Documentation

LSTM

Input

  • x:
    • Required: yes
    • Type: float16
    • Description: time change input x, with shape [T, N, input_size]
    • Restriction: Must be within 256.
  • cont:
    • Required: yes
    • Type: float16, float32
    • Description: sequence continuity flag, with shape [T, N]
    • Restriction: Must be within 256.
  • w_x:
    • Required: yes
    • Type: float16
    • Description: x's weight, with shape [input_size, 4 × num_output]
    • Restrictions: none
  • bias:
    • Required: yes
    • Type: float16, float32
    • Description: bias, with shape [4 × num_output]
    • Restrictions: none
  • w_h:
    • Required: yes
    • Type: float16
    • Description: H weight, with shape [4 × num_output, num_output]
    • Restrictions: none
  • x_static:
    • Required: no
    • Type: float16
    • Description: static data with time unchanged, with shape [N, input_size]. The parameter is reserved.
    • Restriction: When static is specified, inputFM and HiddenSize must be multiples of 16.
  • h_0:
    • Required: no
    • Type: float16, float32
    • Description: the initial hidden status, with shape [1,N,num_output]. Required when expose_hidden = True.
    • Restrictions: none
  • c_0:
    • Required: no
    • Type: float16, float32
    • Description: the initial cell status, with shape [1,N,num_output]. Required when expose_hidden = True.
    • Restrictions: none
  • w_x_static:
    • Required: no
    • Type: float16
    • Description: x_static weight, with shape [4 × num_output, input_size]. The parameter is reserved.
    • Restrictions: none

Attribute

  • num_output:
    • Required: yes
    • Type: uint
    • Description: number of output features. Must be greater than 0.
    • Restriction: Must be greater than 0.
  • expose_hidden:
    • Required: no
    • Type: bool
    • Description: whether the initial hidden status and initial cell status exist (default = False)
    • Restrictions: none

Output

  • h:
    • Required: yes
    • Type: float16, float32
    • Description: output tensor (T × N × num_output)
    • Restrictions: none
  • h_t:
    • Required: no
    • Type: float16, float32
    • Description: hidden status of the last time step (1 × N × num_output)
    • Restrictions: none
  • c_t:
    • Required: no
    • Type: float16, float32
    • Description: cell status of the last time step (1 × N × num_output)
    • Restrictions: none