Transposed Conv

• use a transformation going in the opposite direction of a normal convolution, i.e., from something that has the shape of the output of some convolution to something that has the shape of its input while maintaining a connectivity pattern that is compatible with said convolution.
• Upsampling
• Input i, kernel k, padding p, stride s , $o=(i-1)\times s+k-2p$
• Steps
  • Calculate new Param’s z and p’
  • Between each row and columns of the input, insert z number of zeros. This increases the size of the input to $(2\ast i-1)\times (2\ast i-1)$
  • Pad the modified image with p’ no of zeros
  • Standard conv with stride of 1
•