Training for grey images

[bzisl]

I know that we can train without problem from gray images, but the size of the image is 3x the grey one, and then, the batches are 3x smaller that it could be. Smaller batches impact quite negatively on the convergence of the process. With bigger batches, the output will be much better and we will reduce also the conversion requirement, thinking in small devices in mind.
Do you believe that the changes are medium-easy or will it be hard?