yolov8seg的640x640输入的推理能正常跑了，但是onnx在服务器4个目标结果，在mnn只出来2个。什么原因？

[jamesBaiyuhao]

其实input就是 640x640的bgr的输入。 我的模型的input是 1，3，640，640， data应该是int的。 output是1，42，8400，应该是float。 您看是我input不对吗？ pretreat->convert((uint8_t*)inputImage, width, height, 0, input); 还是output不对？ 我直接output_t->host() 直接处理输出的tensor？

int main(int argc, const char* argv[]) { if (argc < 4) { MNN_PRINT("Usage: ./segment.out model.mnn input.jpg output.jpg\n"); return 0; } std::shared_ptr net; net.reset(Interpreter::createFromFile(argv[1])); if (net == nullptr) { MNN_ERROR("Invalid Model\n"); return 0; } ScheduleConfig config; auto session = net->createSession(config); auto input = net->getSessionInput(session, "images"); MNN_PRINT("origin size: line %d\n", LINE); // input->print(); MNN_PRINT("origin size: line %d\n", LINE); input->printShape(); MNN_PRINT("origin size: line %d\n", LINE);

auto shape = input->shape();
{
    int input_h = 160;
    int input_w = 160;
    if(shape[2] > 0 && shape[2] > 0)
    {

    }
    else
    {
        shape[2] = input_h;
        shape[3] = input_w;
    }
}
net->resizeTensor(input, shape);
net->resizeSession(session);
input = net->getSessionInput(session, "images");
    MNN_PRINT("origin size: line %d\n", __LINE__);
// input->print();
    MNN_PRINT("origin size: line %d\n", __LINE__);
input->printShape();
    MNN_PRINT("origin size: line %d\n", __LINE__);

{
    auto ow              = input->width();
    auto oh              = input->height();
    auto ochannel             = input->channel();
    auto dimensionFormat = input->getDimensionType();
    MNN_PRINT("input---: ow:%d , oh:%d, ochannel: %d, dimensionFormat %d\n", 
        ow, oh, ochannel, dimensionFormat);

    int size_w   = 0;
    int size_h   = 0;
    int bpp      = 3;

    if (bpp == 0)
        bpp = 1;
    size_w = shape[2] ;
    size_h = shape[3] ;
    MNN_PRINT("input---: w:%d , h:%d, bpp: %d\n", size_w, size_h, bpp);

    auto inputPatch = argv[2];
    int width, height, channel;
    auto inputImage = stbi_load(inputPatch, &width, &height, &channel, 4);
    if (nullptr == inputImage) {
        MNN_ERROR("Can't open %s\n", inputPatch);
        return 0;
    }
    MNN_PRINT("origin size: %d, %d\n", width, height);
    Matrix trans;
    // Set scale, from dst scale to src
    trans.setScale((float)(width-1) / (size_w-1), (float)(height-1) / (size_h-1));
    ImageProcess::Config config;
    config.filterType = CV::BILINEAR;
    //        float mean[3]     = {103.94f, 116.78f, 123.68f};
    //        float normals[3] = {0.017f, 0.017f, 0.017f};
    // float mean[3]     = {127.5f, 127.5f, 127.5f};
    // float normals[3] = {0.00785f, 0.00785f, 0.00785f};
    const float mean[3] = {127.5f, 127.5f, 127.5f};
    const float normals[3] = {2.0f / 255.0f, 2.0f / 255.0f, 2.0f / 255.0f};
    ::memcpy(config.mean, mean, sizeof(mean));
    ::memcpy(config.normal, normals, sizeof(normals));
    config.sourceFormat = RGBA;
    config.destFormat   = BGR;

    MNN_PRINT("origin size: line %d %d, %d\n", __LINE__,width, height);
    std::shared_ptr<ImageProcess> pretreat(ImageProcess::create(config));
    pretreat->setMatrix(trans);
    pretreat->convert((uint8_t*)inputImage, width, height, 0, input);
    stbi_image_free(inputImage);
    MNN_PRINT("origin size: line %d %d, %d\n", __LINE__,width, height);
}
// Run model
    MNN_PRINT("origin size: line %d\n", __LINE__);
net->runSession(session);
    MNN_PRINT("origin size: line %d\n", __LINE__);

int numClasses = 6;//4;
float  confThreshold = 0.35;
float nmsThreshold = 0.5;

// get output
auto output_t = net->getSessionOutput(session, "output0");
    MNN_PRINT("origin size: line %d\n", __LINE__);
printf_value(output_t->host<float>(),"output_ptr",10);
    MNN_PRINT("origin size: line %d\n", __LINE__);
output_t->printShape();
    MNN_PRINT("origin size: line %d\n", __LINE__);

