Thank you for your work.
Is there any benefit to this operation?
for i in range(N/S):
step 1 (slicing): slicing the input data at the batch dimension from iS to (i+1)S, input (NxCxHxW) -> sliced input (SxCxHxW)
step 2 (deformable im2col): sliced input (SxCxHxW)+sliced offset (Sx18xHxW) -> column (Cx9xSxHxW)
step 3 (MatMul&reshape): weight matrix (C'x 9C) * column (9CxSHW) -> temp sliced output (C'xSxHxW) -> sliced output (SxC'xHxW)
step 4 (Merge): merge sliced output to form the whole output data (NxC'xHxW)
end
Thank you for your work. Is there any benefit to this operation?
for i in range(N/S): step 1 (slicing): slicing the input data at the batch dimension from iS to (i+1)S, input (NxCxHxW) -> sliced input (SxCxHxW) step 2 (deformable im2col): sliced input (SxCxHxW)+sliced offset (Sx18xHxW) -> column (Cx9xSxHxW) step 3 (MatMul&reshape): weight matrix (C'x 9C) * column (9CxSHW) -> temp sliced output (C'xSxHxW) -> sliced output (SxC'xHxW) step 4 (Merge): merge sliced output to form the whole output data (NxC'xHxW) end