Python math_ops.reciprocal函数代码示例

 def _grad_and_hess_for_logloss(logits, labels):
   # TODO(youngheek): add weights handling.
   predictions = math_ops.reciprocal(math_ops.exp(-logits) + 1.0)
   normalizer = math_ops.reciprocal(
       math_ops.cast(array_ops.size(predictions), dtypes.float32))
   gradients = (predictions - labels) * normalizer
   hessians = predictions * (1.0 - predictions) * normalizer
   return gradients, hessians

 def _grad_and_hess_for_logloss(logits, labels):
   """A closed form gradient and hessian for logistic loss."""
   # TODO(youngheek): add weights handling.
   predictions = math_ops.reciprocal(math_ops.exp(-logits) + 1.0)
   normalizer = math_ops.reciprocal(
       math_ops.cast(array_ops.size(predictions), dtypes.float32))
   labels = math_ops.cast(labels, dtypes.float32)
   labels = head_lib._check_dense_labels_match_logits_and_reshape(  # pylint: disable=protected-access
       labels, logits, head.logits_dimension)
   gradients = (predictions - labels) * normalizer
   hessians = predictions * (1.0 - predictions) * normalizer
   return gradients, hessians

def normalize_moments(counts, mean_ss, variance_ss, shift, name=None):
  """Calculate the mean and variance of based on the sufficient statistics.

  Args:
    counts: A `Tensor` containing a the total count of the data (one value).
    mean_ss: A `Tensor` containing the mean sufficient statistics: the (possibly
      shifted) sum of the elements to average over.
    variance_ss: A `Tensor` containing the variance sufficient statistics: the
      (possibly shifted) squared sum of the data to compute the variance over.
    shift: A `Tensor` containing the value by which the data is shifted for
      numerical stability, or `None` if no shift was performed.
    name: Name used to scope the operations that compute the moments.

  Returns:
    Two `Tensor` objects: `mean` and `variance`.
  """
  with ops.name_scope(name, "normalize", [counts, mean_ss, variance_ss, shift]):
    divisor = math_ops.reciprocal(counts, name="divisor")
    if shift is not None:
      shifted_mean = math_ops.mul(mean_ss, divisor, name="shifted_mean")
      mean = math_ops.add(shifted_mean, shift, name="mean")
    else:  # no shift.
      shifted_mean = math_ops.mul(mean_ss, divisor, name="mean")
      mean = shifted_mean
    variance = math_ops.sub(math_ops.mul(variance_ss, divisor),
                            math_ops.square(shifted_mean),
                            name="variance")
  return (mean, variance)

def _TanGrad(op, grad):
  """Returns grad * 1/sec^2(x)."""
  x = op.inputs[0]
  with ops.control_dependencies([grad]):
    x = math_ops.conj(x)
    secx = math_ops.reciprocal(math_ops.cos(x))
    secx2 = math_ops.square(secx)
    return grad * secx2

def _AtanGrad(op, grad):
  """Returns grad * 1/ (1 + x^2)."""
  x = op.inputs[0]
  with ops.control_dependencies([grad]):
    x = math_ops.conj(x)
    x2 = math_ops.square(x)
    one = constant_op.constant(1, dtype=grad.dtype)
    inv = math_ops.reciprocal(math_ops.add(one, x2))
    return grad * inv

def _AngleGrad(op, grad):
  """Returns -grad / (Im(x) + iRe(x))"""
  x = op.inputs[0]
  with ops.control_dependencies([grad]):
    re = math_ops.real(x)
    im = math_ops.imag(x)
    z = math_ops.reciprocal(math_ops.complex(im, re))
    zero = constant_op.constant(0, dtype=grad.dtype)
    complex_grad = math_ops.complex(grad, zero)
    return -complex_grad * z

def _AcosGrad(op, grad):
  """Returns grad * -1/sqrt(1-x^2)."""
  x = op.inputs[0]
  with ops.control_dependencies([grad]):
    x = math_ops.conj(x)
    x2 = math_ops.square(x)
    one = constant_op.constant(1, dtype=grad.dtype)
    den = math_ops.sqrt(math_ops.subtract(one, x2))
    inv = math_ops.reciprocal(den)
    return -grad * inv