{
    struct timeval tstartNms, tendNms;
    gettimeofday(&tstartNms, NULL);

    int channel  = output_t->channel();
    int height   = output_t->height();
    int width    = output_t->width();
    printf("[%s:%d] channel %d height %d width %d\n", __FUNCTION__, __LINE__, channel, height, width);

    width = height;//8400
    height = channel;//42

    auto scoresPtr = output_t->host<float>();
    float *scoreHead = NULL;
    int iLayer = 0;
    std::vector<BoxInfo_t> outBoxes;

    float cx, cy, w, h;
    float *cxHead = (float *)scoresPtr;
    float *cyHead = (float *)scoresPtr + 1 * width;
    float *wHead = (float *)scoresPtr + 2 * width;
    float *hHead = (float *)scoresPtr + 3 * width;

    printf("[%s:%d] width %d\n", __FUNCTION__, __LINE__, width);
    for (int iBox = 0; iBox < width; iBox++)
    {
        scoreHead = (float *)scoresPtr + 4 * width + iBox;

        float maxScore = -INFINITY;
        int maxInd = -1;

        for (int iClass = 0; iClass < numClasses; iClass++)
        {
            //printf("[%s:%d] iBox %d iClass %d maxScore %f *scoreHead %f\n", __FUNCTION__, __LINE__, iBox, iClass, maxScore, *scoreHead);
            if (*scoreHead >= confThreshold && *scoreHead > maxScore) {
                maxScore = *scoreHead;
                maxInd = iClass;
                printf("[%s:%d] iBox %d iClass %d maxScore %f maxInd %d\n", __FUNCTION__, __LINE__, iBox, iClass, maxScore, maxInd);
            }
            scoreHead += width;
        }

        if (maxInd >= 0)
        {
            cx = cxHead[iBox];
            cy = cyHead[iBox];
            w = wHead[iBox];
            h = hHead[iBox];
            BoxInfo_t rBoxInfo;
            rBoxInfo.x1 = cx - w / 2.0;
            rBoxInfo.y1 = cy - h / 2.0;
            rBoxInfo.x2 = cx + w / 2.0;
            rBoxInfo.y2 = cy + h / 2.0;
            rBoxInfo.score = maxScore;
            rBoxInfo.label = maxInd;

            // printf(" iBox %d maxInd %d: x1 %f %f %f %f w: %f %f label %d score %f\n", iBox, maxInd,
            //     rBoxInfo.x1 , rBoxInfo.y1, rBoxInfo.x2 , rBoxInfo.y2,
            //     (rBoxInfo.x2 - rBoxInfo.x1) , (rBoxInfo.y2 - rBoxInfo.y1), rBoxInfo.label , rBoxInfo.score );

            outBoxes.push_back(rBoxInfo);
        }
    }

    gettimeofday(&tendNms, NULL);
    printf("loop time: %f ms\n",
           (tendNms.tv_sec - tstartNms.tv_sec) * 1000.0 + (tendNms.tv_usec - tstartNms.tv_usec) / 1000.0);
   CheckBoxes(outBoxes, shape[2], shape[3]);
    gettimeofday(&tendNms, NULL);
    printf("CheckBoxes time: %f ms\n",
           (tendNms.tv_sec - tstartNms.tv_sec) * 1000.0 + (tendNms.tv_usec - tstartNms.tv_usec) / 1000.0);
    Nms(outBoxes, nmsThreshold);
    gettimeofday(&tendNms, NULL);
    printf("Nms time: %f ms\n",
           (tendNms.tv_sec - tstartNms.tv_sec) * 1000.0 + (tendNms.tv_usec - tstartNms.tv_usec) / 1000.0);

    for (int i = 0; i < int(outBoxes.size()); ++i)
    {
        printf(" %d: x1 %f %f %f %f w: %f %f label %d score %f\n", i, 
              outBoxes[i].x1 , outBoxes[i].y1, outBoxes[i].x2 , outBoxes[i].y2,
               (outBoxes[i].x2 - outBoxes[i].x1) , (outBoxes[i].y2 - outBoxes[i].y1), outBoxes[i].label , outBoxes[i].score );
    }
}

return 0;

}

平台(如果交叉编译请再附上交叉编译目标平台):

Platform(Include target platform as well if cross-compiling):

macOS

Github版本:

Github Version:

20240912最新版本

直接下载ZIP包请提供下载日期以及压缩包注释里的git版本(可通过7z l zip包路径命令并在输出信息中搜索Comment 获得，形如Comment = bc80b11110cd440aacdabbf59658d630527a7f2b)。 git clone请提供 git commit 第一行的commit id

Provide date (or better yet, git revision from the comment section of the zip. Obtainable using 7z l PATH/TO/ZIP and search for Comment in the output) if downloading source as zip,otherwise provide the first commit id from the output of git commit

编译方式:

Compiling Method

cmake -DMNN_BUILD_DEMO=ON -DMNN_BUILD_CONVERTER=ON -DMNN_BUILD_TOOL=ON -DMNN_BUILD_BENCHMARK=ON -DMNN_BUILD_QUANTOOLS=ON -DMNN_BUILD_OPENCV=ON -DMNN_IMGCODECS=ON ..

