François CholletVerified account

A misconception I see is that Keras is made or owned by me. I may have started it a long time ago, but nowadays I do very little compared to everyone else. keras-team/keras has 750 contributors, including a dozen of very active ones. The Keras repo exists thanks to them.

At Google, tf.keras (the internal TensorFlow implementation of the Keras spec) is made by 8 awesome engineers (come work with us!), and many more people contribute to it. And more great folks are working on the Keras TypeScript implementation as part of TensorFlow.js.

And then there are many more people elsewhere working on 3rd-party Keras implementations: the Keras-MXNet team at AWS, the Scala implementation from SkyMind, PlaidML, CoreML... As for me, I mostly do code reviews and I participate in design discussions.

Another misconception I see is that I have a leadership position inside TensorFlow. I do not. I don’t actually make any important decisions with regard to TF, and never have. That’s the role of the TF leads, 5 people, of which I am not one.

In general, the TF team is quite large, with over 100 people (I think?), and none of them report to me. My scope is very narrow: looking at API changes within tf.keras, reviewing tf.keras code. That's really it.

One last misconception I see is that Keras is a TensorFlow wrapper library. In fact, Keras is an API standard for defining & training deep learning models, similar to what Gluon is to MXNet. Like Gluon, it's meant not to be tied to a specific implementation.