Method buckets

tfjs-node/src/nodejs_kernel_backend.ts:630–674

* Group data into histogram buckets. * * @param data A `Tensor` of any shape. Must be castable to `float32` * @param bucketCount Optional positive `number` * @returns A `Tensor` of shape `[k, 3]` and type `float32`. The `i`th row * is * a triple `[leftEdge, rightEdge, count]`

(data: Tensor, bucketCount?: number)

   * of `k` is either `bucketCount`, `1` or `0`.
   */
  private buckets(data: Tensor, bucketCount?: number): Tensor<tf.Rank> {
    if (data.size === 0) {
      return tf.tensor([], [0, 3], 'float32');
    }

    // 30 is the default number of buckets in the TensorFlow Python
    // implementation. See
    // https://github.com/tensorflow/tensorboard/blob/master/tensorboard/plugins/histogram/summary_v2.py
    bucketCount = bucketCount !== undefined ? bucketCount : 30;
    util.assert(
        Number.isInteger(bucketCount) && bucketCount > 0,
        () =>
            `Expected bucket count to be a strictly positive integer, but it was ` +
            `${bucketCount}`);
    data = data.flatten();
    data = data.cast('float32');
    const min: Scalar = data.min();
    const max: Scalar = data.max();
    const range: Scalar = max.sub(min);
    const isSingular = range.equal(0).arraySync() !== 0;

    if (isSingular) {
      const center = min;
      const bucketStart: Scalar = center.sub(0.5);
      const bucketEnd: Scalar = center.add(0.5);
      const bucketCounts = tf.scalar(data.size, 'float32');
      return tf.concat([bucketStart, bucketEnd, bucketCounts]).reshape([1, 3]);
    }

    const bucketWidth = range.div(bucketCount);
    const offsets = data.sub(min);
    const bucketIndices = offsets.floorDiv(bucketWidth).cast('int32');
    const clampedIndices =
        tf.minimum(bucketIndices, bucketCount - 1).cast('int32');
    const oneHots = tf.oneHot(clampedIndices, bucketCount);
    const bucketCounts = oneHots.sum(0).cast('int32');
    let edges = tf.linspace(min.arraySync(), max.arraySync(), bucketCount + 1);
    // Ensure last value in edges is max (TF's linspace op doesn't do this)
    edges = tf.concat([edges.slice(0, bucketCount), max.reshape([1])], 0) as
        tf.Tensor1D;
    const leftEdges = edges.slice(0, bucketCount);
    const rightEdges = edges.slice(1, bucketCount);
    return tf.stack([leftEdges, rightEdges, bucketCounts.cast('float32')])
        .transpose();
  }

  // ~ TensorBoard-related (tfjs-node-specific) backend kernels.
  // ------------------------------------------------------------