tf.raw_ops.FusedBatchNormGrad

Gradient for batch normalization.

tf.raw_ops.FusedBatchNormGrad(
    y_backprop, x, scale, reserve_space_1, reserve_space_2, epsilon=0.0001,
    data_format='NHWC', is_training=True, name=None
)

Note that the size of 4D Tensors are defined by either "NHWC" or "NCHW". The size of 1D Tensors matches the dimension C of the 4D Tensors.

Args:

  • y_backprop: A Tensor. Must be one of the following types: float32. A 4D Tensor for the gradient with respect to y.
  • x: A Tensor. Must have the same type as y_backprop. A 4D Tensor for input data.
  • scale: A Tensor. Must have the same type as y_backprop. A 1D Tensor for scaling factor, to scale the normalized x.
  • reserve_space_1: A Tensor. Must have the same type as y_backprop. When is_training is True, a 1D Tensor for the computed batch mean to be reused in gradient computation. When is_training is False, a 1D Tensor for the population mean to be reused in both 1st and 2nd order gradient computation.
  • reserve_space_2: A Tensor. Must have the same type as y_backprop. When is_training is True, a 1D Tensor for the computed batch variance (inverted variance in the cuDNN case) to be reused in gradient computation. When is_training is False, a 1D Tensor for the population variance to be reused in both 1st and 2nd order gradient computation.
  • epsilon: An optional float. Defaults to 0.0001. A small float number added to the variance of x.
  • data_format: An optional string from: "NHWC", "NCHW". Defaults to "NHWC". The data format for y_backprop, x, x_backprop. Either "NHWC" (default) or "NCHW".
  • is_training: An optional bool. Defaults to True. A bool value to indicate the operation is for training (default) or inference.
  • name: A name for the operation (optional).

Returns:

A tuple of Tensor objects (x_backprop, scale_backprop, offset_backprop, reserve_space_3, reserve_space_4).

  • x_backprop: A Tensor. Has the same type as y_backprop.
  • scale_backprop: A Tensor. Has the same type as y_backprop.
  • offset_backprop: A Tensor. Has the same type as y_backprop.
  • reserve_space_3: A Tensor. Has the same type as y_backprop.
  • reserve_space_4: A Tensor. Has the same type as y_backprop.