def _BesselI1eGrad(op, grad):
  """Compute gradient of bessel_i1e(x) with respect to its argument."""
  x = op.inputs[0]
  y = op.outputs[0]
  with ops.control_dependencies([grad]):
    # For x = 0, the correct gradient is 0.5.
    # However, the main branch gives NaN because of the division by x, so
    # we impute the gradient manually.
    # An alternative solution is to express the gradient via bessel_i0e and
    # bessel_i2e, but the latter is not yet implemented in Eigen.
    eps = np.finfo(x.dtype.as_numpy_dtype).eps
    zeros = array_ops.zeros_like(x)
    x_is_not_tiny = math_ops.abs(x) > eps
    safe_x = array_ops.where(x_is_not_tiny, x, eps + zeros)
    dy_dx = math_ops.bessel_i0e(safe_x) - y * (
        math_ops.sign(safe_x) + math_ops.reciprocal(safe_x))
    return grad * array_ops.where(x_is_not_tiny, dy_dx, 0.5 + zeros)

def _BatchNormGrad(grad_y, x, scale, epsilon, data_format):
  """Returns the gradients for the 3 inputs of BatchNorm.

  Args:
    grad_y: A `Tensor` of 4 dimensions for gradient for y.
    x: A `Tensor` of 4 dimensions for x.
    scale: A `Tensor` of 1 dimension for scaling.
    epsilon: A small float number added to the variance of x.
    data_format: The data format for input. Either b"NHWC" or b"NCHW".

  Returns:
    A tuple (grad_x, grad_scale, grad_offset), where grad_x is the gradient
    for x, grad_scale the gradient for scale, and grad_offset the gradient
    for offset.
  """
  if data_format == b"NHWC":
    keep_dims = False
    reduce_axis = [0, 1, 2]
  else:
    keep_dims = True
    reduce_axis = [0, 2, 3]
    shape = [1, array_ops.size(scale), 1, 1]
    scale = array_ops.reshape(scale, shape)
  mean_grad_y = math_ops.reduce_mean(grad_y, reduce_axis, keep_dims=keep_dims)
  mean_x = math_ops.reduce_mean(x, reduce_axis, keep_dims=keep_dims)
  var_x = math_ops.reduce_mean(
      math_ops.squared_difference(x, array_ops.stop_gradient(mean_x)),
      reduce_axis,
      keep_dims=keep_dims)
  grad_y_offset = grad_y - mean_grad_y
  x_offset = x - mean_x
  mean = math_ops.reduce_mean(
      grad_y * x_offset, axis=reduce_axis, keep_dims=keep_dims)
  grad_x = scale * math_ops.rsqrt(var_x + epsilon) * (
      grad_y_offset - math_ops.reciprocal(var_x + epsilon) * mean * x_offset)
  grad_scale = math_ops.rsqrt(var_x + epsilon) * math_ops.reduce_sum(
      grad_y * x_offset, axis=reduce_axis, keep_dims=keep_dims)
  if data_format == b"NCHW":
    grad_scale = array_ops.squeeze(grad_scale)
  grad_offset = math_ops.reduce_sum(grad_y, axis=reduce_axis)
  return grad_x, grad_scale, grad_offset

  def _Grad(op, grad):
    """A gradient function for IRFFT with the provided `rank` and `rfft_fn`."""
    # Generate a simple mask like [1.0, 2.0, ..., 2.0, 1.0] for even-length FFTs
    # and [1.0, 2.0, ..., 2.0] for odd-length FFTs. To reduce extra ops in the
    # graph we special-case the situation where the FFT length and last
    # dimension of the input are known at graph construction time.
    fft_length = op.inputs[1]
    is_odd = math_ops.mod(fft_length[-1], 2)
    input_last_dimension = array_ops.shape(op.inputs[0])[-1]
    mask = array_ops.concat(
        [[1.0], 2.0 * array_ops.ones([input_last_dimension - 2 + is_odd]),
         array_ops.ones([1 - is_odd])], 0)

    rsize = math_ops.reciprocal(math_ops.to_float(_FFTSizeForGrad(grad, rank)))

