Function timeInference

e2e/benchmarks/benchmark_util.js:246–286

* Executes `predict()` and times the inference process for `numRuns` rounds. * Then returns a promise that resolves with information about the inference * time: * - `times`: an array of inference time for each inference * - `averageTime`: the average time of all inferences * - `averageTimeExclF

(predict, numRuns = 1)

 * @param numRuns The number of rounds for `predict` to execute and time.
 */
async function timeInference(predict, numRuns = 1) {
  if (typeof predict !== 'function') {
    throw new Error(
        'The first parameter should be a function, while ' +
        `a(n) ${typeof predict} is found.`);
  }

  const times = [];
  for (let i = 0; i < numRuns; i++) {
    const start = performance.now();
    const res = await predict();
    // Prediction from tflite backend generates in the worker thread,
    // we don't post the result back to main thread to avoid unnecessary
    // overhead in transferring between worker and main thread.
    if (!isTflite()) {
      // The prediction can be tf.Tensor|tf.Tensor[]|{[name: string]:
      // tf.Tensor}.
      const value = await downloadValuesFromTensorContainer(res);
    }
    const elapsedTime = performance.now() - start;

    tf.dispose(res);
    times.push(elapsedTime);
  }

  const averageTime = times.reduce((acc, curr) => acc + curr, 0) / times.length;
  const averageTimeExclFirst = times.length > 1 ?
      times.slice(1).reduce((acc, curr) => acc + curr, 0) / (times.length - 1) :
      'NA';
  const minTime = Math.min(...times);
  const maxTime = Math.max(...times);
  const timeInfo = {
    times,
    averageTime,
    averageTimeExclFirst,
    minTime,
    maxTime

  };
  return timeInfo;
}

/**
 * Time one model inference with parallel compilation feature, based on the