TensorFlow.js-Beispiele

// TensorFlow.js Hello World

// Import TensorFlow.js
// In browser: <script src="https://cdn.jsdelivr.net/npm/@tensorflow/tfjs@latest/dist/tf.min.js"></script>
// In Node.js: const tf = require('@tensorflow/tfjs');

// 1. Basic tensor creation
console.log('=== Basic Tensors ===');
const scalar = tf.scalar(5);
console.log('Scalar:', scalar.toString());

const vector = tf.tensor1d([1, 2, 3, 4]);
console.log('Vector:', vector.toString());

const matrix = tf.tensor2d([[1, 2], [3, 4]]);
console.log('Matrix:', matrix.toString());

// 2. Tensor operations
console.log('\n=== Basic Operations ===');
const a = tf.tensor1d([1, 2, 3]);
const b = tf.tensor1d([4, 5, 6]);

const add = a.add(b);
console.log('Add:', add.toString());

const multiply = a.mul(b);
console.log('Multiply:', multiply.toString());

const squared = a.square();
console.log('Square:', squared.toString());

// 3. Matrix multiplication
console.log('\n=== Matrix Operations ===');
const mat1 = tf.tensor2d([[1, 2], [3, 4]]);
const mat2 = tf.tensor2d([[5, 6], [7, 8]]);

const matMul = mat1.matMul(mat2);
console.log('Matrix Multiplication:', matMul.toString());

// 4. Mathematical functions
console.log('\n=== Mathematical Functions ===');
const x = tf.tensor1d([-1, 0, 1, 2]);

const relu = x.relu();
console.log('ReLU:', relu.toString());

const sigmoid = tf.sigmoid(x);
console.log('Sigmoid:', sigmoid.toString());

const tanh = tf.tanh(x);
console.log('Tanh:', tanh.toString());

// 5. Reshaping tensors
console.log('\n=== Reshaping ===');
const tensor = tf.tensor1d([1, 2, 3, 4, 5, 6]);
const reshaped = tensor.reshape([2, 3]);
console.log('Original:', tensor.toString());
console.log('Reshaped (2x3):', reshaped.toString());

// 6. Reduction operations
console.log('\n=== Reductions ===');
const data = tf.tensor1d([1, 2, 3, 4, 5]);

const sum = data.sum();
console.log('Sum:', sum.toString());

const mean = data.mean();
console.log('Mean:', mean.toString());

const max = data.max();
console.log('Max:', max.toString());

// 7. Creating tensors with specific values
console.log('\n=== Special Tensors ===');
const zeros = tf.zeros([3, 3]);
console.log('Zeros (3x3):', zeros.toString());

const ones = tf.ones([2, 4]);
console.log('Ones (2x4):', ones.toString());

const identity = tf.eye(3);
console.log('Identity (3x3):', identity.toString());

const random = tf.randomNormal([2, 3]);
console.log('Random Normal (2x3):', random.toString());

// 8. Using variables (mutable tensors)
console.log('\n=== Variables ===');
const variable = tf.variable(tf.tensor1d([1, 2, 3]));
console.log('Initial variable:', variable.toString());

variable.assign(tf.tensor1d([4, 5, 6]));
console.log('Updated variable:', variable.toString());

// 9. Memory management
console.log('\n=== Memory Management ===');
// Use tf.tidy() to automatically dispose intermediate tensors
const result = tf.tidy(() => {
    const x = tf.tensor1d([1, 2, 3]);
    const y = tf.tensor1d([4, 5, 6]);
    return x.add(y).square();
});
console.log('Complex operation result:', result.toString());

// Don't forget to dispose tensors when done
scalar.dispose();
vector.dispose();
matrix.dispose();
add.dispose();
multiply.dispose();
squared.dispose();
matMul.dispose();
relu.dispose();
sigmoid.dispose();
tanh.dispose();
reshaped.dispose();
sum.dispose();
mean.dispose();
max.dispose();
zeros.dispose();
ones.dispose();
identity.dispose();
random.dispose();
variable.dispose();
result.dispose();

console.log('\nAll tensors disposed successfully!');

// 10. Asynchronous operations
async function asyncExample() {
    console.log('\n=== Async Operations ===');

    // Use tf.data API for data processing
    const data = tf.data.array([1, 2, 3, 4, 5]);
    const processed = await data.map(x => x * 2).batch(2).toArray();
    console.log('Batched data:', processed);
}

asyncExample().catch(console.error);

