assertThat(TWO_VALUES_STATS.mean()).isWithin(ALLOWED_ERROR).of(TWO_VALUES_MEAN);
    // For datasets of many double values created from an array, we test many combinations of finite
    // and non-finite values:
    for (ManyValues values : ALL_MANY_VALUES) {
      double mean = Stats.of(values.asArray()).mean();
      if (values.hasAnyNaN()) {
        assertWithMessage("mean of %s", values).that(mean).isNaN();

  /**
   * Helper class for testing with non-finite values. {@link #ALL_MANY_VALUES} gives a number of
   * instances with many combinations of finite and non-finite values. All have {@link
   * #MANY_VALUES_COUNT} values. If all the values are finite then the mean is {@link
   * #MANY_VALUES_MEAN} and the sum-of-squares-of-deltas is {@link
   * #MANY_VALUES_SUM_OF_SQUARES_OF_DELTAS}. The smallest and largest finite values are always

   *   computeInPlace;
   * - (except that, for non-finite values, we don't do all combinations exhaustively);
   * - percentiles with index and with indexes taking int-varargs, and with compute taking a
   *   double-collection and with computeInPlace.
   */

  private static final double ALLOWED_ERROR = 1.0e-10;

  /**
   * A {@link Correspondence} which accepts finite values within {@link #ALLOWED_ERROR} of each
   * other.
   */

   *   computeInPlace;
   * - (except that, for non-finite values, we don't do all combinations exhaustively);
   * - percentiles with index and with indexes taking int-varargs, and with compute taking a
   *   double-collection and with computeInPlace.
   */

  private static final double ALLOWED_ERROR = 1.0e-10;

  /**
   * A {@link Correspondence} which accepts finite values within {@link #ALLOWED_ERROR} of each
   * other.
   */

        .isWithin(ALLOWED_ERROR)
        .of(MANY_VALUES_MEAN);
    // For datasets of many double values created from an iterable, we test many combinations of
    // finite and non-finite values:
    for (ManyValues values : ALL_MANY_VALUES) {
      StatsAccumulator accumulator = new StatsAccumulator();
      StatsAccumulator accumulatorByAddAllStats = new StatsAccumulator();

  static long getSignificand(double d) {
    checkArgument(isFinite(d), "not a normal value");
    int exponent = getExponent(d);
    long bits = doubleToRawLongBits(d);
    bits &= SIGNIFICAND_MASK;
    return (exponent == MIN_EXPONENT - 1) ? bits << 1 : bits | IMPLICIT_BIT;
  }

  static boolean isFinite(double d) {
    return getExponent(d) <= MAX_EXPONENT;
  }

 * infinite value and itself, the infinite value is returned. If required to do a weighted average
 * between {@link Double#NEGATIVE_INFINITY NEGATIVE_INFINITY} and {@link Double#POSITIVE_INFINITY
 * POSITIVE_INFINITY}, {@link Double#NaN NaN} is returned (note that this will only happen if the
 * dataset contains no finite values).
 *
 * <h3>Performance</h3>
 *

    ANY(Timeout.values()),
    PAST(Timeout.MIN, Timeout.MINUS_SMALL, Timeout.ZERO),
    FUTURE(Timeout.SMALL, Timeout.MAX),
    SMALL(Timeout.SMALL),
    FINITE(Timeout.MIN, Timeout.MINUS_SMALL, Timeout.ZERO, Timeout.SMALL),
    INFINITE(Timeout.LARGE, Timeout.MAX);

    final ImmutableList<Timeout> timeouts;

    TimeoutsToUse(Timeout... timeouts) {
      this.timeouts = ImmutableList.copyOf(timeouts);
    }

 * that provide readers that are:
 *
 * <ul>
 *   <li><b>Finite:</b> Many operations, such as {@link #length()} and {@link #read()}, will either
 *       block indefinitely or fail if the source creates an infinite reader.
 *   <li><b>Non-destructive:</b> A <i>destructive</i> reader will consume or otherwise alter the

        and `tf.data.Dataset.with_options()` respectively.
    *   New `tf.data.Dataset.reduce()` API allows users to reduce a finite
        dataset to a single element using a user-provided reduce function.
    *   New `tf.data.Dataset.window()` API allows users to create finite windows
        of input dataset; when combined with the `tf.data.Dataset.reduce()` API,
        this allows users to implement customized batching.