tensorflow_privacy/privacy/analysis/rdp_accountant.py [287:348]:
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
def _stable_inplace_diff_in_log(vec, signs, n=-1):
  """Replaces the first n-1 dims of vec with the log of abs difference operator.

  Args:
    vec: numpy array of floats with size larger than 'n'
    signs: Optional numpy array of bools with the same size as vec in case one
      needs to compute partial differences vec and signs jointly describe a
      vector of real numbers' sign and abs in log scale.
    n: Optonal upper bound on number of differences to compute. If negative, all
      differences are computed.

  Returns:
    The first n-1 dimension of vec and signs will store the log-abs and sign of
    the difference.

  Raises:
    ValueError: If input is malformed.
  """

  assert vec.shape == signs.shape
  if n < 0:
    n = np.max(vec.shape) - 1
  else:
    assert np.max(vec.shape) >= n + 1
  for j in range(0, n, 1):
    if signs[j] == signs[j + 1]:  # When the signs are the same
      # if the signs are both positive, then we can just use the standard one
      signs[j], vec[j] = _log_sub_sign(vec[j + 1], vec[j])
      # otherwise, we do that but toggle the sign
      if not signs[j + 1]:
        signs[j] = ~signs[j]
    else:  # When the signs are different.
      vec[j] = _log_add(vec[j], vec[j + 1])
      signs[j] = signs[j + 1]


def _get_forward_diffs(fun, n):
  """Computes up to nth order forward difference evaluated at 0.

  See Theorem 27 of https://arxiv.org/pdf/1808.00087.pdf

  Args:
    fun: Function to compute forward differences of.
    n: Number of differences to compute.

  Returns:
    Pair (deltas, signs_deltas) of the log deltas and their signs.
  """
  func_vec = np.zeros(n + 3)
  signs_func_vec = np.ones(n + 3, dtype=bool)

  # ith coordinate of deltas stores log(abs(ith order discrete derivative))
  deltas = np.zeros(n + 2)
  signs_deltas = np.zeros(n + 2, dtype=bool)
  for i in range(1, n + 3, 1):
    func_vec[i] = fun(1.0 * (i - 1))
  for i in range(0, n + 2, 1):
    # Diff in log scale
    _stable_inplace_diff_in_log(func_vec, signs_func_vec, n=n + 2 - i)
    deltas[i] = func_vec[0]
    signs_deltas[i] = signs_func_vec[0]
  return deltas, signs_deltas
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -



tensorflow_privacy/privacy/analysis/rdp_privacy_accountant.py [252:313]:
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
def _stable_inplace_diff_in_log(vec, signs, n=-1):
  """Replaces the first n-1 dims of vec with the log of abs difference operator.

  Args:
    vec: numpy array of floats with size larger than 'n'
    signs: Optional numpy array of bools with the same size as vec in case one
      needs to compute partial differences vec and signs jointly describe a
      vector of real numbers' sign and abs in log scale.
    n: Optonal upper bound on number of differences to compute. If negative, all
      differences are computed.

  Returns:
    The first n-1 dimension of vec and signs will store the log-abs and sign of
    the difference.

  Raises:
    ValueError: If input is malformed.
  """

  assert vec.shape == signs.shape
  if n < 0:
    n = np.max(vec.shape) - 1
  else:
    assert np.max(vec.shape) >= n + 1
  for j in range(0, n, 1):
    if signs[j] == signs[j + 1]:  # When the signs are the same
      # if the signs are both positive, then we can just use the standard one
      signs[j], vec[j] = _log_sub_sign(vec[j + 1], vec[j])
      # otherwise, we do that but toggle the sign
      if not signs[j + 1]:
        signs[j] = ~signs[j]
    else:  # When the signs are different.
      vec[j] = _log_add(vec[j], vec[j + 1])
      signs[j] = signs[j + 1]


def _get_forward_diffs(fun, n):
  """Computes up to nth order forward difference evaluated at 0.

  See Theorem 27 of https://arxiv.org/pdf/1808.00087.pdf

  Args:
    fun: Function to compute forward differences of.
    n: Number of differences to compute.

  Returns:
    Pair (deltas, signs_deltas) of the log deltas and their signs.
  """
  func_vec = np.zeros(n + 3)
  signs_func_vec = np.ones(n + 3, dtype=bool)

  # ith coordinate of deltas stores log(abs(ith order discrete derivative))
  deltas = np.zeros(n + 2)
  signs_deltas = np.zeros(n + 2, dtype=bool)
  for i in range(1, n + 3, 1):
    func_vec[i] = fun(1.0 * (i - 1))
  for i in range(0, n + 2, 1):
    # Diff in log scale
    _stable_inplace_diff_in_log(func_vec, signs_func_vec, n=n + 2 - i)
    deltas[i] = func_vec[0]
    signs_deltas[i] = signs_func_vec[0]
  return deltas, signs_deltas
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -