It might be an idea to train the same data with fast.ai and see where it leads you. If you have the data in the right structure, it should not take too long. Let me know if you have questions... :)
I guess titles are meant to attract and provoke a little at the same time...
Fast.ai uses a top-down approach, allowing students to get awesome results with a minimal set of instructions, and start to dig deeper from there. If you look at the takeaways from the first lesson, you can see it's more than just three lines of Python and you're done, It's a solid and fresh approach to learn some actual practical deep learning.
Additionally, it kind of depends on what you want to do, we are all using things other people built, depending on the level of abstraction we care for or are willing to deal with...
I found/find fast.ai to be incredibly useful for its practicality, good results, and top-down approach, however, it is sometimes hard to reproduce the results as well as clearly distilling what it is you can actually learn from each lesson. Writing this post blew my socks off as to what was taught in the video, yet it took me quite some time to get it all. So I hope the posts help people with that aspect of the course.
At the moment, I'm learning fast.ai/PyTorch in parallel with Keras/Tensorflow, so at this point, I have no definitive answer to your question which one is preferable. It will probably depend and they will most likely have their own benefits (I know that the boring answer, but I need to get more experience to give you a better answer).
As an exercise I'm trying to write the fast.ai notebooks in Keras, to see how they stack up. Might need to do a post on that as well.
I hope to answer your question better in the future. Could you tell me more about what you want to achieve, I might be of more assistance?
Have a look at our Quick Guides (https://github.com/zerotosingularity/seeme-quick-guides). These are literally a work in progress, but we are working on hard to get it right for early customers...