Algorithm-Repo/SourceCode/CVAlgorithmUtils/YOLODetectPostProcessUtils/YOLODetectProcessLib/InferenceNetworkUtils.cs

using System;
using System.IO;
using System.Collections.Generic;
using Microsoft.ML.OnnxRuntime;

namespace YOLODetectProcessLib
{    
     /// <summary>
     /// 推理网络共用的一些辅助的功能函数
     /// </summary>
    public class InferenceNetworkUtils
    {

        #region public 

        /// <summary>
        /// 从文件中读出网络参数
        /// </summary>
        /// <param name="netsDir"></param>
        /// <param name="netFileName"></param>
        /// <param name="netHashCode"></param>
        /// <returns></returns>
        public static byte[] ReadNetworkDataFromFile(string netsDir, string netFileName, string netHashCode)
        {
            // 网络权值文件是否存在
            string netFilePath;
            if (Directory.Exists(netsDir))
            {
                netFilePath = Path.Combine(netsDir, netFileName);
            }
            else
            {
                netFilePath = netFileName;
            }
            if (!File.Exists(netFilePath))
            {
                throw new FileNotFoundException("Failed to load network:" + netFileName.ToString() + " from:" + netFilePath + ".");
            }
            // 读入（只读不写，不需要加锁）
            byte[] fileDataEncrypted = File.ReadAllBytes(netFilePath);
            // 检查模型文件的哈希值是否为当前版本所需
            string hashstr = HashCode.ComputeHashCode(fileDataEncrypted);
            if (hashstr != netHashCode)
            {
                throw new ArgumentException("netFilePath", "Unexpected parameter data file(" + netFilePath +
                    ") for the current version of AIDiagSystem");
            }
            // 解密
            var fileDataDecrypted = AES.AESDecrypt(fileDataEncrypted);
            // 返回
            return fileDataDecrypted;
        }


        /// <summary>
        /// 检查Onnx模型的输入输出tensor名和尺寸是否符合要求
        /// </summary>
        /// <param name="inferSession"></param>
        /// <param name="inputTensorNames"></param>
        /// <param name="outputTensorNames"></param>
        /// <param name="inputTensorDimensions"></param>
        /// <param name="outputTensorDimensions"></param>
        public static void CheckOnnxModel(InferenceSession inferSession, string[] inputTensorNames, string[] outputTensorNames,
            List<int[]> inputTensorDimensions, List<int[]> outputTensorDimensions)
        {
            // 检查模型输入输出tensor名
            if (inferSession.InputMetadata.Count != inputTensorNames.Length)
            {
                throw new ArgumentOutOfRangeException("input", "The expected model input number is " + inputTensorNames.Length +
                    ", but got " + inferSession.InputMetadata.Count + ".");
            }
            foreach (string inputTensorName in inputTensorNames)
            {
                if (!inferSession.InputMetadata.ContainsKey(inputTensorName))
                {
                    throw new ArgumentOutOfRangeException("inputTensorName", "Find no input tensor with expected name of " + inputTensorName + ".");
                }
                if (inferSession.InputMetadata[inputTensorName].ElementType != typeof(float))
                {
                    throw new ArgumentException("inputTensorType", "The expected input tensor type is float, but got " +
                        inferSession.InputMetadata[inputTensorName].ElementType.ToString() + ".");
                }
            }
            //if (inferSession.OutputMetadata.Count != outputTensorNames.Length)
            //{
            //    throw new ArgumentOutOfRangeException("output", "The expected model output number is " + outputTensorNames.Length +
            //        ", but got " + inferSession.OutputMetadata.Count + ".");
            //}
            foreach (string outputTensorName in outputTensorNames)
            {
                if (!inferSession.OutputMetadata.ContainsKey(outputTensorName))
                {
                    throw new ArgumentOutOfRangeException("outputTensorName", "Find no output tensor with expected name of " + outputTensorName + ".");
                }
                if (inferSession.OutputMetadata[outputTensorName].ElementType != typeof(float))
                {
                    throw new ArgumentException("outputTensorType", "The expected output tensor type if float, but got " +
                        inferSession.OutputMetadata[outputTensorName].ElementType.ToString() + ".");
                }
            }
            // 检查输入尺寸
            if (inputTensorDimensions.Count != inputTensorNames.Length)
            {
                throw new ArgumentException("inputTensorDimensions", "The inputTensorDimensions and inputTensorNames should have the same length.");
            }
            for (int ni = 0; ni < inputTensorNames.Length; ni++)
            {
                var inputTensorDimension = inferSession.InputMetadata[inputTensorNames[ni]].Dimensions;
                var inputTensorDimensionExpected = inputTensorDimensions[ni];
                if (inputTensorDimension.Length != inputTensorDimensionExpected.Length)
                {
                    throw new ArgumentException("inputTensorDimensionLength", "The expected input tensor dimension length is " +
                        inputTensorDimensionExpected.Length + ", but got " + inputTensorDimension.Length + ".");
                }
                if (inputTensorDimension.Length <= 1)
                {
                    continue;
                }
                for (int nj = 0; nj < inputTensorDimension.Length; nj++)
                {
                    if (inputTensorDimension[nj] != inputTensorDimensionExpected[nj] && inputTensorDimension[nj] != -1)
                    {
                        throw new ArgumentException("inputTensorDimension", "The expected input tensor dimension is " +
                            inputTensorDimensionExpected[nj] + ", but got " + inputTensorDimension[nj] + " for the " + nj + "th axis.");
                    }
                }
            }
            // 检查输出尺寸
            if (outputTensorDimensions.Count != outputTensorNames.Length)
            {
                throw new ArgumentException("outputTensorDimensions", "The outputTensorDimensions and outputTensorNames should have the same length.");
            }
            for (int ni = 0; ni < outputTensorNames.Length; ni++)
            {
                var outputTensorDimension = inferSession.OutputMetadata[outputTensorNames[ni]].Dimensions;
                var outputTensorDimensionExpected = outputTensorDimensions[ni];
                if (outputTensorDimension.Length != outputTensorDimensionExpected.Length)
                {
                    throw new ArgumentException("outputTensorDimensionLength", "The expected output tensor dimension length is " +
                        outputTensorDimensionExpected.Length + ", but got " + outputTensorDimension.Length + ".");
                }
                if (outputTensorDimension.Length <= 1)
                {
                    continue;
                }
                for (int nj = 0; nj < outputTensorDimension.Length; nj++)
                {
                    if (outputTensorDimension[nj] != outputTensorDimensionExpected[nj] && outputTensorDimension[nj] != -1)
                    {
                        throw new ArgumentException("outputTensorDimension", "The expected output tensor dimension is" +
                            outputTensorDimensionExpected[nj] + ", but got " + outputTensorDimension[nj] + " for the " + nj + "th axis.");
                    }
                }
            }
        }

        #endregion

    }

}