Documentation

RNN

Input

  • x:
    • Required: yes
    • Type: fp16
    • Description: time change input, (T × N × ...)
    • Restriction: Must be within 256.
  • cont:
    • Required: yes
    • Type: fp16
    • Description: sequence continuity flag (T × N)
    • Restriction: Must be within 256.
  • x_static (optional):
    • Required: no
    • Type: fp16
    • Description: static data with time unchanged (N × ...)
    • Restrictions: none
  • h_0 (optional):
    • Required: no
    • Type: fp32, fp16
    • Description: initial hidden state (1 × N × num_output)
    • Restrictions: none
  • w_xh:
    • Required: yes
    • Type: fp16
    • Description: xh weight, with shape [input_size, num_output]
    • Restrictions: none
  • bias_h:
    • Required: yes
    • Type: fp32, fp16
    • Description: bias, with shape [num_output,]
    • Restrictions: none
  • w_sh:
    • Required: no
    • Type: fp16
    • Description: sh weight, with shape [input_size, num_output]
    • Restrictions: none
  • w_hh:
    • Required: yes
    • Type: fp16
    • Description: hh weight, with shape [num_output, num_output]
    • Restrictions: none
  • w_ho:
    • Required: yes
    • Type: fp16
    • Description: ho weight, with shape [num_output, num_output]
    • Restrictions: none
  • bias_o:
    • Required: yes
    • Type: fp32, fp16
    • Description: bias, with shape [num_output,]
    • Restrictions: none

Attribute

  • expose_hidden:
    • Required: no
    • Type: bool
    • Description: whether to expose the hidden status (default = false)
    • Restrictions: none
  • num_output:
    • Required: no
    • Type: int
    • Description: number of output features. If the value is 0, the value is obtained from the shape.
    • Restrictions: none

Output

  • o:
    • Required: yes
    • Type: fp32, fp16
    • Description: output tensor (T × N × num_output)
    • Restrictions: none
  • h_t:
    • Required: yes
    • Type: fp32, fp16
    • Description: hidden status of the last time step (1 × N * num_output)
    • Restrictions: none