function decodeBoxes()

in face-landmarks-detection/src/shared/calculators/tensors_to_detections.ts [138:229]
82 lines of code
6 McCabe index (conditional complexity)

function decodeBoxes(
    rawBoxes: tf.Tensor2D, anchor: AnchorTensor,
    config: TensorsToDetectionsConfig): tf.Tensor2D {
  return tf.tidy(() => {
    let yCenter;
    let xCenter;
    let h;
    let w;

    if (config.reverseOutputOrder) {
      // Shape [numOfBoxes, 1].
      xCenter = tf.squeeze(
          tf.slice(rawBoxes, [0, config.boxCoordOffset + 0], [-1, 1]));
      yCenter = tf.squeeze(
          tf.slice(rawBoxes, [0, config.boxCoordOffset + 1], [-1, 1]));
      w = tf.squeeze(
          tf.slice(rawBoxes, [0, config.boxCoordOffset + 2], [-1, 1]));
      h = tf.squeeze(
          tf.slice(rawBoxes, [0, config.boxCoordOffset + 3], [-1, 1]));
    } else {
      yCenter = tf.squeeze(
          tf.slice(rawBoxes, [0, config.boxCoordOffset + 0], [-1, 1]));
      xCenter = tf.squeeze(
          tf.slice(rawBoxes, [0, config.boxCoordOffset + 1], [-1, 1]));
      h = tf.squeeze(
          tf.slice(rawBoxes, [0, config.boxCoordOffset + 2], [-1, 1]));
      w = tf.squeeze(
          tf.slice(rawBoxes, [0, config.boxCoordOffset + 3], [-1, 1]));
    }

    xCenter =
        tf.add(tf.mul(tf.div(xCenter, config.xScale), anchor.w), anchor.x);
    yCenter =
        tf.add(tf.mul(tf.div(yCenter, config.yScale), anchor.h), anchor.y);

    if (config.applyExponentialOnBoxSize) {
      h = tf.mul(tf.exp(tf.div(h, config.hScale)), anchor.h);
      w = tf.mul(tf.exp(tf.div(w, config.wScale)), anchor.w);
    } else {
      h = tf.mul(tf.div(h, config.hScale), anchor.h);
      w = tf.mul(tf.div(w, config.wScale), anchor.h);
    }

    const yMin = tf.sub(yCenter, tf.div(h, 2));
    const xMin = tf.sub(xCenter, tf.div(w, 2));
    const yMax = tf.add(yCenter, tf.div(h, 2));
    const xMax = tf.add(xCenter, tf.div(w, 2));

    // Shape [numOfBoxes, 4].
    let boxes = tf.concat(
        [
          tf.reshape(yMin, [config.numBoxes, 1]),
          tf.reshape(xMin, [config.numBoxes, 1]),
          tf.reshape(yMax, [config.numBoxes, 1]),
          tf.reshape(xMax, [config.numBoxes, 1])
        ],
        1);

    if (config.numKeypoints) {
      for (let k = 0; k < config.numKeypoints; ++k) {
        const keypointOffset =
            config.keypointCoordOffset + k * config.numValuesPerKeypoint;
        let keypointX;
        let keypointY;
        if (config.reverseOutputOrder) {
          keypointX =
              tf.squeeze(tf.slice(rawBoxes, [0, keypointOffset], [-1, 1]));
          keypointY =
              tf.squeeze(tf.slice(rawBoxes, [0, keypointOffset + 1], [-1, 1]));
        } else {
          keypointY =
              tf.squeeze(tf.slice(rawBoxes, [0, keypointOffset], [-1, 1]));
          keypointX =
              tf.squeeze(tf.slice(rawBoxes, [0, keypointOffset + 1], [-1, 1]));
        }
        const keypointXNormalized = tf.add(
            tf.mul(tf.div(keypointX, config.xScale), anchor.w), anchor.x);
        const keypointYNormalized = tf.add(
            tf.mul(tf.div(keypointY, config.yScale), anchor.h), anchor.y);
        boxes = tf.concat(
            [
              boxes, tf.reshape(keypointXNormalized, [config.numBoxes, 1]),
              tf.reshape(keypointYNormalized, [config.numBoxes, 1])
            ],
            1);
      }
    }

    // Shape [numOfBoxes, 4] || [numOfBoxes, 12].
    return boxes as tf.Tensor2D;
  });
}