    # The gradient of IRFFT is the RFFT of the incoming gradient times a scaling
    # factor and a mask. The mask scales the gradient for the Hermitian
    # symmetric components of the RFFT by a factor of two, since these
    # components are de-duplicated in the RFFT.
    rfft = rfft_fn(grad, fft_length)
    return rfft * math_ops.cast(rsize * mask, dtypes.complex64), None

def _SelfAdjointEigV2Grad(op, grad_e, grad_v):
  """Gradient for SelfAdjointEigV2."""
  e = op.outputs[0]
  compute_v = op.get_attr("compute_v")
  # a = op.inputs[0], which satisfies
  # a[...,:,:] * v[...,:,i] = e[...,i] * v[...,i]
  with ops.control_dependencies([grad_e, grad_v]):
    if compute_v:
      v = op.outputs[1]
      # Construct the matrix f(i,j) = (i != j ? 1 / (e_i - e_j) : 0).
      # Notice that because of the term involving f, the gradient becomes
      # infinite (or NaN in practice) when eigenvalues are not unique.
      # Mathematically this should not be surprising, since for (k-fold)
      # degenerate eigenvalues, the corresponding eigenvectors are only defined
      # up to arbitrary rotation in a (k-dimensional) subspace.
      f = array_ops.matrix_set_diag(
          math_ops.reciprocal(
              array_ops.expand_dims(e, -2) - array_ops.expand_dims(e, -1)),
          array_ops.zeros_like(e))
      grad_a = math_ops.matmul(
          v,
          math_ops.matmul(
              array_ops.matrix_diag(grad_e) +
              f * math_ops.matmul(v, grad_v, adjoint_a=True),
              v,
              adjoint_b=True))
    else:
      _, v = linalg_ops.self_adjoint_eig(op.inputs[0])
      grad_a = math_ops.matmul(v,
                               math_ops.matmul(
                                   array_ops.matrix_diag(grad_e),
                                   v,
                                   adjoint_b=True))
    # The forward op only depends on the lower triangular part of a, so here we
    # symmetrize and take the lower triangle
    grad_a = array_ops.matrix_band_part(
        grad_a + math_ops.conj(array_ops.matrix_transpose(grad_a)), -1, 0)
    grad_a = array_ops.matrix_set_diag(grad_a,
                                       0.5 * array_ops.matrix_diag_part(grad_a))
    return grad_a

  def _mini_batch_training_op(self, inputs, cluster_idx_list, cluster_centers,
                              cluster_centers_var, total_counts):
    """Creates an op for training for mini batch case.

    Args:
      inputs: list of input Tensors.
      cluster_idx_list: A vector (or list of vectors). Each element in the
        vector corresponds to an input row in 'inp' and specifies the cluster id
        corresponding to the input.
      cluster_centers: Tensor of cluster centers, possibly normalized.
      cluster_centers_var: Tensor Ref of cluster centers.
      total_counts: Tensor Ref of cluster counts.

    Returns:
      An op for doing an update of mini-batch k-means.
    """
    update_ops = []
    for inp, cluster_idx in zip(inputs, cluster_idx_list):
      with ops.colocate_with(inp):
        assert total_counts is not None
        cluster_idx = array_ops.reshape(cluster_idx, [-1])
        # Dedupe the unique ids of cluster_centers being updated so that updates
        # can be locally aggregated.
        unique_ids, unique_idx = array_ops.unique(cluster_idx)
        num_unique_cluster_idx = array_ops.size(unique_ids)
        # Fetch the old values of counts and cluster_centers.
        with ops.colocate_with(total_counts):
          old_counts = array_ops.gather(total_counts, unique_ids)
        with ops.colocate_with(cluster_centers):
          old_cluster_centers = array_ops.gather(cluster_centers, unique_ids)
        # Locally aggregate the increment to counts.
        count_updates = math_ops.unsorted_segment_sum(
            array_ops.ones_like(
                unique_idx, dtype=total_counts.dtype),
            unique_idx,
            num_unique_cluster_idx)
        # Locally compute the sum of inputs mapped to each id.
        # For a cluster with old cluster value x, old count n, and with data
        # d_1,...d_k newly assigned to it, we recompute the new value as
        # x += (sum_i(d_i) - k * x) / (n + k).
        # Compute sum_i(d_i), see comment above.
        cluster_center_updates = math_ops.unsorted_segment_sum(
            inp, unique_idx, num_unique_cluster_idx)
        # Shape to enable broadcasting count_updates and learning_rate to inp.
        # It extends the shape with 1's to match the rank of inp.
        broadcast_shape = array_ops.concat(
            [
                array_ops.reshape(num_unique_cluster_idx, [1]), array_ops.ones(
                    array_ops.reshape(array_ops.rank(inp) - 1, [1]),
                    dtype=dtypes.int32)
            ],
            0)
        # Subtract k * x, see comment above.
        cluster_center_updates -= math_ops.cast(
            array_ops.reshape(count_updates, broadcast_shape),
            inp.dtype) * old_cluster_centers
        learning_rate = math_ops.reciprocal(
            math_ops.cast(old_counts + count_updates, inp.dtype))
        learning_rate = array_ops.reshape(learning_rate, broadcast_shape)
        # scale by 1 / (n + k), see comment above.
        cluster_center_updates *= learning_rate
        # Apply the updates.
      update_counts = state_ops.scatter_add(total_counts, unique_ids,
                                            count_updates)
      update_cluster_centers = state_ops.scatter_add(cluster_centers_var,
                                                     unique_ids,
                                                     cluster_center_updates)
      update_ops.extend([update_counts, update_cluster_centers])
    return control_flow_ops.group(*update_ops)

