TypeScript deeplearn.tidy函数代码示例

export function predict(x: dl.Tensor2D): number[] {
  const pred = dl.tidy(() => {
    const axis = 1;
    return model(x).argMax(axis);
  });
  return Array.from(pred.dataSync());
}

  xhr.onload = async () => {
    const data = JSON.parse(xhr.responseText) as SampleData;

    // Wrap everything in a dl.tidy so we clean up intermediate Tensors.
    dl.tidy(async () => {
      console.log(`Evaluation set: n=${data.images.length}.`);

      let numCorrect = 0;
      for (let i = 0; i < data.images.length; i++) {
        const x = dl.tensor1d(data.images[i]);

        // Infer through the model to get a prediction.
        const predictedLabel = Math.round(await infer(x, vars).val());
        console.log(`Item ${i}, predicted label ${predictedLabel}.`);

        // Aggregate correctness to show accuracy.
        const label = data.labels[i];
        if (label === predictedLabel) {
          numCorrect++;
        }

        // Show the image.
        const result =
            renderResults(dl.tensor1d(data.images[i]), label, predictedLabel);
        document.body.appendChild(result);
      }

      // Compute final accuracy.
      const accuracy = numCorrect * 100 / data.images.length;
      document.getElementById('accuracy').innerHTML = `${accuracy}%`;
    });
  };

 async runInference() {
   await dl.tidy(async () => {
     const preprocessed = dl.fromPixels(this.contentImgElement);
     const inferenceResult = await this.transformNet.predict(preprocessed);
     this.setCanvasShape(inferenceResult.shape);
     renderToCanvas(inferenceResult, this.canvas);
   });
 }

  async run(size: number, option: string): Promise<number> {
    dl.setBackend('cpu');

    const input: dl.Tensor2D = dl.randomUniform([size, size], -1, 1);
    const op = getReductionOp(option);
    const start = performance.now();

    dl.tidy(() => {
      op(input).get();
    });

    const end = performance.now();
    return end - start;
  }

  async run(size: number, option: string): Promise<number> {
    const safeMode = false;
    const math = new dl.NDArrayMath('cpu', safeMode);
    dl.ENV.setMath(math);

    const input: dl.Tensor2D = dl.randomUniform([size, size], -1, 1);
    const op = getReductionOp(option);
    const start = performance.now();

    dl.tidy(() => {
      op(input).get();
    });

    math.dispose();

    const end = performance.now();
    return end - start;
  }

function getConditioning(): dl.Tensor1D {
  return dl.tidy(() => {
    if (!conditioned) {
      // TODO(nsthorat): figure out why we have to cast these shapes to numbers.
      // The linter is complaining, though VSCode can infer the types.
      const size = 1 + (noteDensityEncoding.shape[0] as number) +
          (pitchHistogramEncoding.shape[0] as number);
      const conditioning: dl.Tensor1D =
          dl.oneHot(dl.tensor1d([0]), size).as1D();
      return conditioning;
    } else {
      const axis = 0;
      const conditioningValues =
          noteDensityEncoding.concat(pitchHistogramEncoding, axis);
      return dl.tensor1d([0]).concat(conditioningValues, axis);
    }
  });
}

  /**
   * Infer through TransformNet, assumes variables have been loaded.
   * Original Tensorflow version of model can be found at
   * https://github.com/lengstrom/fast-style-transfer
   *
   * @param preprocessedInput preprocessed input Array.
   * @return dl.Tensor3D containing pixels of output img
   */
  predict(preprocessedInput: dl.Tensor3D): dl.Tensor3D {
    const img = dl.tidy(() => {
      const conv1 = this.convLayer(preprocessedInput.toFloat(), 1, true, 0);
      const conv2 = this.convLayer(conv1, 2, true, 3);
      const conv3 = this.convLayer(conv2, 2, true, 6);
      const resid1 = this.residualBlock(conv3, 9);
      const resid2 = this.residualBlock(resid1, 15);
      const resid3 = this.residualBlock(resid2, 21);
      const resid4 = this.residualBlock(resid3, 27);
      const resid5 = this.residualBlock(resid4, 33);
      const convT1 = this.convTransposeLayer(resid5, 64, 2, 39);
      const convT2 = this.convTransposeLayer(convT1, 32, 2, 42);
      const convT3 = this.convLayer(convT2, 1, false, 45);

      return convT3.tanh()
                 .mul(this.timesScalar)
                 .add(this.plusScalar)
                 .clip(0, 255)
                 .div(dl.scalar(255)) as dl.Tensor3D;
    });

    return img;
  }

// 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)}`);
}

export const learnXOR = async () => {
  const iterations = getRandomIntegerInRange(800, 1000);
  const timeStart: number = performance.now();
  let loss: number;
  let cost: dl.Scalar;

  const graph = new dl.Graph();

  const input = graph.placeholder('input', [2]);
  const y = graph.placeholder('y', [1]);

