Deconvolutions checker boad

To avoid a checker board pattern resulting from many deconvolutions (as described by Jeremy in lesson 12) would it be a good solution to just use more random numbers (say 4 x 100 x 64 x 64 as the input noise tensor for the GAN generator in wgan1.ipynb) to match the height and width of the desired output and to modify the architecture to just do convolutions (with padding to avoid grid size shrinkage) instead of deconvolutions?