TypeScript train.sgd方法代碼示例

 private createOptimizer() {
   switch (this.selectedOptimizerName) {
     case 'sgd': {
       return dl.train.sgd(this.learningRate);
     }
     case 'momentum': {
       return dl.train.momentum(+this.learningRate, +this.momentum);
     }
     case 'rmsprop': {
       return new dl.RMSPropOptimizer(+this.learningRate, +this.gamma);
     }
     case 'adagrad': {
       return new dl.AdagradOptimizer(+this.learningRate);
     }
     case 'adadelta': {
       return new dl.AdadeltaOptimizer(+this.learningRate, +this.gamma);
     }
     case 'adam': {
       return new dl.AdamOptimizer(
           +this.learningRate, +this.beta1, +this.beta2);
     }
     case 'adamax': {
       return new dl.AdamaxOptimizer(
           +this.learningRate, +this.beta1, +this.beta2);
     }
     default: {
       throw new Error(`Unknown optimizer "${this.selectedOptimizerName}"`);
     }
   }
 }

    await dl.tidy(async () => {
      /**
       * Inference
       */
      // Now we ask the dl.Graph to evaluate (infer) and give us the result when
      // providing a value 4 for "x".
      // NOTE: "a", "b", and "c" are randomly initialized, so this will give us
      // something random.
      let result = session.eval(y, [{tensor: x, data: dl.scalar(4)}]);
      console.log(await result.data());

      /**
       * Training
       */
      // Now let's learn the coefficients of this quadratic given some data.
      // To do this, we need to provide examples of x and y.
      // The values given here are for values a = 3, b = 2, c = 1, with random
      // noise added to the output so it's not a perfect fit.
      const xs = [dl.scalar(0), dl.scalar(1), dl.scalar(2), dl.scalar(3)];
      const ys =
          [dl.scalar(1.1), dl.scalar(5.9), dl.scalar(16.8), dl.scalar(33.9)];
      // When training, it's important to shuffle your data!
      const shuffledInputProviderBuilder =
          new dl.InCPUMemoryShuffledInputProviderBuilder([xs, ys]);
      const [xProvider, yProvider] =
          shuffledInputProviderBuilder.getInputProviders();

      // Training is broken up into batches.
      const NUM_BATCHES = 20;
      const BATCH_SIZE = xs.length;
      // Before we start training, we need to provide an optimizer. This is the
      // object that is responsible for updating weights. The learning rate
      // param is a value that represents how large of a step to make when
      // updating weights. If this is too big, you may overstep and oscillate.
      // If it is too small, the model may take a long time to train.
      const LEARNING_RATE = .01;
      const optimizer = dl.train.sgd(LEARNING_RATE);
      for (let i = 0; i < NUM_BATCHES; i++) {
        // Train takes a cost dl.Tensor to minimize; this call trains one batch
        // and returns the average cost of the batch as a dl.Scalar.
        const costValue = session.train(
            cost,
            // Map input providers to Tensors on the dl.Graph.
            [{tensor: x, data: xProvider}, {tensor: yLabel, data: yProvider}],
            BATCH_SIZE, optimizer, dl.CostReduction.MEAN);

        console.log(`average cost: ${await costValue.data()}`);
      }

      // Now print the value from the trained model for x = 4, should be ~57.0.
      result = session.eval(y, [{tensor: x, data: dl.scalar(4)}]);
      console.log('result should be ~57.0:');
      console.log(await result.data());
    });

// This file parallels (some of) the code in the introduction tutorial.