请在这里粘贴cmake参数或使用的cmake脚本路径以及完整输出
Paste cmake arguments or path of the build script used here as well as the full log of the cmake proess here or pastebin

编译日志:

Build Log:

默认 2%] Built target MNNCore [ 2%] Built target MNNCV [ 2%] Built target MNNMath [ 14%] Built target MNNTransform [ 14%] Built target MNNUtils [ 25%] Built target MNNCPU [ 25%] Built target MNNX8664 [ 26%] Built target MNNAVX [ 27%] Built target MNNAVXFMA [ 28%] Built target MNNSSE [ 28%] Built target MNN [ 30%] Built target MNN_Express [ 34%] Built target MNNTrain [ 36%] Built target MNNTrainUtils [ 37%] Built target pictureRecognition.out [ 37%] Built target pictureRecognition_module.out [ 37%] Built target multithread_imgrecog.out [ 37%] Built target pictureRecognition_batch.out [ 37%] Built target pictureRotate.out [ 37%] Built target multiPose.out [ 37%] Built target segment.out [ 37%] Building CXX object CMakeFiles/yolov8s-seg.out.dir/demo/exec/yolov8s-seg.cpp.o [ 38%] Linking CXX executable yolov8s-seg.out [ 38%] Built target yolov8s-seg.out [ 38%] Built target expressDemo.out [ 38%] Built target expressMakeModel.out [ 38%] Built target transformerDemo.out [ 38%] Built target nluDemo.out [ 38%] Built target GetMNNInfo [ 38%] Built target ModuleBasic.out [ 38%] Built target SequenceModuleTest.out [ 38%] Built target mergeInplaceForCPU [ 39%] Built target modelCompare.out [ 40%] Built target MNNV2Basic.out [ 40%] Built target mobilenetTest.out [ 40%] Built target backendTest.out [ 41%] Built target testModel.out [ 41%] Built target testModel_expr.out [ 41%] Built target testModelWithDescribe.out [ 42%] Built target getPerformance.out [ 43%] Built target checkInvalidValue.out [ 44%] Built target timeProfile.out [ 44%] Built target testTrain.out [ 44%] Built target fuseTest [ 44%] Built target LoRA [ 44%] Built target checkDir.out [ 44%] Built target checkFile.out [ 44%] Built target winogradExample.out [ 44%] Built target benchmark.out [ 45%] Built target benchmarkExprModels.out [ 46%] Built target quantized.out [ 49%] Built target libprotobuf-lite [ 58%] Built target libprotobuf [ 58%] Built target transformer [ 58%] Built target extractForInfer [ 59%] Built target runTrainDemo.out [ 59%] Built target MNNCompress [ 67%] Built target MNNConverterTF [ 72%] Built target MNNConverterONNX [ 72%] Built target OnnxClip [ 76%] Built target MNNConverterCaffe [ 76%] Built target MNNConverterMNN [ 92%] Built target MNNConverterOpt [ 96%] Built target MNNConverterTFL [ 97%] Built target MNNConvertDeps [ 98%] Built target MNNConvert [ 98%] Built target MNNRevert2Buffer [ 98%] Built target MNNDump2Json [ 99%] Built target TestConvertResult [ 99%] Built target TestPassManager [100%] Built target MNNOpenCV

粘贴在这里
Paste log here or pastebin

[jamesBaiyuhao]

转模型是： ./MNNConvert -f ONNX --modelFile yolov8s-seg-bedNurse913_1.onnx --MNNModel yyolov8s-seg-bedNurse913_1.mnn --bizCode MNN --keepInputFormat=0

[jxt1234]

1. 换用 Module API ，在 load module 时传入 onnx 里面的4个输出名字 https://mnn-docs.readthedocs.io/en/latest/inference/module.html
2. 创建 Session 传入 saveTensors 或者设定 path ，指定输出 https://mnn-docs.readthedocs.io/en/latest/inference/session.html

建议换用 Module API