I heard about new version of TensorFlow and that it is more simple, like 'keras'. But does it lose it terms of performance and flexibility to tf1? If not, are there any benefits to tf1 right now?
First of all, you should know tf1 will not be further developed. All the development is now going on in tf2. Secondly, no, tf2 is in no way worse than tf1 in terms of performance or flexibility. Moreover, you can easily migrate your tf1 code to tf2. Any code. All you need is 'disablev2behavior()' function to use namespace names from tf1. There are practically no benefits in staying on tf1 code.