Deconvolution

General Restrictions

If:

Ny, Cy, Hy, Wy [x1] = x.shape

Nk, Ck/group, Hk, Wk = filter.shape (k=kernel)

stride, stride_h, stride_w, pad, pad_h, pad_w:

if stride.size() == 0:
   stride_h, stride_w
elif stride.size() == 1:
   stride_h, stride_w = stride[0]
else:
   stride_h = stride[0]
   stride_w = stride[1]

if pad.size() == 0:
   pad_h, pad_w
elif pad.size() == 1:
   pad_h = pad_w = pad[0]
else:
   pad_h = pad[0]
   pad_w = pad[1]
if dilation.size() == 1:
   dilation_h = dilation_w = dilation[0]
else:
   dilation_h = dilation[0]
   dilation_w = dilation[1]

If the output shape is [Nx,Cx,Hx,Wx]:

• Nx=Ny, Cx=Ck
• Wx = (Wy-1)*stride_w - 2*pad_w + dilation_h*(Wk-1)+1
• Hx = (Hy-1)*stride_h - 2*pad_h + dilation_w*(Hk-1)+1

The general restrictions are as follows:

1. 1<=stride_h*stride_w<256
2. Hy + 2 * pad_h >= dilation_h * (Hk – 1) + 1, Wy + 2 * pad_w >= dilation_w * (Wk – 1) + 1
3. Hy * stride_h and Wy * stride_w must be in the range [2, 4096]:
   • When Hk==Wk==1:

     When Wy * stride_w * stride_h is in the range [2, 4096] and stride_h or stride_w is greater than 1, Wy is less than or equal to 2032.

   • Atlas Training Series Product: Hx! = Hk, Wx! = Wk, Hy or Wy can be 1.
   • When Hy=1, Hk=1, pad_h=0, stride_h=1, dilation_h=1, Wx is in the range [1, 2147483647] (representable by int32).
4. AL1Size:

   If w_value = Wy * stride_w:

       if Wx>16:
           h_value_max = Hk + 1
       else if 16% Wx== 0: 
           h_value_max = Hk + 16//Wx - 1
       else: 
           h_value_max = Hk + 16//Wx +1

   AL1Size = h_value_max * w_value * 32 bytes, BL1Size = Hk * Wk * 512 bytes

   AL1Size + BL1Size ≤ l1_size (For details about l1_size, see the configuration file of the corresponding SoC, that is, platform_config/${soc_version}.ini.)

   Hy >= dilation * (Hk – 1) + 1, Wy >= dilation * (Wk – 1) + 1

   

   When Hy = 1, Hk = 1, pad_h = 0, stride_h = 1, dilation_h = 1, then, AL1Size = [15 * stride_w + (Wk – 1) * dilation_w + 1] * 32 bytes

Input

• x:
  • Required: yes
  • Type: float16
  • Description: input tensor
  • Restrictions: none

Attribute

• pad:
  • Required: no
  • Type: ListInt
  • Description: padding size (default = 0). pad and pad_h/pad_w are mutually exclusive. pad is a list of length up to 2.
  • Restriction: When the length of the list is 1, pad < Hk and pad < Wk. When the length of the list is 2, pad[0] < Hk and pad[1] < Wk.

Output