/**
 * 'Math with WebGL backend' section of tutorial
 */
async function intro() {
  const a = dl.tensor2d([1.0, 2.0, 3.0, 4.0], [2, 2]);
  const b = dl.tensor2d([0.0, 2.0, 4.0, 6.0], [2, 2]);

  const size = dl.scalar(a.size);

  // Non-blocking math calls.
  const average = a.sub(b).square().sum().div(size);

  console.log(`mean squared difference: ${await average.val()}`);

  /**
   * 'Graphs and Tensors' section of tutorial
   */

  const g = new dl.Graph();

  // Placeholders are input containers. This is the container for where we
  // will feed an input Tensor when we execute the graph.
  const inputShape = [3];
  const inputTensor = g.placeholder('input', inputShape);

  const labelShape = [1];
  const labelTensor = g.placeholder('label', labelShape);

  // Variables are containers that hold a value that can be updated from
  // training.
  // Here we initialize the multiplier variable randomly.
  const multiplier = g.variable('multiplier', dl.randomNormal([1, 3]));

  // Top level graph methods take Tensors and return Tensors.
  const outputTensor = g.matmul(multiplier, inputTensor);
  const costTensor = g.meanSquaredCost(labelTensor, outputTensor);

  // Tensors, like Tensors, have a shape attribute.
  console.log(outputTensor.shape);

  /**
   * 'dl.Session and dl.FeedEntry' section of the tutorial.
   */

  const learningRate = .00001;
  const batchSize = 3;

  const session = new dl.Session(g, dl.ENV.math);
  const optimizer = dl.train.sgd(learningRate);

  const inputs: dl.Tensor1D[] = [
    dl.tensor1d([1.0, 2.0, 3.0]), dl.tensor1d([10.0, 20.0, 30.0]),
    dl.tensor1d([100.0, 200.0, 300.0])
  ];

  const labels: dl.Tensor1D[] =
      [dl.tensor1d([4.0]), dl.tensor1d([40.0]), dl.tensor1d([400.0])];

  // Shuffles inputs and labels and keeps them mutually in sync.
  const shuffledInputProviderBuilder =
      new dl.InCPUMemoryShuffledInputProviderBuilder([inputs, labels]);
  const [inputProvider, labelProvider] =
      shuffledInputProviderBuilder.getInputProviders();

  // Maps tensors to InputProviders.
  const feedEntries: dl.FeedEntry[] = [
    {tensor: inputTensor, data: inputProvider},
    {tensor: labelTensor, data: labelProvider}
  ];

  const NUM_BATCHES = 10;
  for (let i = 0; i < NUM_BATCHES; i++) {
    // Wrap session.train in a scope so the cost gets cleaned up
    // automatically.
    await dl.tidy(async () => {
      // Train takes a cost tensor to minimize. Trains one batch. Returns the
      // average cost as a dl.Scalar.
      const cost = session.train(
          costTensor, feedEntries, batchSize, optimizer, dl.CostReduction.MEAN);

      console.log(`last average cost (${i}): ${await cost.val()}`);
    });
  }

  const testInput = dl.tensor1d([0.1, 0.2, 0.3]);

  // session.eval can take Tensors as input data.
  const testFeedEntries: dl.FeedEntry[] =
      [{tensor: inputTensor, data: testInput}];

  const testOutput = session.eval(outputTensor, testFeedEntries);

  console.log('---inference output---');
  console.log(`shape: ${testOutput.shape}`);
  console.log(`value: ${await testOutput.val(0)}`);
}

import * as dl from 'deeplearn';
import {MnistData} from './data';

// Hyperparameters.
const LEARNING_RATE = .05;
const BATCH_SIZE = 64;
const TRAIN_STEPS = 100;

// Data constants.
const IMAGE_SIZE = 784;
const LABELS_SIZE = 10;

const optimizer = dl.train.sgd(LEARNING_RATE);

// Set up the model and loss function.
const weights: dl.Tensor2D = dl.variable(
    dl.randomNormal([IMAGE_SIZE, LABELS_SIZE], 0, 1 / Math.sqrt(IMAGE_SIZE)));

const model = (xs: dl.Tensor2D) => xs.matMul(weights);

const loss = (labels: dl.Tensor2D, ys: dl.Tensor2D) =>
    dl.losses.softmaxCrossEntropy(labels, ys).mean() as dl.Scalar;

// Train the model.
export async function train(data: MnistData, log: (message: string) => void) {
  const returnCost = true;
  for (let i = 0; i < TRAIN_STEPS; i++) {
    const cost = optimizer.minimize(() => {
      const batch = data.nextTrainBatch(BATCH_SIZE);

      return loss(batch.labels, model(batch.xs));