TensorFlow is no longer what it used to be.

Let’s have a quick history of development overview.

TensorFlow 1 is one of the most widely used deep learning packages.

It is very versatile and that is why many practitioners like it.

However, it has a major disadvantage – it is very hard to learn and use.

Many people become disheartened after seeing even a couple of lines of TensorFlow code.

Not only its methods are strange, but the whole logic of coding is unlike most libraries

out there.

This led to the development and popularization of higher-level packages such as PyTorch and

Keras.

Keras is especially interesting as in 2017 it was integrated in the core TensorFlow – a

feat that may sound a bit strange.

In reality though, both TensorFlow and Keras are open source, so such things do happen

in the programming world.

In fact, Keras’ author claims that Keras is conceived as “an interface for TensorFlow

rather than a different library”, making this integration even easier to digest and

implement.

So far so good.

However, even with Keras as a part of TF, TensorFlow was still losing popularity.

This was addressed in 2019, when TensorFlow 2.0 came on the horizon, or at least its alpha

version.

It is TensorFlow’s effort to catch up with the current demand for higher-level programming.

Interestingly, instead of creating their own high-level syntax, the TF developers chose

to borrow that of Keras.

This decision made sense as Keras was widely adopted already and people generally love

it.

On that note, you may hear people saying: TensorFlow 2 is basically Keras.

In fact, TF 2 has the best of both worlds – most of the versatility of TF 1 and the

high-level simplicity of Keras.

And that’s not all.

There are also other major advantages of TF 2 over TF 1 – they simplified the API, removed

duplicate and deprecated functions and added some new to the core TensorFlow.

Most importantly for us – TensorFlow 2 boasts ‘eager execution’ or in other words – allowing

standard Python “rules of physics” to apply to it, rather than complex computational

graphs, that you don’t really want to know about.