// Simple Neural Network with TensorFlow.js

// 1. Create synthetic training data
function createTrainingData() {
    const data = [];
    const labels = [];

    // Generate data for two classes
    for (let i = 0; i < 100; i++) {
        // Class 0: points near (0, 0)
        data.push([Math.random() * 2 - 1, Math.random() * 2 - 1]);
        labels.push([1, 0]);

        // Class 1: points near (5, 5)
        data.push([Math.random() * 2 + 4, Math.random() * 2 + 4]);
        labels.push([0, 1]);
    }

    return {
        data: tf.tensor2d(data),
        labels: tf.tensor2d(labels)
    };
}

// 2. Create the model
function createModel() {
    const model = tf.sequential();

    // Input layer
    model.add(tf.layers.dense({
        units: 8,
        activation: 'relu',
        inputShape: [2]
    }));

    // Hidden layer
    model.add(tf.layers.dense({
        units: 8,
        activation: 'relu'
    }));

    // Output layer (2 classes)
    model.add(tf.layers.dense({
        units: 2,
        activation: 'softmax'
    }));

    // Compile the model
    model.compile({
        optimizer: tf.train.adam(0.01),
        loss: 'categoricalCrossentropy',
        metrics: ['accuracy']
    });

    return model;
}

// 3. Train the model
async function trainModel(model, trainingData) {
    console.log('Starting training...');

    const history = await model.fit(trainingData.data, trainingData.labels, {
        epochs: 50,
        batchSize: 16,
        validationSplit: 0.2,
        shuffle: true,
        callbacks: {
            onEpochEnd: (epoch, logs) => {
                console.log(`Epoch ${epoch + 1}: loss = ${logs.loss.toFixed(4)}, accuracy = ${(logs.acc * 100).toFixed(2)}%`);
            }
        }
    });

    return history;
}

// 4. Make predictions
function makePredictions(model) {
    console.log('\nMaking predictions...');

    // Test points
    const testPoints = [
        [0, 0],    // Should be class 0
        [5, 5],    // Should be class 1
        [1, 1],    // Should be class 0
        [4.5, 4.5] // Should be class 1
    ];

    testPoints.forEach(point => {
        const prediction = model.predict(tf.tensor2d([point]));
        const probabilities = prediction.dataSync();
        const predictedClass = probabilities[0] > probabilities[1] ? 0 : 1;

        console.log(`Point [${point[0]}, ${point[1]}]: Class ${predictedClass} (Probabilities: [${probabilities[0].toFixed(3)}, ${probabilities[1].toFixed(3)}])`);

        prediction.dispose();
    });
}

// 5. Save and load model
async function saveAndLoadModel(model) {
    console.log('\nSaving and loading model...');

    // Save model to local storage
    await model.save('localstorage://my-model');
    console.log('Model saved to local storage');

    // Load model from local storage
    const loadedModel = await tf.loadLayersModel('localstorage://my-model');
    console.log('Model loaded from local storage');

    return loadedModel;
}

// Main execution
async function main() {
    try {
        // Create and prepare data
        const trainingData = createTrainingData();
        console.log('Training data created:', trainingData.data.shape, trainingData.labels.shape);

        // Create model
        const model = createModel();
        console.log('Model architecture:');
        model.summary();

        // Train model
        await trainModel(model, trainingData);

        // Make predictions
        makePredictions(model);

        // Save and load model
        const loadedModel = await saveAndLoadModel(model);
        console.log('\nPrediction with loaded model:');
        makePredictions(loadedModel);

        // Clean up
        trainingData.data.dispose();
        trainingData.labels.dispose();
        model.dispose();
        loadedModel.dispose();

    } catch (error) {
        console.error('Error:', error);
    }
}

// Run the example
main();

// Linear Regression with TensorFlow.js

// 1. Generate synthetic data
function generateLinearData(numPoints = 100) {
    const x = [];
    const y = [];

    // y = 2x + 1 + noise
    for (let i = 0; i < numPoints; i++) {
        const xVal = Math.random() * 10;
        const noise = (Math.random() - 0.5) * 2;
        const yVal = 2 * xVal + 1 + noise;

        x.push(xVal);
        y.push(yVal);
    }

    return {
        x: tf.tensor2d(x, [x.length, 1]),
        y: tf.tensor2d(y, [y.length, 1])
    };
}

// 2. Create linear regression model
function createLinearModel() {
    const model = tf.sequential();

