Python utils.constant_value方法代码示例

# 需要导入模块: from tensorflow.python.layers import utils [as 别名]
# 或者: from tensorflow.python.layers.utils import constant_value [as 别名]
def _smart_select(pred, fn_then, fn_else):
  """Selects fn_then() or fn_else() based on the value of pred.

  The purpose of this function is the same as `utils.smart_cond`. However, at
  the moment there is a bug (b/36297356) that seems to kick in only when
  `smart_cond` delegates to `tf.cond`, which sometimes results in the training
  hanging when using parameter servers. This function will output the result
  of `fn_then` or `fn_else` if `pred` is known at graph construction time.
  Otherwise, it will use `tf.where` which will result in some redundant work
  (both branches will be computed but only one selected). However, the tensors
  involved will usually be small (means and variances in batchnorm), so the
  cost will be small and will not be incurred at all if `pred` is a constant.

  Args:
    pred: A boolean scalar `Tensor`.
    fn_then: A callable to use when pred==True.
    fn_else: A callable to use when pred==False.

  Returns:
    A `Tensor` whose value is fn_then() or fn_else() based on the value of pred.
  """
  pred_value = utils.constant_value(pred)
  if pred_value:
    return fn_then()
  elif pred_value is False:
    return fn_else()
  t_then = array_ops.expand_dims(fn_then(), 0)
  t_else = array_ops.expand_dims(fn_else(), 0)
  pred = array_ops.reshape(pred, [1])
  result = array_ops.where(pred, t_then, t_else)
  return array_ops.squeeze(result, [0]) 

# 需要导入模块: from tensorflow.python.layers import utils [as 别名]
# 或者: from tensorflow.python.layers.utils import constant_value [as 别名]
def _fused_batch_norm(self, inputs, training):
    """Returns the output of fused batch norm."""
    # TODO(reedwm): Add support for fp16 inputs.
    beta = self.beta if self.center else self._beta_const
    gamma = self.gamma if self.scale else self._gamma_const

    def _fused_batch_norm_training():
      return nn.fused_batch_norm(
          inputs,
          gamma,
          beta,
          epsilon=self.epsilon,
          data_format=self._data_format)

    def _fused_batch_norm_inference():
      return nn.fused_batch_norm(
          inputs,
          gamma,
          beta,
          mean=self.moving_mean,
          variance=self.moving_variance,
          epsilon=self.epsilon,
          is_training=False,
          data_format=self._data_format)

    output, mean, variance = utils.smart_cond(
        training, _fused_batch_norm_training, _fused_batch_norm_inference)
    if not self._bessels_correction_test_only:
      # Remove Bessel's correction to be consistent with non-fused batch norm.
      # Note that the variance computed by fused batch norm is
      # with Bessel's correction.
      sample_size = math_ops.cast(
          array_ops.size(inputs) / array_ops.size(variance), variance.dtype)
      factor = (sample_size - math_ops.cast(1.0, variance.dtype)) / sample_size
      variance *= factor

    training_value = utils.constant_value(training)
    if training_value is None:
      one_minus_decay = _smart_select(training,
                                      lambda: self._one_minus_decay,
                                      lambda: 0.)
    else:
      one_minus_decay = ops.convert_to_tensor(self._one_minus_decay)
    if training_value or training_value is None:
      mean_update = self._assign_moving_average(self.moving_mean, mean,
                                                one_minus_decay)
      variance_update = self._assign_moving_average(self.moving_variance,
                                                    variance, one_minus_decay)
      if context.in_graph_mode():
        # Note that in Eager mode, the updates are already executed when running
        # assign_moving_averages. So we do not need to put them into
        # collections.
        self.add_update(mean_update, inputs=inputs)
        self.add_update(variance_update, inputs=inputs)

    return output