useing mnncv image process get wrong result

[jfkkf123]

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):

• x84_64 ubuntu20.04
• raspberry 3b+

mnn版本:

mnn Version:

2.9.5

I use pytorch->onnx->mnn to get .mnn model. I use two different ways to preprocess image,but get different result:

• same init

    ultraface_interpreter_ = std::shared_ptr<MNN::Interpreter>(MNN::Interpreter::createFromFile(mnn_path.c_str()));
    MNN::ScheduleConfig config;
    config.numThread = 4;
    MNN::BackendConfig backendConfig;
    backendConfig.precision = (MNN::BackendConfig::PrecisionMode) 1;
    config.backendConfig = &backendConfig;
    ultraface_session_ = ultraface_interpreter_->createSession(config);
    input_tensor_ = ultraface_interpreter_->getSessionInput(ultraface_session_, nullptr);

• opencv pre-process

   cv::Mat raw_image=cv::imread("./image.jpg");
    cv::Mat image;
    cv::resize(raw_image, image, cv::Size(320, 240));

    cv::Mat rgb_img;
    cv::cvtColor(image, rgb_img, cv::COLOR_BGR2RGB);
    rgb_img.convertTo(rgb_img,CV_32FC3,1.0);
    rgb_img=(rgb_img-127.0)/128.0;
    cv::Mat nchwMat;
    hwc_to_chw(rgb_img, nchwMat);

    MNN::Tensor *inhostTensor=MNN::Tensor::create<float>(std::vector<int>{1,3,240,320},nullptr,MNN::Tensor::CAFFE);
    std::memcpy(inhostTensor->host<float>(),nchwMat.data,nchwMat.total()*sizeof(float));
    input_tensor_->copyFromHostTensor(inhostTensor);

    ultraface_interpreter_->resizeTensor(input_tensor_, {1, 3, 240, 320});
    ultraface_interpreter_->resizeSession(ultraface_session_);
   // run net
    ultraface_interpreter_->runSession(ultraface_session_);

• mnn cv pre-process

   // opencv function
   cv::Mat raw_image=cv::imread("./image.jpg");
    cv::Mat image;
    cv::resize(raw_image, image, cv::Size(320, 240));

   //mnn cv function
   const float mean_vals[3] = {127, 127, 127};
    const float norm_vals[3] = {1.0 / 128, 1.0 / 128, 1.0 / 128};
    std::shared_ptr<MNN::CV::ImageProcess> pretreat(
             MNN::CV::ImageProcess::create(MNN::CV::BGR, MNN::CV::RGB, mean_vals, 3,norm_vals, 3));
     pretreat->convert(image.data, 320, 240, image.step[0], input_tensor_);  // or pretreat->convert(image.data, 320, 240, 0, input_tensor_);
    ultraface_interpreter_->runSession(ultraface_session_);

• get output data

   string scores = "scores";
    string boxes = "boxes";
    MNN::Tensor *tensor_scores = ultraface_interpreter_->getSessionOutput(ultraface_session_, scores.c_str());
    MNN::Tensor *tensor_boxes = ultraface_interpreter_->getSessionOutput(ultraface_session_, boxes.c_str());

     MNN::Tensor tensor_scores_host(tensor_scores, tensor_scores->getDimensionType());
     tensor_scores->copyToHostTensor(&tensor_scores_host);
    MNN::Tensor tensor_boxes_host(tensor_boxes, tensor_boxes->getDimensionType());
    tensor_boxes->copyToHostTensor(&tensor_boxes_host);

using opencv pre-process, I can get the same result as onnx model, but mnn cv can not.

PS: an old version mnn can get same result, but I do not know the version number. ( only have .so)

Uploading slim-Epoch-500_320_8c_1.onnx.tar.gz…

[jxt1234]

It appears in x86_64 ? If close MNN_USE_SSE, what's the result?

[jfkkf123]

It appears in x86_64 ? If close MNN_USE_SSE, what's the result?

yes. It appears in x86_64. I close MNN_USE_SSE to build a new libmnn.so. It also get the wrong result.
I think there is something deference between old version and 2.9.5 in imagepro cess.
When I use old version libmnn. It tell me Imageprocess don't support caffe dimension type,please create tensor with type TensorFlow or caffe_c4. So I convert onnx model to an nc4hw4.mnn model, old version mnn give me right result. But I use nc4hw4.mnn in 2.9.5 version mnn, it gives wrong result.

[jxt1234]

The onnx file is not upload successfully. Could you upload it again?

[jfkkf123]

slim-Epoch-500_320_8c_1.onnx.tar.gz

[jxt1234]

Could you upload correct input and output txt ?

[jfkkf123]

Could you upload correct input and output txt ?

which part data do you need？ before OpenCV pre-process or after pre-process？or others.
I think correct input data can't help you find out reason. I give you the all code and test image file?

[jxt1234]

Could you upload correct input and output txt ?

which part data do you need？ before OpenCV pre-process or after pre-process？or others. I think correct input data can't help you find out reason. I give you the all code and test image file?

OK. I use a random picture can't reproduce the error. It seem the same as the expect value.

[jfkkf123]

test.tar.gz