def _SvdGrad(op, grad_s, grad_u, grad_v):
  """Gradient for the singular value decomposition."""

  # The derivation for the compute_uv=False case, and most of
  # the derivation for the full_matrices=True case, are in
  # Giles' paper (see reference at top of file).  A derivation for
  # the full_matrices=False case is available at
  # https://j-towns.github.io/papers/svd-derivative.pdf
  a = op.inputs[0]
  a_shape = a.get_shape().with_rank_at_least(2)
  grad_s_mat = array_ops.matrix_diag(grad_s)

  if not op.get_attr("compute_uv"):
    s, u, v = linalg_ops.svd(a, compute_uv=True)
    grad_a = math_ops.matmul(u, math_ops.matmul(grad_s_mat, v, adjoint_b=True))
    grad_a.set_shape(a_shape)
    return grad_a

  full_matrices = op.get_attr("full_matrices")

  # TODO(rmlarsen): Make this work with complex types.
  if a.dtype.is_complex:
    raise NotImplementedError(
        "SVD gradient is not implemented for complex types and "
        "compute_uv=True.")
  grad_u_shape = grad_u.get_shape().with_rank_at_least(2)
  grad_v_shape = grad_v.get_shape().with_rank_at_least(2)
  m = a_shape.dims[-2].merge_with(grad_u_shape[-2])
  n = a_shape.dims[-1].merge_with(grad_v_shape[-2])
  batch_shape = a_shape[:-2].merge_with(grad_u_shape[:-2]).merge_with(
      grad_v_shape[:-2])
  a_shape = batch_shape.concatenate([m, n])

  m = a_shape.dims[-2].value
  n = a_shape.dims[-1].value
  # TODO(rmlarsen): Make this work with placeholders.
  if m is None or n is None:
    raise NotImplementedError(
        "SVD gradient has not been implemented for input with unknown "
        "inner matrix shape.")

  s = op.outputs[0]
  u = op.outputs[1]
  v = op.outputs[2]

  use_adjoint = False
  if m > n:
    # Compute the gradient for A^H = V * S^T * U^H, and (implicitly) take the
    # Hermitian transpose of the gradient at the end.
    use_adjoint = True
    m, n = n, m
    u, v = v, u
    grad_u, grad_v = grad_v, grad_u

  with ops.control_dependencies([grad_s, grad_u, grad_v]):
    if full_matrices and abs(m - n) > 1:
      raise NotImplementedError(
          "svd gradient is not implemented for abs(m - n) > 1 "
          "when full_matrices is True")
    s_mat = array_ops.matrix_diag(s)
    s2 = math_ops.square(s)

    # NOTICE: Because of the term involving f, the gradient becomes
    # infinite (or NaN in practice) when singular values are not unique.
    # Mathematically this should not be surprising, since for (k-fold)
    # degenerate singular values, the corresponding singular vectors are
    # only defined up a (k-dimensional) subspace. In practice, this can
    # lead to numerical instability when singular values are close but not
    # exactly equal.
    f = array_ops.matrix_set_diag(
        math_ops.reciprocal(
            array_ops.expand_dims(s2, -2) - array_ops.expand_dims(s2, -1)),
        array_ops.zeros_like(s))
    s_inv_mat = array_ops.matrix_diag(math_ops.reciprocal(s))

    v1 = v[..., :, :m]
    grad_v1 = grad_v[..., :, :m]

    u_gu = math_ops.matmul(u, grad_u, adjoint_a=True)
    v_gv = math_ops.matmul(v1, grad_v1, adjoint_a=True)

    f_u = f * u_gu
    f_v = f * v_gv

    term1_nouv = (
        grad_s_mat + math_ops.matmul(f_u + _linalg.adjoint(f_u), s_mat) +
        math_ops.matmul(s_mat, f_v + _linalg.adjoint(f_v)))

    term1 = math_ops.matmul(u, math_ops.matmul(term1_nouv, v1, adjoint_b=True))

    if m == n:
      grad_a_before_transpose = term1
    else:
      gv1t = array_ops.matrix_transpose(grad_v1)
      gv1t_v1 = math_ops.matmul(gv1t, v1)
      term2_nous = gv1t - math_ops.matmul(gv1t_v1, v1, adjoint_b=True)

      if full_matrices:
        v2 = v[..., :, m:n]
        grad_v2 = grad_v[..., :, m:n]
#.........这里部分代码省略.........

