I feel like I'm trying to create some pipelines that should be fairly common among Tensorflow users. There seems to be multiple ways of doing everything, each method with it's own nuance, that is not officially documented anywhere. You have to dig around on their github issues and sometimes the info will be there, if you're lucky.
There's a lot of unofficial medium blogs out there, but I've noticed that often they have inaccurate information. A lot of the times it's contractors creating marketing materials to get hired.
A lot of the design choices I google lead me to unanswered stackoverflow questions, asked several months or a year ago. I think this is the easiest fix. A lot of these unanswered questions are good questions, that are not covered anywhere in Tensorflow's official documentation.
I am looking at the backlog of unaswered questions for both. Pytorch has 2,101, Tensorflow has 24,066
Pytorch, on the other hand, has a forum where you often get your questions answered by someone on the Pytorch developers team.
I hear a lot of complains about Tensorflow, but I personally don't see the structure itself as poor. I think it's pretty good actually, but what's the point of creating something awesome if you don't enough info for people to optimally use it.