Changing Model Architecture for Deep Learning

I think understanding stuff builds your craft in the field. And it's naturally rewarding if you become better and better, you just feel good. But there's another aspect of ML that some data scientists are not wholly ignorant of, but it's not a priority for them. You have to go into the source code and dig through like, how does this implementation actually work? And why is this part of it so slow? Or why is there conflict here? If I set this parameter, why does it throw an exception here? We've all been there when we started playing around with these libraries. It's going to be a very painful experience,. Like trying to troubleshoot what's going on