### 在 C# 中使用 TensorFlow 将张量数据拆分为训练集和测试集 在 C# 的开发环境下，通过 TensorFlow.NET 库可以轻松实现张量数据的划分。以下是详细的解决方案以及代码示例。 #### 使用随机索引方法拆分张量 一种常见的做法是利用随机打乱后的索引来分割原始张量。这种方法能够确保训练集和测试集中的样本分布均匀且不重叠。 ```csharp using System; using System.Linq; using Tensorflow; public class TensorSplitter { /// <summary> /// Splits the given tensor into training and testing tensors based on specified ratio. /// </summary> /// <param name="tensor">Input tensor to be split.</param> /// <param name="testRatio">Proportion of dataset allocated for testing (e.g., 0.2).</param> /// <returns>Tuple containing two tensors: training set and testing set.</returns> public static (TFTensor trainTensor, TFTensor testTensor) SplitTensor(TFTensor tensor, double testRatio = 0.2) { if (tensor == null || tensor.shape.num_elements() == 0) throw new ArgumentException("Input tensor cannot be empty or null."); int totalElements = (int)tensor.shape.num_elements(); int testSize = (int)(totalElements * testRatio); int trainSize = totalElements - testSize; Random random = new Random(); var shuffledIndices = Enumerable.Range(0, totalElements).OrderBy(_ => random.Next()).ToArray(); // Extract values using reshaped indices var flatArray = tensor.numpy().flatten<float>(); float[] trainData = shuffledIndices.Take(trainSize).Select(index => flatArray[index]).ToArray(); float[] testData = shuffledIndices.Skip(trainSize).Take(testSize).Select(index => flatArray[index]).ToArray(); // Reshape back to original dimensions excluding batch size dimension long[] shapeWithoutBatch = tensor.shape.dims.Skip(1).ToArray(); return ( TFTensor.FromArray(ReshapeToOriginalDimensions(trainData, shapeWithoutBatch)), TFTensor.FromArray(ReshapeToOriginalDimensions(testData, shapeWithoutBatch)) ); } private static float[][] ReshapeToOriginalDimensions(float[] dataArray, long[] targetShape) { int batchSize = dataArray.Length / (targetShape.Aggregate(1L, (a, b) => a * b)).ToInt32Checked(); var result = new float[batchSize][]; for(int i=0;i<batchSize;i++) { int startIdx=i*(int)(targetShape.Aggregate(1,(p,q)=>p*q)); result[i]=dataArray[startIdx..(startIdx+(int)(targetShape.Aggregate(1,(r,s)=>r*s)))].ToArray(); } return result; } } ``` 此代码片段展示了如何将一维或多维张量按照一定比例划分为训练集与测试集[^1]。它首先计算总的元素数量并决定每组应包含多少条记录；接着创建了一个随机序列用于重新安排位置顺序以便后续选取对应范围内的项目形成最终的结果对象[^2]。 #### 集成到完整的流程中 下面是一个简单的例子展示怎样调用上面定义的方法来进行实际操作： ```csharp static void Main(string[] args) { // Create sample tensor with shape [100, 4] TFShape shape = new TFShape(new long[]{100, 4}); TFTensor fullTensor = tf.random_uniform(shape, minval:0f,maxval:1f,dtype:TFTypes.Float32); // Perform splitting operation var (trainingSet, testingSet) = TensorSplitter.SplitTensor(fullTensor, testRatio:0.3); Console.WriteLine($"Training Set Shape:{trainingSet.shape}"); Console.WriteLine($"Testing Set Shape:{testingSet.shape}"); // Continue building your model here... } ``` 以上演示了从构建初始张量到最后打印出各部分形状的过程[^3]。可以根据自己的需求继续添加更多功能或者直接投入建模阶段。 --- ###