Placeholder Image

Subtitles section Play video

  • ♪ (intro music) ♪

  • Welcome to Ask TensorFlow where we answer the questions

  • that you submitted with the hashtag #AskTensorFlow.

  • I'm Paige and I'm a developer advocate on the TensorFlow team.

  • And I'm Alex and I'm an engineer.

  • So for the first question today, Codigo asked,

  • "What will be the support model for stand-alone Keras?

  • Excellent, so that's a phenomenal question, Codigo

  • and we're really excited to answer it.

  • Keras has been integrated as part of TensorFlow 2.0

  • as tf.keras.

  • But stand-alone Keras still certainly exists.

  • A number of the optimizations that we've included

  • for tf.keras haven't yet made their way back

  • to the stand-alone package but we are setting up

  • a special interest group specifically for Keras users

  • for TensorFlow.

  • So if you have any questions, comments, concerns,

  • feel free to join the SIG Keras group and you can find that

  • on the TensorFlow community section.

  • Yeah, thanks for the question, Codigo.

  • Our second question is from Gurjeet, who asked,

  • "Does tf.keras include everything that stand-alone Keras includes?"

  • That's a great question, Gurjeet, we get that one a lot,

  • and I understand where you're coming from.

  • The simple answer is no.

  • There are lots of things in stand-alone Keras

  • that TensorFlow has no business having,

  • like the Theano backend or the CNTK backend.

  • But all the useful pieces, all of the optimizers, the metrics,

  • the losses, the layers, the model building API,

  • all the things that you need to use Keras

  • to train your models, to deploy your models,

  • they're all a part of tf.keras.

  • Absolutely, and if you're familiar with stand-alone Keras,

  • the syntax is very similar, if not identical using tf.keras.

  • For the next question, Clement asked,

  • "What will TensorFlow 2.0 change to stand-alone Keras?"

  • That's a great question.

  • And the answer is that that's still yet to be defined.

  • I mentioned before that we have a special interest group,

  • it's going to be specifically focused on Keras

  • and that will include a lot of the interplay

  • between tf.keras and stand-alone Keras.

  • The ideal is that all of the great optimizations

  • that we've made for tf.keras

  • will find their way back to the open source community

  • and we would love to have your help with that.

  • So if you have interest, join the SIG Keras group.

  • Yeah, you can find all the TensorFlow SIGs

  • in github.com/tensorflow/community.

  • Our next question is also from Gurjeet, who asked,

  • "Is there support for Bayesian layers in tf.keras?"

  • So in a way, while tf.keras itself

  • does not ship with any Bayesian layers.

  • TensorFlow Probability does.

  • They have a full set of Bayesian layers,

  • you can do all sorts of cool things with it,

  • including turning your deep neural network

  • into a Gaussian process with just one line of code change.

  • They also have made a port for Bayesian methods for hackers

  • and it's available on their website.

  • So go and take a look.

  • So for our next question, from Siby,

  • "Can I create custom layers through tf.keras?"

  • And yes, absolutely you can.

  • Do you want to go a little bit more into that, Alex?

  • Yeah, so you just do the same thing you'd do

  • if you were using keras-team/keras

  • So you can inherit from the layer class and make your own layer.

  • Same way you can inherit from the metric class

  • and make your own metric.

  • Or the same thing for the losses.

  • So shipping with a default set of layers, losses, metrics, optimizers, models--

  • all the things that you're already used to from Keras--

  • but we also make it easy to extend, right?

  • Absolutely.

  • So if you want to add your own custom loss function,

  • your own custom layer, you can contribute it

  • to the open source community through TensorFlow add-ons

  • and we would love to have you submit it as a PR.

  • Yeah, thanks for the question, Siby.

  • Our next question, from Sharavsambuu, is

  • "Will the Keras namespace be removed in future releases of TF 2.0?"

  • Please tell me no.

  • Of course not.

  • TF 2.0 has a stable public API.

  • Not just Keras, but any symbol that is in TF 2.0 API

  • is going to stay there until at least TF 3.0.

  • And this even includes the symbols that we removed from TF 1.0

  • that are still available in tf.compat.v1

  • So if you're using anything from TF 2.0 now,

  • you can keep using it,

  • for as long as there is a TF 2.X at least.

  • Absolutely, and I am super delighted to hear that.

  • Yeah, thanks for the question, Sharavsambuu.

  • And out next question is also from Siby, who asked,

  • "Can we use SavedModel for a Keras model?"

  • Yes.

  • You can take a Keras model and export it to the SavedModel format

  • and then use it with the entire rest of the TensorFlow ecosystem

  • like TF Lite, TFJS, TF Serving.

  • If you have a Keras model that was saved to SavedModel,

  • you can load it back into Python and get a full Keras model back

  • with exactly the same API.

  • So it's completely seamless and fun to use.

  • Absolutely, and SavedModel really is a first-class citizen

  • as part of TensorFlow 2.0.

  • So we're trying to make it as easy as possible

  • for you to interact with your models and to export them

  • to any location that you would like to have them,

  • to any sort of platform, any kind of device.

  • Yeah, so thanks for the great question, Siby.

  • So if you have anymore questions, please use #AskTensorFlow

  • on social media and we'll answer it for you

  • in a future video.

  • Excellent, thank you so much to everyone who submitted

  • the questions that we had for this episode

  • and we'll see you next time.

  • ♪ (ending music) ♪

♪ (intro music) ♪

Subtitles and vocabulary

Click the word to look it up Click the word to find further inforamtion about it