Python nn.moments方法代码示例

# 需要导入模块: from tensorflow.python.ops import nn [as 别名]
# 或者: from tensorflow.python.ops.nn import moments [as 别名]
def _data_dep_init(self, inputs):
        """Data dependent initialization for eager execution"""
        from tensorflow.python.ops.nn import moments
        from tensorflow.python.ops.math_ops import sqrt

        with variable_scope.variable_scope("data_dep_init"):
            # Generate data dependent init values
            activation = self.layer.activation
            self.layer.activation = None
            x_init = self.layer.call(inputs)
            m_init, v_init = moments(x_init, self.norm_axes)
            scale_init = 1.0 / sqrt(v_init + 1e-10)

        # Assign data dependent init values
        self.layer.g = self.layer.g * scale_init
        self.layer.bias = -1 * m_init * scale_init
        self.layer.activation = activation
        self.initialized = True

    # pylint: disable=signature-differs 

# 需要导入模块: from tensorflow.python.ops import nn [as 别名]
# 或者: from tensorflow.python.ops.nn import moments [as 别名]
def normalize(self, inputs):
    """Apply normalization to input.

    The shape must match the declared shape in the constructor.
    [This is copied from tf.contrib.rnn.LayerNormBasicLSTMCell.]

    Args:
      inputs: Input tensor

    Returns:
      Normalized version of input tensor.

    Raises:
      ValueError: if inputs has undefined rank.
    """
    inputs_shape = inputs.get_shape()
    inputs_rank = inputs_shape.ndims
    if inputs_rank is None:
      raise ValueError('Inputs %s has undefined rank.' % inputs.name)
    axis = range(1, inputs_rank)

    beta = self._component.get_variable('beta_%s' % self._name)
    gamma = self._component.get_variable('gamma_%s' % self._name)

    with tf.variable_scope('layer_norm_%s' % self._name):
      # Calculate the moments on the last axis (layer activations).
      mean, variance = nn.moments(inputs, axis, keep_dims=True)

      # Compute layer normalization using the batch_normalization function.
      variance_epsilon = 1E-12
      outputs = nn.batch_normalization(
          inputs, mean, variance, beta, gamma, variance_epsilon)
      outputs.set_shape(inputs_shape)
      return outputs 

# 需要导入模块: from tensorflow.python.ops import nn [as 别名]
# 或者: from tensorflow.python.ops.nn import moments [as 别名]
def normalize_batch_in_training(x, gamma, beta, reduction_axes, epsilon=1e-3):
  """Computes mean and std for batch then apply batch_normalization on batch.

  Arguments:
      x: Input tensor or variable.
      gamma: Tensor by which to scale the input.
      beta: Tensor with which to center the input.
      reduction_axes: iterable of integers,
          axes over which to normalize.
      epsilon: Fuzz factor.

  Returns:
      A tuple length of 3, `(normalized_tensor, mean, variance)`.
  """
  mean, var = nn.moments(
      x, reduction_axes, shift=None, name=None, keep_dims=False)
  if sorted(reduction_axes) == list(range(ndim(x)))[:-1]:
    normed = nn.batch_normalization(x, mean, var, beta, gamma, epsilon)
  else:
    # need broadcasting
    target_shape = []
    for axis in range(ndim(x)):
      if axis in reduction_axes:
        target_shape.append(1)
      else:
        target_shape.append(array_ops.shape(x)[axis])
    target_shape = array_ops.stack(target_shape)

    broadcast_mean = array_ops.reshape(mean, target_shape)
    broadcast_var = array_ops.reshape(var, target_shape)
    if gamma is None:
      broadcast_gamma = None
    else:
      broadcast_gamma = array_ops.reshape(gamma, target_shape)
    if beta is None:
      broadcast_beta = None
    else:
      broadcast_beta = array_ops.reshape(beta, target_shape)
    normed = nn.batch_normalization(x, broadcast_mean, broadcast_var,
                                    broadcast_beta, broadcast_gamma, epsilon)
  return normed, mean, var 

# 需要导入模块: from tensorflow.python.ops import nn [as 别名]
# 或者: from tensorflow.python.ops.nn import moments [as 别名]
def normalize(self, inputs):
    """Apply normalization to input.

    The shape must match the declared shape in the constructor.
    [This is copied from tf.contrib.rnn.LayerNormBasicLSTMCell.]

    Args:
      inputs: Input tensor

    Returns:
      Normalized version of input tensor.

    Raises:
      ValueError: if inputs has undefined rank.
    """
    inputs_shape = inputs.get_shape()
    inputs_rank = inputs_shape.ndims
    if inputs_rank is None:
      raise ValueError('Inputs %s has undefined rank.' % inputs.name)
    axis = range(1, inputs_rank)

    beta = self._component.get_variable('beta_%s' % self._name)
    gamma = self._component.get_variable('gamma_%s' % self._name)

    with tf.variable_scope('layer_norm_%s' % self._name):
      # Calculate the moments on the last axis (layer activations).
      mean, variance = nn.moments(inputs, axis, keep_dims=True)

      # Compute layer normalization using the batch_normalization function.
      variance_epsilon = 1E-12
      outputs = nn.batch_normalization(inputs, mean, variance, beta, gamma,
                                       variance_epsilon)
      outputs.set_shape(inputs_shape)
      return outputs 

