问题描述
我正在使用 ONNXRuntime 来推断 UNet 模型,作为预处理的一部分,我必须将 EMGU OpenCV 矩阵转换为 OnnxRuntime.Tensor。
我使用两个嵌套的 for 循环实现了它,不幸的是它很慢:
var data = new DenseTensor<float>(new[] { 1,3,WIDTH,HEIGHT});
for (int y = 0; y < HEIGHT; y++)
{
for (int x = 0; x < WIDTH; x++)
{
data[0,x,y] = image.GetValue(2,y,x)/255.0;
data[0,1,y] = image.GetValue(1,2,y] = image.GetValue(0,x)/255.0;
}
}
然后我发现存在一种将 Array 转换为 DenseTensor 的方法。我想使用这种方法如下:
var imgToPredictFloat = new Mat(image.Height,image.Width,DepthType.Cv32F,3);
image.ConvertTo(imgToPredictFloat,1/255.0);
CvInvoke.CvtColor(imgToPredictFloat,imgToPredictFloat,ColorConversion.Bgra2Rgb);
var data = image.GetData().ToTensor<float>;
var reshaped = data.Reshape(new int[] { 1,HEIGHT});
这将大大提高性能,但是输出张量的布局不正确(与 for 循环相同)并且模型显然无法工作。有什么建议可以将数组重塑为正确的布局吗?
在代码中还执行了将 int 0-255 转换为 float 0-1 并将 BGR 布局转换为 RGB 布局。
解决方法
这就是我使用 cv::Mat 和 ONNX Runtime (C++) 的方式:
const wchar_t* model_path = L"C:/data/DNN/ONNX/ResNet/resnet152v2/resnet152-v2-7.onnx";
printf("Using Onnxruntime C++ API\n");
Ort::Session session(env,model_path,session_options);
//*************************************************************************
// print model input layer (node names,types,shape etc.)
Ort::AllocatorWithDefaultOptions allocator;
size_t num_output_nodes = session.GetOutputCount();
std::vector<char*> outputNames;
for (size_t i = 0; i < num_output_nodes; ++i)
{
char* name = session.GetOutputName(i,allocator);
std::cout << "output: " << name << std::endl;
outputNames.push_back(name);
}
// print number of model input nodes
size_t num_input_nodes = session.GetInputCount();
std::vector<const char*> input_node_names(num_input_nodes);
std::vector<int64_t> input_node_dims; // simplify... this model has only 1 input node {1,3,224,224}.
// Otherwise need vector<vector<>>
printf("Number of inputs = %zu\n",num_input_nodes);
// iterate over all input nodes
for (int i = 0; i < num_input_nodes; i++) {
// print input node names
char* input_name = session.GetInputName(i,allocator);
printf("Input %d : name=%s\n",i,input_name);
input_node_names[i] = input_name;
// print input node types
Ort::TypeInfo type_info = session.GetInputTypeInfo(i);
auto tensor_info = type_info.GetTensorTypeAndShapeInfo();
ONNXTensorElementDataType type = tensor_info.GetElementType();
printf("Input %d : type=%d\n",type);
// print input shapes/dims
input_node_dims = tensor_info.GetShape();
printf("Input %d : num_dims=%zu\n",input_node_dims.size());
for (int j = 0; j < input_node_dims.size(); j++)
printf("Input %d : dim %d=%jd\n",j,input_node_dims[j]);
}
cv::Size dnnInputSize;
cv::Scalar mean;
cv::Scalar std;
bool rgb = true;
//cv::Mat inputImage = cv::imread("C:/TestImages/kitten_01.jpg");
cv::Mat inputImage = cv::imread("C:/TestImages/slug_01.jpg");
rgb = true;
dnnInputSize = cv::Size(224,224);
mean[0] = 0.485;
mean[1] = 0.456;
mean[2] = 0.406;
std[0] = 0.229;
std[1] = 0.224;
std[2] = 0.225;
cv::Mat blob;
// ONNX: (N x 3 x H x W)
cv::dnn::blobFromImage(inputImage,blob,1.0 / 255.0,dnnInputSize,mean,rgb,false);
size_t input_tensor_size = blob.total();
std::vector<float> input_tensor_values(input_tensor_size);
for (size_t i = 0; i < input_tensor_size; ++i)
{
input_tensor_values[i] = blob.at<float>(i);
}
std::vector<const char*> output_node_names = { outputNames.front() };
// create input tensor object from data values
auto memory_info = Ort::MemoryInfo::CreateCpu(OrtArenaAllocator,OrtMemTypeDefault);
Ort::Value input_tensor = Ort::Value::CreateTensor<float>(memory_info,input_tensor_values.data(),input_tensor_size,input_node_dims.data(),4);
assert(input_tensor.IsTensor());
// score model & input tensor,get back output tensor
auto output_tensors = session.Run(Ort::RunOptions{ nullptr },input_node_names.data(),&input_tensor,1,output_node_names.data(),1);
assert(output_tensors.size() == 1 && output_tensors.front().IsTensor());
// Get pointer to output tensor float values
float* floatarr = output_tensors.front().GetTensorMutableData<float>();
assert(abs(floatarr[0] - 0.000045) < 1e-6);
cv::Mat1f result = cv::Mat1f(1000,floatarr);
cv::Point classIdPoint;
double confidence = 0;
minMaxLoc(result,&confidence,&classIdPoint);
int classId = classIdPoint.y;
std::cout << "confidence: " << confidence << std::endl;
std::cout << "class: " << classId << std::endl;
您需要的实际转换部分是 imho(根据您的网络调整大小和平均值/标准):
cv::Mat inputImage = cv::imread("C:/TestImages/slug_01.jpg");
rgb = true;
dnnInputSize = cv::Size(224,false);