I have a particular Input with Shape = [NxHxWxC_in] and a kernel of Size = [n_h,n_w,stride_h, stride_w] with C_out number of filters (the strides can be 1 and 1 if that simplifies things but a general answer would be even better).
What is the most efficient way in TensorFlow of creating 1D Conv / Affine transformation layer combinations to get the same results as the 2D convolution ?
I have found something called a helix transform but I am unable to see whether this can be extended to multiple channels (with C_in and C_out being fairly large numbers).