TypeScript 深度学习：神经网络类型安全

            
interface ImageData {
  width: number;
  height: number;
  channels: number; // Grayscale: 1, RGB: 3, etc.
  data: number[]; // Pixel data (0-255)
}

function processImage(image: ImageData): void {
  // ... image processing logic ...
}

// Example usage:
const myImage: ImageData = {
  width: 28,
  height: 28,
  channels: 1,
  data: new Array(28 * 28).fill(0) // Initialize with zeros
};

processImage(myImage);

            

              
                Copy
              
            
          

            
interface DenseLayerParams {
  inputUnits: number;
  outputUnits: number;
  activation: 'relu' | 'sigmoid' | 'tanh'; // Restrict activation function choices
  weightInitializer?: 'random' | 'zeros'; // Optional weight initialization strategy
}

class DenseLayer {
  private weights: number[][];
  private biases: number[];

  constructor(params: DenseLayerParams) {
    // ... weight and bias initialization logic based on params ...
    this.weights = Array(params.inputUnits).fill(null).map(() => Array(params.outputUnits).fill(0)); // Example initialization
    this.biases = Array(params.outputUnits).fill(0);
  }

  forward(input: number[]): number[] {
    // ... forward propagation logic ...
    return []; // Replace with actual output
  }
}

// Example usage:
const denseLayerParams: DenseLayerParams = {
  inputUnits: 784,
  outputUnits: 128,
  activation: 'relu',
  weightInitializer: 'random'
};

const denseLayer = new DenseLayer(denseLayerParams);

            

              
                Copy
              
            
          

            
// Simple Tensor type (can be expanded for multi-dimensional tensors)
type Tensor = number[];

function matrixMultiply(a: Tensor, b: Tensor, aRows: number, aCols: number, bRows: number, bCols: number): Tensor {
  if (aCols !== bRows) {
    throw new Error("Matrix dimensions are incompatible for multiplication.");
  }

  const result: Tensor = new Array(aRows * bCols).fill(0);

  for (let i = 0; i < aRows; i++) {
    for (let j = 0; j < bCols; j++) {
      for (let k = 0; k < aCols; k++) {
        result[i * bCols + j] += a[i * aCols + k] * b[k * bCols + j];
      }
    }
  }

  return result;
}

// Example Usage:
const matrixA: Tensor = [1, 2, 3, 4, 5, 6]; // 2x3 matrix
const matrixB: Tensor = [7, 8, 9, 10, 11, 12]; // 3x2 matrix

try {
  const resultMatrix = matrixMultiply(matrixA, matrixB, 2, 3, 3, 2);
  console.log("Result Matrix:", resultMatrix);
} catch (error: any) {
  console.error("Error during matrix multiplication:", error.message);
}

            

              
                Copy
              
            
          

            
import * as tf from '@tensorflow/tfjs';

interface NetworkConfig {
  inputShape: number[];
  layers: LayerConfig[];
  optimizer?: tf.Optimizer;
}

interface LayerConfig {
  type: 'dense';
  units: number;
  activation: 'relu' | 'sigmoid' | 'softmax';
}

class NeuralNetwork {
  private model: tf.Sequential;
  private config: NetworkConfig;

  constructor(config: NetworkConfig) {
    this.config = config;
    this.model = tf.sequential();
    this.buildModel();
  }

  private buildModel(): void {
    this.config.layers.forEach((layerConfig) => {
      if (layerConfig.type === 'dense') {
        this.model.add(tf.layers.dense({
          units: layerConfig.units,
          activation: layerConfig.activation,
          inputShape: this.config.inputShape
        }));
      }
    });

    this.model.compile({
      optimizer: this.config.optimizer || 'adam',
      loss: 'categoricalCrossentropy',
      metrics: ['accuracy']
    });
  }

  async train(xTrain: tf.Tensor, yTrain: tf.Tensor, epochs: number): Promise {
    const history = await this.model.fit(xTrain, yTrain, {
      epochs: epochs,
      validationSplit: 0.1
    });
    return history;
  }

  predict(input: tf.Tensor): tf.Tensor {
    return this.model.predict(input) as tf.Tensor;
  }
}

// Example Usage:
const config: NetworkConfig = {
  inputShape: [784], // MNIST image size (28x28)
  layers: [
    { type: 'dense', units: 128, activation: 'relu' },
    { type: 'dense', units: 10, activation: 'softmax' } // 10 output classes (digits 0-9)
  ]
};

const model = new NeuralNetwork(config);

// Dummy Data (replace with actual MNIST data)
const xTrain = tf.randomNormal([100, 784]);
const yTrain = tf.oneHot(tf.randomUniform([100], 0, 10, 'int32'), 10);

model.train(xTrain, yTrain, 10).then((history) => {
  console.log("Training complete:", history);
  const prediction = model.predict(xTrain.slice([0], [1]));
  console.log("Prediction:", prediction.toString());
});

            

              
                Copy