How to use @tensorflow/tfjs-node-gpu - 10 common examples

tf = require('@tensorflow/tfjs-node-gpu');
  } else {
    console.log('Using CPU');
    tf = require('@tensorflow/tfjs-node');
  }

  if (!fs.existsSync(path.dirname(args.generatorSavePath))) {
    fs.mkdirSync(path.dirname(args.generatorSavePath));
  }
  const saveURL = `file://${args.generatorSavePath}`;
  const metadataPath = path.join(args.generatorSavePath, 'acgan-metadata.json');

  // Build the discriminator.
  const discriminator = buildDiscriminator();
  discriminator.compile({
    optimizer: tf.train.adam(args.learningRate, args.adamBeta1),
    loss: ['binaryCrossentropy', 'sparseCategoricalCrossentropy']
  });
  discriminator.summary();

  // Build the generator.
  const generator = buildGenerator(args.latentSize);
  generator.summary();

  const optimizer = tf.train.adam(args.learningRate, args.adamBeta1);
  const combined = buildCombinedModel(
      args.latentSize, generator, discriminator, optimizer);

  await data.loadData();
  let {images: xTrain, labels: yTrain} = data.getTrainData();
  yTrain = tf.expandDims(yTrain.argMax(-1), -1);

const saveURL = `file://${args.generatorSavePath}`;
  const metadataPath = path.join(args.generatorSavePath, 'acgan-metadata.json');

  // Build the discriminator.
  const discriminator = buildDiscriminator();
  discriminator.compile({
    optimizer: tf.train.adam(args.learningRate, args.adamBeta1),
    loss: ['binaryCrossentropy', 'sparseCategoricalCrossentropy']
  });
  discriminator.summary();

  // Build the generator.
  const generator = buildGenerator(args.latentSize);
  generator.summary();

  const optimizer = tf.train.adam(args.learningRate, args.adamBeta1);
  const combined = buildCombinedModel(
      args.latentSize, generator, discriminator, optimizer);

  await data.loadData();
  let {images: xTrain, labels: yTrain} = data.getTrainData();
  yTrain = tf.expandDims(yTrain.argMax(-1), -1);

  // Save the generator model once before starting the training.
  await generator.save(saveURL);

  let numTensors;
  let logWriter;
  if (args.logDir) {
    console.log(`Logging to tensorboard at logdir: ${args.logDir}`);
    logWriter = tf.node.summaryFileWriter(args.logDir);
  }

const optimizer = tf.train.adam(args.learningRate, args.adamBeta1);
  const combined = buildCombinedModel(
      args.latentSize, generator, discriminator, optimizer);

  await data.loadData();
  let {images: xTrain, labels: yTrain} = data.getTrainData();
  yTrain = tf.expandDims(yTrain.argMax(-1), -1);

  // Save the generator model once before starting the training.
  await generator.save(saveURL);

  let numTensors;
  let logWriter;
  if (args.logDir) {
    console.log(`Logging to tensorboard at logdir: ${args.logDir}`);
    logWriter = tf.node.summaryFileWriter(args.logDir);
  }

  let step = 0;
  for (let epoch = 0; epoch < args.epochs; ++epoch) {
    // Write some metadata to disk at the beginning of every epoch.
    fs.writeFileSync(
        metadataPath,
        JSON.stringify(makeMetadata(args.epochs, epoch, false)));

    const tBatchBegin = tf.util.now();

    const numBatches = Math.ceil(xTrain.shape[0] / args.batchSize);

    for (let batch = 0; batch < numBatches; ++batch) {
      const actualBatchSize = (batch + 1) * args.batchSize >= xTrain.shape[0] ?
          (xTrain.shape[0] - batch * args.batchSize) :

let numTensors;
  let logWriter;
  if (args.logDir) {
    console.log(`Logging to tensorboard at logdir: ${args.logDir}`);
    logWriter = tf.node.summaryFileWriter(args.logDir);
  }

  let step = 0;
  for (let epoch = 0; epoch < args.epochs; ++epoch) {
    // Write some metadata to disk at the beginning of every epoch.
    fs.writeFileSync(
        metadataPath,
        JSON.stringify(makeMetadata(args.epochs, epoch, false)));

