It has a lot of features, I love that TensorFlow builds its eager execution, but that’s happening in TensorFlow version 2. Its happen after the wave of PyTorch come, and at its first release they said that it's not recommended, but the newest version did not say that.

There are many things that I did not cover in this article because of time did not permit me to do that. But, both of them do their best, but if you are more to pythonian engineer side like me and at the same time love to do the math and science. I think you would choose pytorch, the reason is I like the way pytorch was built from its foundation (more math API naming, more modularity, more easily debugged), it's intuitively compatible with me, its feels like home.

For disclaimer, I’ve been switch to pytorch since version 0.3.0, if you see the growth of PyTorch today, it's really big. I follow the progress of pytorch from 0.3.0 and I’ve lot of confident it will be the next most usage deep learning frameworks in research and industry, I trust my guts.