Get started with TensorFlow's High-Level APIs (Google I/O '18)

High-level APIs like tf.keras enable developers to train models easily and effectively. This session will introduce these APIs, and notebooks you can run live in the browser to get started using Colab. We'll walk you through writing your first neural network in TensorFlow using just 10 lines of code with tf.keras, and then we’ll introduce you to Eager execution. We'll close with educational resources you can use to learn more about ML. By releasing easier and more intuitive APIs, we hope to make TensorFlow, an open-source machine learning framework more accessible for all.
Rate this session by signing-in on the I/O website here → goo.gl/fZTwce
Try TensorFlow with zero install → goo.gl/NrJAEz
Train your first neural network with just 10 lines of code → goo.gl/6SRkzf
Use the same Keras-compatible API with TensorFlow.js! → goo.gl/ZBbzJH
Learn more about ML → goo.gl/36baeH
Follow Josh on Twitter → goo.gl/WQ4EFs
Follow Laurence on Twitter → goo.gl/ZsvMuf event: Google I/O 2018; re_ty: Publish; product: TensorFlow - General; fullname: Laurence Moroney, Josh Gordan; event: Google I/O 2018;

[@bakanndy]

This is the best free course on AI I've ever seen

[@melihaslan9509]

This is great introduction to Tensor Flow and Keras. Thank you everyone and Josh...Keep up the great work!

[@AkashKamal1998]

This video opens me into a wide range of information. Thanks a lot.

[@wakeatmethree4023]

This guy loves his coffee.

[@laurencemoroney655]

Haha! I couldn't get it off him! :)

[@wakeatmethree4023]

Both of you were great, thanks a lot :-)

[@rebeccacaley5706]

Security coffee

[@RadioNul]

I kept expecting it to be a part of the presentation

[@surajkushwah3221]

I was noticing this for long.

[@koushik7604]

You're just awesome Josh. Very helpful and informative. Thanks a lot.

[@LaurenceMoroney]

What am I? Chopped Liver? Hehehehe

[@fateriddle14]

Wow, this is gold. I love google and how they enable us with less of CS background to go in and get stuff done (or at least having fun trying)! Thank you for this talk. Btw, where's the place cool ppl in ML/AI hanging around? Would love a list of ecosystem.

[@sbarter]

Fate Riddle stackoverflow datascience group

[@nceevij]

Fun started when Laurence got in, i can connect well what he shared related to devs transition to ML .. Awesome intro video. Pycharm is awesome was looking something like that since coming from VS background

[@LaurenceMoroney]

Haha! Thanks! :) Glad you enjoyed!

[@simplegupta7154]

Finally there it is .. the most awaited video

[@laurencemoroney655]

Thanks!

[@briankenny94]

Laurence Moroney tensor flow is awesome Laurence. I am from Dublin, I think you are too, I can pick up the accent mixed in there. I was wondering if you could do a tensor flow video on career paths with tensor flow or something along those lines

[@simplegupta7154]

Brian Kenny Thanks Brian. 🙂 I'm from India. I would be nice to work together

[@LaurenceMoroney]

Hey Brian! Close. I grew up between Drogheda and Cardiff, Wales. :) Good idea about career videos. Let me look into it! :)

[@thedanzjl]

Very well explained. Thank you so much!

[@yuehernkang]

wow thanks! great talk. I am just getting started with TensorFlow and this talk really helped me alot.

[@prithwishdas8294]

Excellent presentation and concepts were covered well through Hands-on codes.

[@deepwebtube]

Josh, you are the God of ML!

[@thecandel5479]

Always, your's clips are nice! 🌹

[@bruhaspati560]

tysm fo this video its really simple and gr8!

[@user-xw6kn5vl6b]

Thank you for your recommending.I appreciated.

[@MarshalHayes]

Great video guys!

[@northernmun]

Hello from Anjin Games in sunny Hove UK. Nice demos +1 on the pose estimation. This is enabling my animation work at the moment.

[@laurencemoroney655]

Great! And say Hi to Brighton for me ! :D

[@northernmun]

Laurence Moroney, the world is built on a wall... LOL will do.

[@baselghaibour6948]

Thank you, Laurance

[@madmenhitbooker3383]

How to show tensorflow,keras,colab,how a notebook works, mnist, eager exec and cnn in 30 minutes. Bravo

[@laurencemoroney655]

Thanks!! :)

[@anupamgh0sh]

Eager to use eager :)

[@VladyVeselinov]

Stellar talk.

[@2009BDiddy]

nice gremlins reference

[@LaurenceMoroney]

hehehehe

Can this method use keras layers with CTC? how to implement? can you give me a sample example?

[@wafi0123]

how will be development of a hybrid programming language such as compiler and interpreter run from same platform with a toggle shifter .So IDE can process bothe from same platform when nessecery .