    // Single neuron for linear regression
    model.add(tf.layers.dense({
        units: 1,
        inputShape: [1]
    }));

    // Compile with appropriate loss function for regression
    model.compile({
        optimizer: tf.train.sgd(0.01),
        loss: 'meanSquaredError'
    });

    return model;
}

// 3. Train the model
async function trainLinearRegression(model, data) {
    console.log('Training linear regression model...');

    const history = await model.fit(data.x, data.y, {
        epochs: 100,
        batchSize: 32,
        callbacks: {
            onEpochEnd: (epoch, logs) => {
                if (epoch % 10 === 0) {
                    console.log(`Epoch ${epoch}: loss = ${logs.loss.toFixed(4)}`);
                }
            }
        }
    });

    return history;
}

// 4. Evaluate and visualize results
function evaluateModel(model, data) {
    console.log('\nEvaluating model...');

    // Get the learned parameters
    const weights = model.getWeights();
    const weight = weights[0].dataSync()[0];
    const bias = weights[1].dataSync()[0];

    console.log(`Learned equation: y = ${weight.toFixed(3)}x + ${bias.toFixed(3)}`);
    console.log(`Expected equation: y = 2.000x + 1.000`);

    // Make predictions on test data
    const testX = tf.tensor2d([[0], [5], [10]]);
    const predictions = model.predict(testX);

    console.log('\nPredictions:');
    testX.dataSync().forEach((x, i) => {
        const pred = predictions.dataSync()[i];
        console.log(`x = ${x}, predicted y = ${pred.toFixed(3)}`);
    });

    // Calculate R-squared
    const yPred = model.predict(data.x);
    const yMean = data.y.mean();
    const totalSumSquares = data.y.sub(yMean).square().sum();
    const residualSumSquares = data.y.sub(yPred).square().sum();
    const rSquared = tf.tensor1d([1]).sub(residualSumSquares.div(totalSumSquares));

    console.log(`R-squared: ${rSquared.dataSync()[0].toFixed(4)}`);

    // Clean up
    testX.dispose();
    predictions.dispose();
    yPred.dispose();
    yMean.dispose();
    totalSumSquares.dispose();
    residualSumSquares.dispose();
    rSquared.dispose();
}

// 5. Plot training progress (browser example)
function setupTrainingVisualization(model, data) {
    console.log('\nSetting up visualization (browser)...');

    // This would typically be used with a charting library like Chart.js
    // For demonstration, we'll just log the data
    const predictions = model.predict(data.x);

    console.log('Sample points for visualization:');
    for (let i = 0; i < Math.min(10, data.x.shape[0]); i++) {
        const x = data.x.dataSync()[i];
        const yActual = data.y.dataSync()[i];
        const yPred = predictions.dataSync()[i];

        console.log(`(${x.toFixed(2)}, ${yActual.toFixed(2)}) -> predicted: ${yPred.toFixed(2)}`);
    }

    predictions.dispose();
}

// 6. Advanced: Polynomial regression
function createPolynomialModel(degree = 2) {
    const model = tf.sequential();

    model.add(tf.layers.dense({
        units: 10,
        activation: 'relu',
        inputShape: [1]
    }));

    model.add(tf.layers.dense({
        units: 10,
        activation: 'relu'
    }));

    model.add(tf.layers.dense({
        units: 1
    }));

    model.compile({
        optimizer: tf.train.adam(0.01),
        loss: 'meanSquaredError'
    });

    return model;
}

// Main execution
async function main() {
    try {
        // Generate linear data
        const data = generateLinearData(100);
        console.log('Generated data shape:', data.x.shape, data.y.shape);

        // Create and train linear model
        const linearModel = createLinearModel();
        console.log('\nLinear Model Architecture:');
        linearModel.summary();

        await trainLinearRegression(linearModel, data);
        evaluateModel(linearModel, data);
        setupTrainingVisualization(linearModel, data);

        // Try polynomial regression for comparison
        console.log('\n=== Polynomial Regression Comparison ===');
        const polyModel = createPolynomialModel(2);
        console.log('Polynomial Model Architecture:');
        polyModel.summary();

        await trainLinearRegression(polyModel, data);
        console.log('\nPolynomial model evaluation:');
        evaluateModel(polyModel, data);

        // Clean up
        data.x.dispose();
        data.y.dispose();
        linearModel.dispose();
        polyModel.dispose();

    } catch (error) {
        console.error('Error:', error);
    }
}

// Run the example
main();