def _Log1pGrad(op, grad):
  """Returns grad * (1/(1 + x))."""
  x = op.inputs[0]
  with ops.control_dependencies([grad]):
    x = math_ops.conj(x)
    return grad * math_ops.reciprocal(1 + x)

def _BatchNormGrad(grad_y,
                   x,
                   scale,
                   pop_mean,
                   pop_var,
                   epsilon,
                   data_format,
                   is_training=True):
  """Returns the gradients for the 3 inputs of BatchNorm.

  Args:
    grad_y: A `Tensor` of 4 dimensions for gradient for y.
    x: A `Tensor` of 4 dimensions for x.
    scale: A `Tensor` of 1 dimension for scaling.
    pop_mean: A `Tensor` of 1 dimension for the population mean. Only used when
      is_training=False.
    pop_var: A `Tensor` of 1 dimension for the population variance. Only used
      when is_training=False.
    epsilon: A small float number added to the variance of x.
    data_format: The data format for input. Either b"NHWC" or b"NCHW".
    is_training: A bool value to indicate the operation is for training
      (default) or inference.

  Returns:
    A tuple (grad_x, grad_scale, grad_offset), where grad_x is the gradient
    for x, grad_scale the gradient for scale, and grad_offset the gradient
    for offset.
  """
  x_dtype = x.dtype.base_dtype
  if x_dtype == dtypes.float16:
    # float16 math is too imprecise, so we do the batch norm gradient
    # computations in float32.
    x = math_ops.cast(x, dtypes.float32)
    grad_y = math_ops.cast(grad_y, dtypes.float32)
  if is_training:
    if data_format == b"NHWC":
      keepdims = False
      reduce_axis = [0, 1, 2]
    else:
      keepdims = True
      reduce_axis = [0, 2, 3]
      shape = [1, array_ops.size(scale), 1, 1]
      scale = array_ops.reshape(scale, shape)
    mean_grad_y = math_ops.reduce_mean(grad_y, reduce_axis, keepdims=keepdims)
    mean_x = math_ops.reduce_mean(x, reduce_axis, keepdims=keepdims)
    var_x = math_ops.reduce_mean(
        math_ops.squared_difference(x, array_ops.stop_gradient(mean_x)),
        reduce_axis,
        keepdims=keepdims)
    grad_y_offset = grad_y - mean_grad_y
    x_offset = x - mean_x
    mean = math_ops.reduce_mean(
        grad_y * x_offset, axis=reduce_axis, keepdims=keepdims)
    grad_x = scale * math_ops.rsqrt(var_x + epsilon) * (
        grad_y_offset - math_ops.reciprocal(var_x + epsilon) * mean * x_offset)
    grad_scale = math_ops.rsqrt(var_x + epsilon) * math_ops.reduce_sum(
        grad_y * x_offset, axis=reduce_axis, keepdims=keepdims)
    if data_format == b"NCHW":
      grad_scale = array_ops.squeeze(grad_scale)
    grad_offset = math_ops.reduce_sum(grad_y, axis=reduce_axis)
    return math_ops.cast(grad_x, x_dtype), grad_scale, grad_offset
  else:
    if data_format == b"NHWC":
      reduce_axis = [0, 1, 2]
    else:
      reduce_axis = [0, 2, 3]
      shape = [1, array_ops.size(pop_mean), 1, 1]
      pop_mean = array_ops.reshape(pop_mean, shape)
      pop_var = array_ops.reshape(pop_var, shape)
      scale = array_ops.reshape(scale, shape)

    grad_offset = math_ops.reduce_sum(grad_y, axis=reduce_axis)
    var_rsqrt = math_ops.rsqrt(pop_var + epsilon)
    grad_scale = math_ops.reduce_sum(
        grad_y * (x - pop_mean) * var_rsqrt, axis=reduce_axis)
    grad_x = grad_y * scale * var_rsqrt
    return math_ops.cast(grad_x, x_dtype), grad_scale, grad_offset