[object Object]

[object Object][object Object]undefined[object Object]

• Description:

  Performs backpropagation of .

• Formula:

$$dx_i= \begin{cases} NaN, & \text{if } x < \text0 \text{ or } x > 1 ,eps <0 \\ 0, & \text{if } x < \text{eps} \text{ or } x > 1 - \text{eps},eps \geq 0 \\ \frac{dy_i}{x_i \cdot (1 - x_i)}, & \text{if } \text{eps} \leq x_i \leq 1 - \text{eps} \\ \end{cases}$$ [object Object]

Each operator has calls. First, aclnnLogitGradGetWorkspaceSize is called to obtain the input parameters and compute the workspace size required by the process. Then, aclnnLogitGrad is called to perform computation.

[object Object]

• Parameters:

  [object Object]

• Returns:

  [object Object]: status code. For details, see .

  The first-phase API implements input parameter verification. The following errors may be thrown:

  [object Object]

[object Object]

• Parameters:

  [object Object]

• Returns:

  [object Object]: status code. For details, see .

[object Object]

• Deterministic compute:

  • aclnnLogitGrad defaults to a deterministic implementation.

• Impact of eps on x and output:

  • If eps is less than 0, the value of x is not in [0, 1], and the output is NaN.
  • If eps is greater than or equal to 0, the value of x is not in [eps, 1 - eps], and the output is 0.
  • If eps is greater than 1, NaN is output. If eps is 1, Inf is output.
  • If eps is Inf, 0 is output.
  • If eps is NaN, NaN is output.

• Impact of x on the output:

  • If x is 0 or 1, Inf is output.
  • If x is Inf or NaN, NaN is output.

• If dy is Inf or NaN, Inf or NaN is output.

[object Object]

The following example is for reference only. For details, see .