Tensor flow 2.0
You might have already heard about TensorFlow: an amazing technology created by Google and made available to everyone in open source.
If you haven't heard about it, you have probably used it already! Nowadays is EVERYWHERE and especially in applications like Google Photos and other tools in the Google suite.
We listened to this amazing podcast with one of the creators of TensorFlow: Rajat Monga head of the TensorFlow team at Google. After a bit of historical background on Tensor Flow, Rajat talks about the future of this great technology.
We are sharing the main points to retain about this fascinating conversation:
TensorFlow is more than tech: it is a community and an ecosystem of tools.
With 41m downloads, 60K commits on Github and 18k contributors, Tensor flow is a "culture" in the deep learning community and beyond.
Great documentation and ease to use is the recipe for success.
Tensor Flow became the most used deep learning framework on the planet because developers with limited knowledge on Machine Learning can integrate it in their Software projects.
Flexibility, flexibility and flexibility.
While the most common case is by far, transfer learning (e.g. adapt a pre-trained model to your problem) TF offers flexible solutions for users with different Deep Learning skills, if you are a researcher you can design your own layer, while if you are a developer with limited time, you can use a pre-trained model (e.g. Inception or Resnet ).
[NEW] - Keras is the easiest way to interact with tensor flow.
The strategy of google is clear: more of the same.. enabling non-expert users to get their transfer learning use cases done will become even easier through Keras.
[NEW] - Monolithic architecture is unbundled without breaking compatibility.
TF 2.0 has better integration among elements, and more harmonization between the different tools and projects so that you will be able to seamlessly run models on cluster or mobile.
[NEW] - Improved Portability.
While you already have a lot of options at the moment, it is still very challenging to move models around. With TF 2.0 there is more cohesiveness in the ecosystem with the ability to port models from TS to JS, to TF extended and TF mobile.
Listen to the entire podcast on overcast