# 需要导入模块: from tensorflow.python.ops import nn [as 别名]
# 或者: from tensorflow.python.ops.nn import moments [as 别名]
def _renorm_correction_and_moments(self, mean, variance, training):
    """Returns the correction and update values for renorm."""
    stddev = math_ops.sqrt(variance + self.epsilon)
    # Compute the average mean and standard deviation, as if they were
    # initialized with this batch's moments.
    mixed_renorm_mean = (self.renorm_mean +
                         (1. - self.renorm_mean_weight) * mean)
    mixed_renorm_stddev = (self.renorm_stddev +
                           (1. - self.renorm_stddev_weight) * stddev)
    # Compute the corrections for batch renorm.
    r = stddev / mixed_renorm_stddev
    d = (mean - mixed_renorm_mean) / mixed_renorm_stddev
    # Ensure the corrections use pre-update moving averages.
    with ops.control_dependencies([r, d]):
      mean = array_ops.identity(mean)
      stddev = array_ops.identity(stddev)
    rmin, rmax, dmax = [self.renorm_clipping.get(key)
                        for key in ['rmin', 'rmax', 'dmax']]
    if rmin is not None:
      r = math_ops.maximum(r, rmin)
    if rmax is not None:
      r = math_ops.minimum(r, rmax)
    if dmax is not None:
      d = math_ops.maximum(d, -dmax)
      d = math_ops.minimum(d, dmax)
    # When not training, use r=1, d=0, and decay=1 meaning no updates.
    r = _smart_select(training, lambda: r, lambda: array_ops.ones_like(r))
    d = _smart_select(training, lambda: d, lambda: array_ops.zeros_like(d))
    decay = _smart_select(training, lambda: self.renorm_momentum, lambda: 1.)
    def _update_renorm_variable(var, weight, value):
      """Updates a moving average and weight, returns the unbiased value."""
      # Update the variables without zero debiasing. The debiasing will be
      # accomplished by dividing the exponential moving average by the weight.
      # For example, after a single update, the moving average would be
      # (1-decay) * value. and the weight will be 1-decay, with their ratio
      # giving value.
      # Make sure the weight is not updated until before r and d computation.
      value = array_ops.identity(value)
      with ops.control_dependencies([value]):
        weight_value = array_ops.constant(1., dtype=weight.dtype)
      new_var = moving_averages.assign_moving_average(
          var, value, decay, zero_debias=False)
      new_weight = moving_averages.assign_moving_average(
          weight, weight_value, decay, zero_debias=False)
      return new_var / new_weight

    with ops.colocate_with(self.moving_mean):
      new_mean = _update_renorm_variable(self.renorm_mean,
                                         self.renorm_mean_weight,
                                         mean)
    with ops.colocate_with(self.moving_variance):
      new_stddev = _update_renorm_variable(self.renorm_stddev,
                                           self.renorm_stddev_weight,
                                           stddev)
      # Make sqrt(moving_variance + epsilon) = new_stddev.
      new_variance = math_ops.square(new_stddev) - self.epsilon

    return (r, d, new_mean, new_variance) 

# 需要导入模块: from tensorflow.python.ops import nn [as 别名]
# 或者: from tensorflow.python.ops.nn import moments [as 别名]
def _renorm_correction_and_moments(self, mean, variance, training):
    """Returns the correction and update values for renorm."""
    stddev = math_ops.sqrt(variance + self.epsilon)
    # Compute the average mean and standard deviation, as if they were
    # initialized with this batch's moments.
    mixed_renorm_mean = (self.renorm_mean +
                         (1. - self.renorm_mean_weight) * mean)
    mixed_renorm_stddev = (self.renorm_stddev +
                           (1. - self.renorm_stddev_weight) * stddev)
    # Compute the corrections for batch renorm.
    r = stddev / mixed_renorm_stddev
    d = (mean - mixed_renorm_mean) / mixed_renorm_stddev
    # Ensure the corrections use pre-update moving averages.
    with ops.control_dependencies([r, d]):
      mean = array_ops.identity(mean)
      stddev = array_ops.identity(stddev)
    rmin, rmax, dmax = [self.renorm_clipping.get(key)
                        for key in ['rmin', 'rmax', 'dmax']]
    if rmin is not None:
      r = math_ops.maximum(r, rmin)
    if rmax is not None:
      r = math_ops.minimum(r, rmax)
    if dmax is not None:
      d = math_ops.maximum(d, -dmax)
      d = math_ops.minimum(d, dmax)
    # When not training, use r=1, d=0, and decay=1 meaning no updates.
    r = _smart_select(training, lambda: r, lambda: array_ops.ones_like(r))
    d = _smart_select(training, lambda: d, lambda: array_ops.zeros_like(d))
    decay = _smart_select(training, lambda: self.renorm_momentum, lambda: 1.)

    def _update_renorm_variable(var, weight, value):
      """Updates a moving average and weight, returns the unbiased value."""
      # Update the variables without zero debiasing. The debiasing will be
      # accomplished by dividing the exponential moving average by the weight.
      # For example, after a single update, the moving average would be
      # (1-decay) * value. and the weight will be 1-decay, with their ratio
      # giving value.
      # Make sure the weight is not updated until before r and d computation.
      value = array_ops.identity(value)
      with ops.control_dependencies([value]):
        weight_value = array_ops.constant(1., dtype=weight.dtype)
      new_var = moving_averages.assign_moving_average(
          var, value, decay, zero_debias=False)
      new_weight = moving_averages.assign_moving_average(
          weight, weight_value, decay, zero_debias=False)
      return new_var / new_weight

    with ops.colocate_with(self.moving_mean):
      new_mean = _update_renorm_variable(self.renorm_mean,
                                         self.renorm_mean_weight,
                                         mean)
    with ops.colocate_with(self.moving_variance):
      new_stddev = _update_renorm_variable(self.renorm_stddev,
                                           self.renorm_stddev_weight,
                                           stddev)
      # Make sqrt(moving_variance + epsilon) = new_stddev.
      new_variance = math_ops.square(new_stddev) - self.epsilon

    return (r, d, new_mean, new_variance)