    const tBatchBegin = tf.util.now();

    const numBatches = Math.ceil(xTrain.shape[0] / args.batchSize);

    for (let batch = 0; batch < numBatches; ++batch) {
      const actualBatchSize = (batch + 1) * args.batchSize >= xTrain.shape[0] ?
          (xTrain.shape[0] - batch * args.batchSize) :
          args.batchSize;

      const dLoss = await trainDiscriminatorOneStep(
          xTrain, yTrain, batch * args.batchSize, actualBatchSize,
          args.latentSize, generator, discriminator);

      // Here we use 2 * actualBatchSize here, so that we have
      // the generator optimizer over an identical number of images
      // as the discriminator.
      const gLoss = await trainCombinedModelOneStep(

optimizer: tf.train.adam(args.learningRate, args.adamBeta1),
    loss: ['binaryCrossentropy', 'sparseCategoricalCrossentropy']
  });
  discriminator.summary();

  // Build the generator.
  const generator = buildGenerator(args.latentSize);
  generator.summary();

  const optimizer = tf.train.adam(args.learningRate, args.adamBeta1);
  const combined = buildCombinedModel(
      args.latentSize, generator, discriminator, optimizer);

  await data.loadData();
  let {images: xTrain, labels: yTrain} = data.getTrainData();
  yTrain = tf.expandDims(yTrain.argMax(-1), -1);

  // Save the generator model once before starting the training.
  await generator.save(saveURL);

  let numTensors;
  let logWriter;
  if (args.logDir) {
    console.log(`Logging to tensorboard at logdir: ${args.logDir}`);
    logWriter = tf.node.summaryFileWriter(args.logDir);
  }

  let step = 0;
  for (let epoch = 0; epoch < args.epochs; ++epoch) {
    // Write some metadata to disk at the beginning of every epoch.
    fs.writeFileSync(
        metadataPath,

`dLoss = ${dLoss[0].toFixed(6)}, gLoss = ${gLoss[0].toFixed(6)}`);
      if (logWriter != null) {
        logWriter.scalar('dLoss', dLoss[0], step);
        logWriter.scalar('gLoss', gLoss[0], step);
        step++;
      }

      // Assert on no memory leak.
      // TODO(cais): Remove this check once the current memory leak in
      //   tfjs-node and tfjs-node-gpu is fixed.
      if (numTensors == null) {
        numTensors = tf.memory().numTensors;
      } else {
        tf.util.assert(
            tf.memory().numTensors === numTensors,
            `Leaked ${tf.memory().numTensors - numTensors} tensors`);
      }
    }

    await generator.save(saveURL);
    console.log(
        `epoch ${epoch + 1} elapsed time: ` +
        `${((tf.util.now() - tBatchBegin) / 1e3).toFixed(1)} s`);
    console.log(`Saved generator model to: ${saveURL}\n`);
  }

  // Write metadata to disk to indicate the end of the training.
  fs.writeFileSync(
      metadataPath,
      JSON.stringify(makeMetadata(args.epochs, args.epochs, true)));
}

console.log(
          `epoch ${epoch + 1}/${args.epochs} batch ${batch + 1}/${
              numBatches}: ` +
          `dLoss = ${dLoss[0].toFixed(6)}, gLoss = ${gLoss[0].toFixed(6)}`);
      if (logWriter != null) {
        logWriter.scalar('dLoss', dLoss[0], step);
        logWriter.scalar('gLoss', gLoss[0], step);
        step++;
      }

      // Assert on no memory leak.
      // TODO(cais): Remove this check once the current memory leak in
      //   tfjs-node and tfjs-node-gpu is fixed.
      if (numTensors == null) {
        numTensors = tf.memory().numTensors;
      } else {
        tf.util.assert(
            tf.memory().numTensors === numTensors,
            `Leaked ${tf.memory().numTensors - numTensors} tensors`);
      }
    }

    await generator.save(saveURL);
    console.log(
        `epoch ${epoch + 1} elapsed time: ` +
        `${((tf.util.now() - tBatchBegin) / 1e3).toFixed(1)} s`);
    console.log(`Saved generator model to: ${saveURL}\n`);
  }

