microsoft / onnxruntime

ONNX Runtime: cross-platform, high performance ML inferencing and training accelerator
https://onnxruntime.ai
MIT License
14.78k stars 2.94k forks source link

the input format and output format not match onnx model #22797

Open Chriost opened 1 week ago

Chriost commented 1 week ago

Describe the issue

auto inputShapeInfo = session_.GetInputTypeInfo(i).GetTensorTypeAndShapeInfo().GetShape();
int ch = inputShapeInfo[1];
input_h = inputShapeInfo[2];
input_w = inputShapeInfo[3];
std::cout << "input format: " << ch << "x" << input_h << "x" << input_w << std::endl;

Image

input_nodes_num:1 output_nodes_num:2 input format: 1x-1x-1 output format: -1x-1 output format: -1x256

To reproduce

`#include "onnxruntime_cxx_api.h"

include "cpu_provider_factory.h"

include <opencv2/opencv.hpp>

include

int main() {

// 测试图片
cv::Mat image = cv::imread("-00000001.jpg");
//cv::imshow("输入图", image);
cv::resize(image, image, cv::Size(image.cols / 4, image.rows / 4));
// 初始化ONNXRuntime环境
Ort::Env env = Ort::Env(ORT_LOGGING_LEVEL_ERROR, "superpoint_v1");

// 设置会话选项
Ort::SessionOptions session_options;
// 优化器级别:基本的图优化级别
session_options.SetGraphOptimizationLevel(ORT_ENABLE_BASIC);
// 线程数:4
session_options.SetIntraOpNumThreads(1);
// 设备使用优先使用GPU而是才是CPU
std::cout << "onnxruntime inference try to use GPU Device" << std::endl;
//OrtSessionOptionsAppendExecutionProvider_CUDA(session_options, 0);
OrtSessionOptionsAppendExecutionProvider_CPU(session_options, 0);
// onnx训练模型文件
std::string onnxpath = "LoadOnnx\\superpoint_v1.onnx";
std::wstring modelPath = std::wstring(onnxpath.begin(), onnxpath.end());

// 加载模型并创建会话
Ort::Session session_(env, modelPath.c_str(), session_options);
// 获取模型输入输出信息
int input_nodes_num = session_.GetInputCount();         // 输入节点输
std::cout << "input_nodes_num:" << input_nodes_num << std::endl;
int output_nodes_num = session_.GetOutputCount();       // 输出节点数
std::cout << "output_nodes_num:" << output_nodes_num << std::endl;
std::vector<std::string> input_node_names;              // 输入节点名称
std::vector<std::string> output_node_names;             // 输出节点名称
Ort::AllocatorWithDefaultOptions allocator;
// 输入图像尺寸
int input_h = 0;
int input_w = 0;

// 获取模型输入信息
for (int i = 0; i < input_nodes_num; i++) {
    // 获得输入节点的名称并存储
    auto input_name = session_.GetInputNameAllocated(i, allocator);
    input_node_names.push_back(input_name.get());
    // 显示输入图像的形状
    auto inputShapeInfo = session_.GetInputTypeInfo(i).GetTensorTypeAndShapeInfo().GetShape();
    int ch = inputShapeInfo[1];
    input_h = inputShapeInfo[2];
    input_w = inputShapeInfo[3];
    std::cout << "input format: " << ch << "x" << input_h << "x" << input_w << std::endl;
}

// 获取模型输出信息
int num = 0;
int nc = 0;
for (int i = 0; i < output_nodes_num; i++) {
    // 获得输出节点的名称并存储
    auto output_name = session_.GetOutputNameAllocated(i, allocator);
    output_node_names.push_back(output_name.get());
    // 显示输出结果的形状
    auto outShapeInfo = session_.GetOutputTypeInfo(i).GetTensorTypeAndShapeInfo().GetShape();
    num = outShapeInfo[0];
    nc = outShapeInfo[1];
    std::cout << "output format: " << num << "x" << nc << std::endl;
}

// 预处理输入数据
cv::Mat rgb, blob;
// 默认是BGR需要转化成RGB
cv::cvtColor(image, rgb, cv::COLOR_BGR2RGB);
// 对图像尺寸进行缩放
cv::resize(rgb, blob, cv::Size(input_w, input_h));
blob.convertTo(blob, CV_32F);
// 对图像进行标准化处理
blob = blob / 255.0;    // 归一化
cv::subtract(blob, cv::Scalar(0.485, 0.456, 0.406), blob);  // 减去均值
cv::divide(blob, cv::Scalar(0.229, 0.224, 0.225), blob);    //除以方差
// CHW-->NCHW 维度扩展
cv::Mat timg = cv::dnn::blobFromImage(blob);
std::cout << timg.size[0] << "x" << timg.size[1] << "x" << timg.size[2] << "x" << timg.size[3] << std::endl;
// 占用内存大小,后续计算是总像素*数据类型大小
size_t tpixels = input_h * input_w * 3;
std::array<int64_t, 4> input_shape_info{ 1, 3, input_h, input_w };

// 准备数据输入
auto allocator_info = Ort::MemoryInfo::CreateCpu(OrtDeviceAllocator, OrtMemTypeCPU);
Ort::Value input_tensor_ = Ort::Value::CreateTensor<float>(allocator_info, timg.ptr<float>(), tpixels, input_shape_info.data(), input_shape_info.size());

// 模型输入输出所需数据(名称及其数量),模型只认这种类型的数组
const std::array<const char*, 1> inputNames = { input_node_names[0].c_str() };
const std::array<const char*, 1> outNames = { output_node_names[0].c_str() };

// 模型推理
std::vector<Ort::Value> ort_outputs;
try {
    ort_outputs = session_.Run(Ort::RunOptions{ nullptr }, inputNames.data(), &input_tensor_, 1, outNames.data(), outNames.size());
}
catch (std::exception e) {
    std::cout << e.what() << std::endl;
}
// 1x5 获取输出数据并包装成一个cv::Mat对象,为了方便后处理
const float* pdata = ort_outputs[0].GetTensorMutableData<float>();
cv::Mat prob(num, nc, CV_32F, (float*)pdata);

// 后处理推理结果
cv::Point maxL, minL;       // 用于存储图像分类中的得分最小值索引和最大值索引(坐标)
double maxv, minv;          // 用于存储图像分类中的得分最小值和最大值
cv::minMaxLoc(prob, &minv, &maxv, &minL, &maxL);

int max_index = maxL.x;     // 获得最大值的索引,只有一行所以列坐标既为索引
std::cout << "label id: " << max_index << std::endl;
// 在测试图像上加上预测的分类标签
//cv::putText(image, labels[max_index], cv::Point(50, 50), cv::FONT_HERSHEY_SIMPLEX, 1.0, cv::Scalar(0, 0, 255), 2, 8);
cv::imshow("输入图像", image);
cv::waitKey(0);

// 释放资源
session_options.release();
session_.release();
return 0;

} `

Urgency

No response

Platform

Windows

OS Version

10

ONNX Runtime Installation

Released Package

ONNX Runtime Version or Commit ID

onnxruntime-win-x64-gpu-1.16.2

ONNX Runtime API

C++

Architecture

X64

Execution Provider

CUDA

Execution Provider Library Version

CUDA11.2