[@aben8763]

this guy is amazing !

[@monstrimmat]

Watched it live during my lunch hour. Was starving, but totally worth it. Now, I'm on colab, and I just tried doing "for image, label in dataset" in eager mode, but it's throwing me a "not an iterable" error. Are they using a newer version than what I have available, or is there just a little more omitted code there?

[@monstrimmat]

stackoverflow.com/questions/49658802/how-can-i-use-tf-data-datasets-in-eager-execution-mode (found an answer, it may well work as they use it in a newer version, but for now there's just one more line involved)

[@akhileshpandey123]

you should upgrade your tensorflow to 1.8

[@randomforests]

Yup exactly. Try !pip install -U tensorflow==1.8.0

[@siloquant]

always awesome!

[@laurencemoroney655]

Thanks!

[@pardhabandaru1105]

Can someone point me to the talk he is referring to at 6:36

[@cuauhtemocjimenezhernandez5778]

I tried to test the neural network in 10 lines, but the code has the following error, ValueError: Error when checking input: expected dense_input to have 2 dimensions, but got array with shape (60000, 28, 28) I think that some code is missing. Please send me the rigth code

[@hamidsayeed9477]

Awesome!

[@fahemhamou6170]

تحياتي الخالصة

[@simplegupta7154]

I want to work with some real time projects .. is there any opportunities?

[@JBGordon]

One of the coolest real-time projects I've seen is PoseNet in TensorFlow.js: medium.com/tensorflow/real-time-human-pose-estimation-in-the-browser-with-tensorflow-js-7dd0bc881cd5

[@NtandoMbele]

what course? he referred to a course. Please drop link. Thanks

[@dtracers]

It is not just pip install tensorflow for python.... Not if you are using the GPU version

[@laurencemoroney655]

Yep -- when using your own deve environment, there's a different install. And GPU needs NVidia stuff installed too. We have a couple of videos and blog posts about that, hope they help :)

[@dtracers]

Laurence Moroney I have definitely got it done but I work on a project with people that are just starting coding and sometimes get tripped up when an error. Happens. Especially matching versions of GPU installs to tensor installs

[@laurencemoroney655]

Agreed, it can be really tricky. If you are learning, I would recommend just going non-GPU with pip install TensorFlow to avoid any gotchas with Cuda CuDNN drivers etc. As you get more experienced with developing in it, then it's worth looking into GPU-based training.

[@DrKhan-hd4cd]

Use Linux/Ubuntu and Anaconda and all you need is to type this in terminal: conda install -c anaconda tensorflow-gpu

[@sbarter]

Installing GPU was Google's form of a gang initiation. Once you survive the 2 day beating you are in the club.

[@ravishankar2180]

FRANÇOIS CHOLLET, the author is.

[@bingeltube]

Recommendable

[@vaibhavshrivastava1781]

I'm facing an issue with the code :( "Error when checking input: expected dense_input to have 2 dimensions, but got array with shape (60000, 28, 28)" can someone help ?

[@alice_in_wonderland42]

It is telling Use a two dimensional array. Just Read The Docs amd check out the function defnition.

[@siirinsahibi]

I exactly get the same error, there is a size mismatch but I can’t figure out what to change yet. I have x train and x label, 60000 is the number of images in mnist and 28 is the pixel amont in x or y axis. What is the connectin between??

[@sanjaykrish8719]

the coffee got cold

[@muhamadabdulsamad2815]

Could anyone give me the link of that manning book of deep learning python!

[@danielm793]

www.deeplearningitalia.com/wp-content/uploads/2017/12/Dropbox_Chollet.pdf

[@aleksanderlukashou137]

Quick question based on what you've said: can I get data if I feed rules and answers? And what is the name of this part of science? :)

[@YasirJan]

Aleksander Lukashou that would be religion

[@babyzone7208]

is the tensorflow use only for deep learning?

[@geoffreyanderson4719]

no. tf works for gradient descent which is more general.

[@user-xo8wu2rp6e]

where i can get that book(deep learning with python)

[@danielm793]

www.deeplearningitalia.com/wp-content/uploads/2017/12/Dropbox_Chollet.pdf

[@adaptivalgorithm2043]

To everybody interested in the theory behind NN's, i highly recommend Haykins "Neural Networks and Learning Machines". Its available for free, just google it. ~900 pages of concentrated knowledge

[@moorthysathishkumar1529]

how to implement code for moving object detection in tensorflow using CNN and python language

[@JBGordon]

Here you go! github.com/tensorflow/models/tree/master/research/object_detection

[@sbarter]

Have fun installing the proto builder :)

[@Datvovan]

For anyone who wants to play with the keras-mnist example : colab.research.google.com/github/CaoManhDat/100DaysOfML/blob/master/tf-high-level/keras-getting-started.ipynb

[@liangli6276]

come on tensorflow

[@mayanksoni1997]

Sir docker is not installed pls help

[@johnsoprych483]

Where is the code for the example he used? :( either show all code OR post a link to it.

