I use fullgrad religiously, although I've removed the multiplication by the original image so that I'm just seeing the model gradients. I don't really use it to debug, it's more useful as a post-facto indication of what the important features in the data were. Every once in a while I'll see a model is overly focused on corners or something obviously wrong, and that can be an indication of too much instability, but aside from that it's more of an explanatory tool than a debugging tool.