SoftmaxCrossEntropyWithLogitsPass

Description

Fuses the Reshape and SoftmaxCrossEntropyWithLogits operators that fit the graph fusion pattern, eliminating the Reshape operator and retaining only the SoftmaxCrossEntropyWithLogits operator.

Restrictions

• Restrictions on the shape of input and output parameters:

  The shapes of input0, input2, and input4 must be 4D.

• Restrictions on the format of input and output parameters:

  The formats of input0 and input2 must be NHWC, and the channel dimension C must be less than 131072.

• Restrictions on the sequence of input parameters:
  • input0 and input2 must be the first inputs of Reshape.
  • The first output of SoftmaxCrossEntropyWithLogits must be the first input of Reshape.
• Performance description:

  The fusion pattern provides improved performance when the channel dimension C equals 1.

  If the performance deteriorates, disable this fusion pattern.

• Restrictions on the data types of the input parameters:

  The data types of input0 and input2 must be the same.

  • Atlas A2 training product/Atlas A2 inference product: The data type can be FLOAT32 or BFLOAT16.

Availability

Atlas A2 training product/Atlas A2 inference product