / scale;
        float y0 = (objects[i].rect.y - (hpad / 2)) / scale;
        float x1 = (objects[i].rect.x + objects[i].rect.width - (wpad / 2)) / scale;
        float y1 = (objects[i].rect.y + objects[i].rect.height - (hpad / 2)) / scale;

        // clip
        x0 = std::max(std::min(x0, (float)(img_w - 1)), 0.f);
        y0 = std::max(std::min(y0, (float)(img_h - 1)), 0.f);
        x1 = std::max(std::min(x1, (float)(img_w - 1)), 0.f);
        y1 = std::max(std::min(y1, (float)(img_h - 1)), 0.f);

        objects[i].rect.x = x0;
        objects[i].rect.y = y0;
        objects[i].rect.width = x1 - x0;
        objects[i].rect.height = y1 - y0;
    }

    return 0;
}

static void draw_objects(const cv::Mat& bgr, const std::vector<Object>& objects)
{
    static const char* class_names[] = {
        "person", "bicycle", "car", "motorcycle", "airplane", "bus", "train", "truck", "boat", "traffic light",
        "fire hydrant", "stop sign", "parking meter", "bench", "bird", "cat", "dog", "horse", "sheep", "cow",
        "elephant", "bear", "zebra", "giraffe", "backpack", "umbrella", "handbag", "tie", "suitcase", "frisbee",
        "skis", "snowboard", "sports ball", "kite", "baseball bat", "baseball glove", "skateboard", "surfboard",
        "tennis racket", "bottle", "wine glass", "cup", "fork", "knife", "spoon", "bowl", "banana", "apple",
        "sandwich", "orange", "broccoli", "carrot", "hot dog", "pizza", "donut", "cake", "chair", "couch",
        "potted plant", "bed", "dining table", "toilet", "tv", "laptop", "mouse", "remote", "keyboard", "cell phone",
        "microwave", "oven", "toaster", "sink", "refrigerator", "book", "clock", "vase", "scissors", "teddy bear",
        "hair drier", "toothbrush"
    };

    cv::Mat image = bgr.clone();

    for (size_t i = 0; i < objects.size(); i++)
    {
        const Object& obj = objects[i];

        fprintf(stderr, "%d = %.5f at %.2f %.2f %.2f x %.2f\n", obj.label, obj.prob,
                obj.rect.x, obj.rect.y, obj.rect.width, obj.rect.height);

        cv::rectangle(image, obj.rect, cv::Scalar(255, 0, 0));

        char text[256];
        sprintf(text, "%s %.1f%%", class_names[obj.label], obj.prob);

        int baseLine = 0;
        cv::Size label_size = cv::getTextSize(text, cv::FONT_HERSHEY_SIMPLEX, 0.5, 1, &baseLine);

        int x = obj.rect.x;
        int y = obj.rect.y - label_size.height - baseLine;
        if (y < 0)
            y = 0;
        if (x + label_size.width > image.cols)
            x = image.cols - label_size.width;

        cv::rectangle(image, cv::Rect(cv::Point(x, y), cv::Size(label_size.width, label_size.height + baseLine)),
                      cv::Scalar(255, 255, 255), -1);

        cv::putText(image, text, cv::Point(x, y + label_size.height),
                    cv::FONT_HERSHEY_SIMPLEX, 0.5, cv::Scalar(0, 0, 0));
    }

    cv::imshow("image", image);
    cv::waitKey(0);
}

int main(int argc, char** argv)
{
    if (argc != 2)
    {
        fprintf(stderr, "Usage: %s [imagepath]\n", argv[0]);
        return -1;
    }

    const char* imagepath = argv[1];

    cv::Mat m = cv::imread(imagepath, 1);
    if (m.empty())
    {
        fprintf(stderr, "cv::imread %s failed\n", imagepath);
        return -1;
    }

    std::vector<Object> objects;
    detect_yolov5(m, objects);

    draw_objects(m, objects);

    return 0;
}

8.ncnn的结果和pt的结果对比，结果一致

三.模型量化

1.将param中YoloV5Focus 层改成Exp（不参与计算的层）

2.回到tools 路径，运行命令

ncnnoptimize.exe yolov5-sim.param yolov5-sim.bin yolov5-sim-opt.param yolov5-sim-opt.bin 65536

//出现下列提示
Input layer images without shape info, shape_inference skipped
Input layer images without shape info, estimate_memory_footprint skipped

3.将生成的yolov5-sin-opt.param文件打开，将之前修改的Exp名称改成YoloV5Focus

4.运行模型测试对比，结果无影响。

参考YOLOv5转NCNN过程