Text this: Faster Training of Large Kernel Convolutions on Smaller Spatial Scales