  // Write metadata to disk to indicate the end of the training.
  fs.writeFileSync(

kernelSize: 5,
    strides: 2,
    padding: 'same',
    activation: 'tanh',
    kernelInitializer: 'glorotNormal'
  }));

  // Unlike most TensorFlow.js models, the generator part of an ACGAN has
  // two inputs:
  //   1. The latent vector that is used as the "seed" of the fake image
  //      generation.
  //   2. A class label that controls which of the ten MNIST digit classes
  //      the generated fake image is meant to belong to.

  // This is the z space commonly referred to in GAN papers.
  const latent = tf.input({shape: [latentSize]});

  // The desired label of the generated image, an integer in the interval
  // [0, NUM_CLASSES).
  const imageClass = tf.input({shape: [1]});

  // The desired label is converted to a vector of length `latentSize`
  // through embedding lookup.
  const classEmbedding = tf.layers.embedding({
    inputDim: NUM_CLASSES,
    outputDim: latentSize,
    embeddingsInitializer: 'glorotNormal'
  }).apply(imageClass);

  // Hadamard product between z-space and a class conditional embedding.
  const h = tf.layers.multiply().apply([latent, classEmbedding]);

kernelInitializer: 'glorotNormal'
  }));

  // Unlike most TensorFlow.js models, the generator part of an ACGAN has
  // two inputs:
  //   1. The latent vector that is used as the "seed" of the fake image
  //      generation.
  //   2. A class label that controls which of the ten MNIST digit classes
  //      the generated fake image is meant to belong to.

  // This is the z space commonly referred to in GAN papers.
  const latent = tf.input({shape: [latentSize]});

  // The desired label of the generated image, an integer in the interval
  // [0, NUM_CLASSES).
  const imageClass = tf.input({shape: [1]});

  // The desired label is converted to a vector of length `latentSize`
  // through embedding lookup.
  const classEmbedding = tf.layers.embedding({
    inputDim: NUM_CLASSES,
    outputDim: latentSize,
    embeddingsInitializer: 'glorotNormal'
  }).apply(imageClass);

  // Hadamard product between z-space and a class conditional embedding.
  const h = tf.layers.multiply().apply([latent, classEmbedding]);

  const fakeImage = cnn.apply(h);
  return tf.model({inputs: [latent, imageClass], outputs: fakeImage});
}

cnn.add(tf.layers.leakyReLU({alpha: 0.2}));
  cnn.add(tf.layers.dropout({rate: 0.3}));

  cnn.add(tf.layers.conv2d(
      {filters: 128, kernelSize: 3, padding: 'same', strides: 2}));
  cnn.add(tf.layers.leakyReLU({alpha: 0.2}));
  cnn.add(tf.layers.dropout({rate: 0.3}));

  cnn.add(tf.layers.conv2d(
      {filters: 256, kernelSize: 3, padding: 'same', strides: 1}));
  cnn.add(tf.layers.leakyReLU({alpha: 0.2}));
  cnn.add(tf.layers.dropout({rate: 0.3}));

  cnn.add(tf.layers.flatten());

  const image = tf.input({shape: [IMAGE_SIZE, IMAGE_SIZE, 1]});
  const features = cnn.apply(image);

  // Unlike most TensorFlow.js models, the discriminator has two outputs.

  // The 1st output is the probability score assigned by the discriminator to
  // how likely the input example is a real MNIST image (as versus
  // a "fake" one generated by the generator).
  const realnessScore =
      tf.layers.dense({units: 1, activation: 'sigmoid'}).apply(features);
  // The 2nd output is the softmax probabilities assign by the discriminator
  // for the 10 MNIST digit classes (0 through 9). "aux" stands for "auxiliary"
  // (the namesake of ACGAN) and refers to the fact that unlike a standard GAN
  // (which performs just binary real/fake classification), the discriminator
  // part of ACGAN also performs multi-class classification.
  const aux = tf.layers.dense({units: NUM_CLASSES, activation: 'softmax'})
                  .apply(features);