[@Theasstasticvillain]

Where is the code from the number classifier?

[@Theasstasticvillain]

As well as the dataset?

[@geoffreyanderson4719]

It's great at 25:40 there is eager execution shown, except you forgot to feed the data from it to any model. SHow us how to do that without a session, as you claimed is possible. All you showed is reading the input files. That's nothing.

[@eVul6]

20:48 Does he say 'air'?

[@Gameboygenius]

eVul6 he said "error".

[@pauldacus4590]

20:08 I've never heard of a "epoch sweet spot" for training... always heard more is better.

[@sbarter]

Paul Dacus Think of it like a ball rolling down a hill. It is possible for that ball to roll out of the "sweet spot" and end up further away from the global/local minimum.

[@pauldacus4590]

Right, I guess his example is guitar tuning, and he states that once you reach a "tuned" state that continuing to tune it will make it go out of tune, which I suppose is true if you are only tuning with constant rotation amounts on the tuning keys and only in one direction. This is sort of unrealistic in deep learning cuz learning rates and back prop take care of both these problems, and for the most part error rates will monotonically decline in a properly functioning deep learning model. If I saw an error graph like the one he is showing, I would know something is terribly wrong with the model...

[@shannonmusic2780]

Well you definitely taught my family how to turn me into their project.. which was not fun at all...

[@speedbumphu]

The files are gone.

[@siirinsahibi]

speedbumphu +1 Trying to test the code but missing

[@gioabashidze6960]

i though that keras and tf was different frameworks

[@2112jonr]

Keras has been integerated into Tensorflow - raw TF code is tedious to write and understand. Keras abstracts it into something more intelligible, and allows other ML frameworks to be plugged into the backend without complete code re-writes.

[@SirajFlorida]

Why does tensor flow hate Java so much?

[@sbarter]

Siraj Florida discrimination in the tech industry is a hot topic these days.

[@LaurenceMoroney]

Ummm....I can't speak for everyone else, but I *love* Java.

[@otroleonarbe]

Good presentation Next time leave the coffee cup on the lectern. It distracts the audience, also you're limiting to engage more the audience

[@mayanksoni1997]

20:48

[@misterxxxxxxxx]

bad guitar tuning analogy :)

[@mohitgupta4539]

Nice shirt.

[@geoffreyanderson4719]

"So when you are building a neural network, start as simply as possible and build from there." No, sorry, you need to develop a model for low bias first. It's a separation of concerns that's paramount during ML project management. There are too many knobs otherwise, as well as model architectures, like number layers, units per layer, and upteen other choices. There are a hundred hyperparameters, so what are you going to do, grid search until the cows come home? No you need to be intelligent. After your model has capacity to get low bias on the training set, then you work on the overfitting as a separate task if you want to remain sane. More data examples, more regularization, dropout, batch norm, initializations and network architecture adjustments are next as you monitor the gap between error due to bias on training set, versus error due to variance on a holdout set (dev set). You certainly do not start out simple, which gets you nothing.

[@RIMJANESSOHMALOOG]

powered by coffee

[@DataSlayerMedia]

Raspberry Pi TensorFlow Object Detection - kzbin.info/www/bejne/sKKsc5ababF-rpo

[@Loppy2345]

TLDR anyone?

[@paulstevenconyngham7880]

please

[@gebiete]

Now you can keras inside your tensorflow while you're tensorflowing.

[@tonycatman]

Am I the only one who loathes and despises Jupyter ? I have no idea where the cursor is going to end up, or what the system is doing. It's also full of indecipherable memory leaks and unwanted object caching.

[@FuZZbaLLbee]

tonycatman but it does allow you to test multiple manipulations on the dataset in memory, without having to restart the code and having to wait a long time. Also there is a good way of documenting what is going on, instead of being restricted to just commented out text

[@sbarter]

I love Jupiter. It's all I code in now. So much easier than a IDE.

[@geoffreyanderson4719]

The slideshow claims that 80% or something of effort is getting the data. That's maybe correct. So then tell me why is another MNIST magical import done to demonstrate how to do get the data in this great new product release of TF? I have a different idea. Please explain how to use, with real datasets like IRIS or something else that is NOT built-in, something else that is NOT going to fit into RAM all at once, the tf.data.Dataset? Better than IRIS which is small, how about HIGGS.CSV? We do not EVER use built-in data for real world projects. My goodness another MNIST to load into memory all at once up front? This is dumber than dumb and does not match the introductory slides that claim getting the data is the hard work. Well it is in TF plenty hard work. Show us the real thing now. Use an external dataset of sufficient size and stop the toy nonsense.

[@joschk8331]

Seriously put down the coffe have some respect