  const hiddenLayer = graph.layers.dense(
      'hiddenLayer', input, 10, (x: dl.SymbolicTensor) => graph.relu(x), true);
  const output = graph.layers.dense(
      'outputLayer', hiddenLayer, 1, (x: dl.SymbolicTensor) => graph.sigmoid(x),
      true);

  const costTensor = graph.reduceSum(graph.add(
      graph.multiply(
          graph.constant([-1]),
          graph.multiply(
              y, graph.log(graph.add(output, graph.constant([EPSILON]))))),
      graph.multiply(
          graph.constant([-1]),
          graph.multiply(
              graph.subtract(graph.constant([1]), y),
              graph.log(graph.add(
                  graph.subtract(graph.constant([1]), output),
                  graph.constant([EPSILON])))))));

  const session = new dl.Session(graph, dl.ENV.math);
  const optimizer = new dl.SGDOptimizer(0.2);

  const inputArray = [
    dl.tensor1d([0, 0]), dl.tensor1d([0, 1]), dl.tensor1d([1, 0]),
    dl.tensor1d([1, 1])
  ];

  const targetArray =
      [dl.tensor1d([0]), dl.tensor1d([1]), dl.tensor1d([1]), dl.tensor1d([0])];

  const shuffledInputProviderBuilder =
      new dl.InCPUMemoryShuffledInputProviderBuilder([inputArray, targetArray]);

  const [inputProvider, targetProvider] =
      shuffledInputProviderBuilder.getInputProviders();

  const feedEntries =
      [{tensor: input, data: inputProvider}, {tensor: y, data: targetProvider}];

  /**
   * Train the model
   */
  await dl.tidy(async () => {
    for (let i = 0; i < iterations; i += 1) {
      cost = session.train(
          costTensor, feedEntries, 4, optimizer, dl.CostReduction.MEAN);
    }
    loss = await cost.val();
  });

  const result = [];

  /**
   * Test the model
   */
  for (let i = 0; i < 4; i += 1) {
    const inputData = inputArray[i];
    const expectedOutput = targetArray[i];

    const val = session.eval(output, [{tensor: input, data: inputData}]);

    result.push({
      input: await inputData.data(),
      expected: await expectedOutput.data(),
      output: await val.data()
    });
  }

  const timeEnd: number = performance.now();
  const time = timeEnd - timeStart;

  return {iterations, loss, time, result};
};

async function generateStep(loopId: number) {
  if (loopId < currentLoopId) {
    // Was part of an outdated generateStep() scheduled via setTimeout.
    return;
  }
  await dl.tidy(async () => {
    const lstm1 = (data: dl.Tensor2D, c: dl.Tensor2D, h: dl.Tensor2D) =>
        dl.basicLSTMCell(forgetBias, lstmKernel1, lstmBias1, data, c, h);
    const lstm2 = (data: dl.Tensor2D, c: dl.Tensor2D, h: dl.Tensor2D) =>
        dl.basicLSTMCell(forgetBias, lstmKernel2, lstmBias2, data, c, h);
    const lstm3 = (data: dl.Tensor2D, c: dl.Tensor2D, h: dl.Tensor2D) =>
        dl.basicLSTMCell(forgetBias, lstmKernel3, lstmBias3, data, c, h);

    const outputs: dl.Scalar[] = [];
    // Generate some notes.
    for (let i = 0; i < STEPS_PER_GENERATE_CALL; i++) {
      // Use last sampled output as the next input.
      const eventInput = dl.oneHot(lastSample.as1D(), EVENT_SIZE).as1D();
      // Dispose the last sample from the previous generate call, since we
      // kept it.
      if (i === 0) {
        lastSample.dispose();
      }
      const conditioning = getConditioning();
      const axis = 0;
      const input = conditioning.concat(eventInput, axis);
      const output =
          dl.multiRNNCell([lstm1, lstm2, lstm3], input.as2D(1, -1), c, h);
      c = output[0];
      h = output[1];

      const outputH = h[2];
      const logits = outputH.matMul(fcW).add(fcB);

      const softmax = logits.as1D().softmax();
      const sampledOutput = dl.multinomial(softmax, 1).asScalar();

      outputs.push(sampledOutput);
      dl.keep(sampledOutput);
      lastSample = sampledOutput;
    }

    c.forEach(val => dl.keep(val));
    h.forEach(val => dl.keep(val));

    await outputs[outputs.length - 1].data();

    for (let i = 0; i < outputs.length; i++) {
      playOutput(await outputs[i].val());
    }

    if (piano.now() - currentPianoTimeSec > MAX_GENERATION_LAG_SECONDS) {
      console.warn(
          `Generation is ${
              piano.now() - currentPianoTimeSec} seconds behind, ` +
          `which is over ${MAX_NOTE_DURATION_SECONDS}. Resetting time!`);
      currentPianoTimeSec = piano.now();
    }
    const delta = Math.max(
        0, currentPianoTimeSec - piano.now() - GENERATION_BUFFER_SECONDS);
    setTimeout(() => generateStep(loopId), delta * 1000);
  });
}