def _normalize_array()

in lucid/misc/io/serialize_array.py [0:0]

• 27 lines of code
• 13 McCabe index (conditional complexity)

def _normalize_array(array, domain=(0, 1)):
  """Given an arbitrary rank-3 NumPy array, produce one representing an image.

  This ensures the resulting array has a dtype of uint8 and a domain of 0-255.

  Args:
    array: NumPy array representing the image
    domain: expected range of values in array,
      defaults to (0, 1), if explicitly set to None will use the array's
      own range of values and normalize them.

  Returns:
    normalized PIL.Image
  """
  # first copy the input so we're never mutating the user's data
  array = np.array(array)
  # squeeze helps both with batch=1 and B/W and PIL's mode inference
  array = np.squeeze(array)
  assert len(array.shape) <= 3
  assert np.issubdtype(array.dtype, np.number)
  assert not np.isnan(array).any()

  low, high = np.min(array), np.max(array)
  if domain is None:
    message = "No domain specified, normalizing from measured (~%.2f, ~%.2f)"
    log.debug(message, low, high)
    domain = (low, high)

  # clip values if domain was specified and array contains values outside of it
  if low < domain[0] or high > domain[1]:
    message = "Clipping domain from (~{:.2f}, ~{:.2f}) to (~{:.2f}, ~{:.2f})."
    log.info(message.format(low, high, domain[0], domain[1]))
    array = array.clip(*domain)

  min_value, max_value = np.iinfo(np.uint8).min, np.iinfo(np.uint8).max  # 0, 255
  # convert signed to unsigned if needed
  if np.issubdtype(array.dtype, np.inexact):
    offset = domain[0]
    if offset != 0:
      array -= offset
      log.debug("Converting inexact array by subtracting -%.2f.", offset)
    if domain[0] != domain[1]:
      scalar = max_value / (domain[1] - domain[0])
      if scalar != 1:
        array *= scalar
        log.debug("Converting inexact array by scaling by %.2f.", scalar)

  return array.clip(min_value, max_value).astype(np.uint8)