DeconvolutionV1

Input

• input_data:
  • Required: yes
  • Type: float16
  • Description: feature map
  • Restrictions:

    One input

    Kw=weight[2]

    Kh=weight[3]

    stride=strides[0]

    pad=pads[0]

    dalition=dalitions[0]=1

    • The dilation value can only be 1.

      • K-stride-2*pad=0
      • 0<=(Kw-1-Pad)//stride<=255
      • 0<=(Kh-1-pad)//stride<=255
      • Kh-pad-1>=0
      • Kw-pad-1>=0

    • group=1

      • Kw*Kh*16*3*unitsize+stride*ceil(input_w, 16.0)*16*2<=245760
      • ceil(input_w/8.0)*8*16*unitsize<=245760
      • ceil(input_w, 16)*16*ks2*unitsize<=65536
      • 16*ks2*16*unitsize<=65536
      • stride*ceil(input_w, 16.0)*16*unitsize<=262144
      • Kw*Kh*16*16+2*ks*input_w*16<=1048576
      • Kw=Kh, Kw*Kh%16=0, Kw%stride=0

    • group=channel

      • ceil(input_w, 16)*16*ks2*unitsize<=65536
      • 16*ks2*16*unitsize<=65536
      • stride*ceil(inpt_w, 16.0)*16*unitsize<=262144
      • Kw*Kh*256*unitsize+2*ks*w*16*unitsize<=1048576
      • Kw*Kh*16*3*unitsize+stride*ceil(input_w, 16.0)*16*2+543<=245760
      • Kw*Kh*16*3*unitsize+16*Kw*Kh*16*unitsize+543<=245760
      • Kw=Kh, Kw*Kh%16=0, Kw%stride=0
      

      • If unitsize is 2, fp16 occupies two bytes.
      • ks is the ratio of kernel size to stride.
      • ks2 is the square of ks.
      • KB stands for kilobyte.

• weight:
  • Required: yes
  • Type: float16
  • Description: deconvolution weight
  • Restriction: For details, see the restrictions of input_data.

• bias:
  • Required: no
  • Type: float16
  • Description: bias
  • Restriction: For details, see the restrictions of input_data.

Attribute

• strides:
  • Required: yes
  • Type: listInt
  • Description: deconvolution stride
  • Restriction: For details, see the restrictions of input_data.

Output