.NET4.7.2+VS2017下Tensorflow.Binding类型初始化异常求解

环境说明 #

• .NET Framework 4.7.2
• Visual Studio 2017
• Tensorflow.NET v0.150.0（Haiping Cheng开发）

自定义TfMatrix类源码 #

using System;
using System.Linq;
using Tensorflow;
using static Tensorflow.Binding;

namespace SimulationLib
{
    public class TfMatrix : IEquatable<TfMatrix>
    {
        private Tensor _matrix;

        public TfMatrix(float[,] values)
        {
            if (values == null || values.GetLength(0) == 0 || values.GetLength(1) == 0)
                throw new ArgumentException("Values array must not be null or empty.");

            _matrix = tf.constant(values);
        }

        public TfMatrix(int rows, int cols)
        {
            if (rows <= 0 || cols <= 0)
                throw new ArgumentException("Number of rows and columns must be positive.");

            _matrix = tf.zeros(new int[] { rows, cols }, tf.float32);
        }

        public TfMatrix(TfMatrix m)
        {
            if (m == null)
                throw new ArgumentNullException(nameof(m));

            _matrix = tf.identity(m._matrix); // Using identity instead of Clone
        }

        public TfMatrix(Tensor tensor)
        {
            if (tensor == null)
                throw new ArgumentNullException(nameof(tensor));

            _matrix = tensor; // Directly assign the provided tensor to the matrix field
        }

        // Indexer for accessing and modifying elements
        public float this[int row, int col]
        {
            get
            {
                var slice = tf.strided_slice(_matrix, new int[] { row, col }, new int[] { row + 1, col + 1 });
                return slice.numpy().ToArray<float>().First();
            }
            set
            {
                // Convert the matrix to a mutable tensor (this will create a copy)
                var updatedMatrix = _matrix.numpy();

                // Update the specific element
                updatedMatrix[row, col] = value;

                // Convert the updated array back to a tensor
                _matrix = tf.constant(updatedMatrix);
            }
        }

        public int Rows => (int)_matrix.shape[0];
        public int Columns => (int)_matrix.shape[1];
    }
}

测试代码 #

using System;
using Microsoft.VisualStudio.TestTools.UnitTesting;
using Tensorflow;
using static Tensorflow.Binding;

namespace SimulationLib.Tests
{
    [TestClass]
    public class TfMatrixTests
    {
        [TestMethod]
        public void TestConstructorWith2DArray()
        {
            float[,] values = new float[,] { { 1, 2 }, { 3, 4 } };
            TfMatrix matrix = new TfMatrix(values);

            Assert.AreEqual(2, matrix.Rows);
            Assert.AreEqual(2, matrix.Columns);
            Assert.AreEqual(1, matrix[0, 0]);
            Assert.AreEqual(4, matrix[1, 1]);
        }
    }
}

报错信息 #

测试名称: TestConstructorWith2DArray
测试完整名称: SimulationLib.Tests.TfMatrixTests.TestConstructorWith2DArray
测试来源: C:\git\Simulation\Simulation\unit_test_3d_for_TfMatrix\UnitTest1.cs : line 12
测试结果: 失败
测试时长: 0:00:01.2855111

调用堆栈跟踪:
at System.Reflection.RuntimeModule.GetTypes(RuntimeModule module)
at System.Reflection.RuntimeModule.GetTypes()
at System.Reflection.Assembly.GetTypes()
at Tensorflow.ops.RegisterFromAssembly()
at Tensorflow.tensorflow.InitGradientEnvironment()
at Tensorflow.tensorflow..ctor()
at Tensorflow.Binding..cctor()
--- 内部异常堆栈跟踪结束 ---
at Tensorflow.Binding.get_tf()
at SimulationLib.TfMatrix..ctor(Single[,] values) in C:\git\Simulation\Simulation\PolymerSimLib\Primitives\TfMatrix.cs:line 17
at SimulationLib.Tests.TfMatrixTests.TestConstructorWith2DArray() in C:\git\Simulation\Simulation\unit_test_3d_for_TfMatrix\UnitTest1.cs:line 15

结果信息:
测试方法SimulationLib.Tests.TfMatrixTests.TestConstructorWith2DArray引发异常:
System.TypeInitializationException: The type initializer for 'Tensorflow.Binding' threw an exception.
System.Reflection.ReflectionTypeLoadException: Unable to load one or more of the requested types. Retrieve the LoaderExceptions property for more information.

解决方法 #

1. 检查并修复NuGet依赖 #

• 确保Tensorflow.NET v0.150.0及其所有依赖包（如Tensorflow.Keras、Google.Protobuf、System.Memory等）都已正确安装，版本完全匹配。
• 清理NuGet缓存后重新安装：打开VS2017的工具→NuGet包管理器→程序包管理器设置，点击清除所有NuGet缓存，然后重新安装Tensorflow.NET v0.150.0。

2. 添加程序集绑定重定向 #

在项目的App.config中添加绑定重定向，解决版本冲突问题：

<runtime>
  <assemblyBinding xmlns="urn:schemas-microsoft-com:asm.v1">
    <dependentAssembly>
      <assemblyIdentity name="Google.Protobuf" publicKeyToken="a7d26565bac4d604" culture="neutral" />
      <bindingRedirect oldVersion="0.0.0.0-3.19.0.0" newVersion="3.19.0.0" />
    </dependentAssembly>
    <dependentAssembly>
      <assemblyIdentity name="System.Memory" publicKeyToken="cc7b13ffcd2ddd51" culture="neutral" />
      <bindingRedirect oldVersion="0.0.0.0-4.0.1.1" newVersion="4.0.1.1" />
    </dependentAssembly>
    <!-- 根据实际LoaderExceptions添加其他程序集的绑定重定向 -->
  </assemblyBinding>
</runtime>

3. 排查具体加载失败的程序集 #

修改测试代码，捕获并输出LoaderExceptions的详细信息，明确问题根源：

[TestMethod]
public void TestConstructorWith2DArray()
{
    try
    {
        float[,] values = new float[,] { { 1, 2 }, { 3, 4 } };
        TfMatrix matrix = new TfMatrix(values);

        Assert.AreEqual(2, matrix.Rows);
        Assert.AreEqual(2, matrix.Columns);
        Assert.AreEqual(1, matrix[0, 0]);
        Assert.AreEqual(4, matrix[1, 1]);
    }
    catch (TypeInitializationException ex)
    {
        if (ex.InnerException is ReflectionTypeLoadException loadEx)
        {
            foreach (var loaderEx in loadEx.LoaderExceptions)
            {
                Console.WriteLine("加载异常：");
                Console.WriteLine(loaderEx.Message);
                Console.WriteLine(loaderEx.StackTrace);
                Console.WriteLine("---");
            }
        }
        throw;
    }
}

运行测试后查看输出，根据提示安装缺失的程序集或调整版本。

4. 验证Tensorflow原生库兼容性 #

• 将项目平台目标设置为x64：右键项目→属性→生成→平台目标，选择x64（Tensorflow.NET v0.150.0对x86支持有限）。
• 安装对应版本的Tensorflow原生库：通过NuGet安装Tensorflow.Runtime.Windows v0.150.0，确保原生库与Tensorflow.NET版本匹配。

5. 调整项目设置 #

• 确认项目语言版本为C# 7.0及以上：右键项目→属性→生成→高级→语言版本，选择C# 7.0。
• 关闭"仅我的代码"：打开调试→选项→调试→常规，取消勾选"启用仅我的代码"，以便查看更完整的异常详情。

内容的提问来源于stack exchange，提问作者user366312