Inside TensorFlow: tf.Keras (Part 1)

Take an inside look into the TensorFlow team’s own internal training sessions--technical deep dives into TensorFlow by the very people who are building it!

On this episode of Inside TensorFlow, creator of Keras, Francois Chollet gives us the overview of tf.Keras.

Let us know what you think about this presentation in the comments below and stay tuned for part 2 coming next week!

[DKL]

"I am Francois and I work on the Keras team."

Wow such a humble guy.

[stephenmugisha]

Thanks for 'open sourcing' your trainings 😀..its been very insightful: am tf.eager() to see more of this..

[JeremyCoppin]

Thank you. Very nicely done. One comment though. Static presentations are nice but nothing beats a live demo where the thought process of what goes where becomes clear. Its the difference between shown a recipe and shown how the ingredients react.

[FatimaCantante]

Very interesting that you are creating this series, kindly keep on doing it. It would be great of course to see the lecturers code while they talk.

[GradientGroup]

For the non-trainable weights example, using x = tf.ones((1, 1)) throws an exception due to incompatible shapes. x = tf.ones((1, )) fixes it. The issue seems to be tf.reduce_sum(inputs, axis=0). Changing the axis parameter from 0 to None fixes it and x = tf.ones((1, 1)) works.

[oliverrosario]

Thanks for the contributions, waiting for part 2.

[kawingchan]

I tried subclassing a model that has a subclass layer and tried to save. It complained it can’t do that for subclass model. I simplified it to have only 1 layer and use save_weights instead. Unfortunately, I ran into AttributeError. It doesn’t look easy to do this.

[allendaniel]

Thank you for this great presentation.

[maziarkasaeiroodsari]

All those who complain about the restriction of Keras functional API, also think that Torch 7 is the pinnacle of deep learning. However, Torch 7 has that same restriction :)

[fuhaoda]

May I know where we can get the slides?

[shinochono]

Learning Keras from Francois is exciting. Can you please share the slide deck for this presentation?

[geoa]

during 4:34- 4:36 by "anything else" does Mr Chollet mean hyperparameters ? Correct me, please if I'n wrong

[victorariasvanegas]

very thanks to @tensorflow team for translate this video.

[congminh]

At 23:45, Does anyone know how to know those blocks below the InceptionNet?

[svitirur]

Why don't you do tensorflow js courses? ?

[freee]

I love how automatic translation on this, machine learning video, doesn't work well, hahaha... No judgement, I just like the irony...
Edit: but it is also good considering the presenter's accent.

[wilamshields]

Find it impossible to understand who is being said by Francois due to his strong accent, even with my MK1 neural network.

[SuperMaDBrothers]

9:19 'Layers can have nontrainable states. So a trainable state is ...' I love how you defined the term after you gave the explanation. Cut the jargon out just tell us what the hell you actually did. So annoying. Other than the jargon, good tutorial