Hands-on TensorBoard (TensorFlow Dev Summit 2017)

Join Dandelion Mané in this talk as they demonstrate all the amazing things you can do with TensorBoard. You'll learn how to visualize your TensorFlow graphs, monitor training performance, and explore how your models represent your data. The code examples shown are available here: goo.gl/ZwGnPE.
Visit the TensorFlow website for all session recordings: goo.gl/bsYmza
Subscribe to the Google Developers channel at goo.gl/mQyv5L
event: TensorFlow Dev Summit 2017; re_ty: Publish;

[@wenzhenzhu8895]

There is a bug: at 2:19, conv2 = conv_layer(pooled, 32, 64) pooled is undefined. It should be pool1.

[@chrisdahms9682]

everybody please vote this to the top, this is definitely a typo in the presentation

[@oneforallah]

Damn Virginia you beast ! You need to be hired by Google or someone good

[@tejasmahajan]

That was an amazing talk!! The debugger option in future upgrades is so promising!!

[@sankothari]

Awesome presentation skills! Wish you had a tutorial series.

[@moanayoon3411]

Thank you for the nice talk, Tensorboard is really something!!

[@user-po3to3zw1k]

Great talk! I like his talking style. My favorite one.

[@terrysatou3344]

Great talk, very clear and informative. The nail polish is cool!

[@firlejczyk]

Thanks for sharing this presentation and code to the community. TF is amazing....

[@yonglelyu4117]

great talk!

[@MdNasirUddinLaskarBD]

Awesome lecture and demo on TensorBoard.

[@cosmos58398]

Wonderful talk!

[@appendix77]

Super awesome! Just what I desperately want right now.

[@chonglv9766]

Benefit a lot from this great talk.

[@robertschneiderman5894]

20:34-20:37 is my favorite!

[@MatthewKleinsmith]

"As we zoom into this particular seven, it really does look like a one. It's probably the one-iest seven that I've seen."

[@frankwitte5170]

20:58 what's that shortcut? the go thingy.

[@heihank9428]

Nice talk as well as his nail-polish 💅

[@yunlongsong7618]

Very nice talk.

[@22691711]

great work guys :)

[@donamincorleone]

Great tutorial! Thanks

[@bashhwu]

Great job!. Thank you very much

[@rohitswami1]

Fantastic!

[@EugeneKrevenets]

Cool! It was very impressive.

[@g0d182]

Tensorflow; the sexiest machine learning library alive

[@janvonschreibe3447]

Why so ?

[@youbrajrai3000]

one of the best tool for machine learinig

[@alesolano5507]

Love it!

[@EvanZamir]

The album you referred to is "Unknown Pleasures" by Joy Division. (In case anyone wants to look it up.)

[@channerduan2797]

I love it!

[@diaozhatushuaide1287]

very nice talk for tensorboard ans it is very interesting

[@MausamJain]

Great explanation !! Excellent tutorial for those starting TensorBoard for the first time !!

[@iliemihai949]

Here is a link to slides: gist.github.com/dandelionmane/4f02ab8f1451e276fea1f165a20336f1 One of the best talks on TensorFlow Dev Summit :D

[@soilblack1992]

Excuse me, Do you have a copy of the slides. Above link seem be broken!

[@spenhouet]

Watching this at the end of 2018, nearly two years after this presentation and tensorboard is nearly the same and didn't evolve a bit. All the features described at the end of the video are still missing. So what did they do in those 2 years? Is there a total overhaul of tensorboard coming oder do I miss something?

[@dimitriosmallios5941]

I have to agree with you. Tensorboard is a big advantage for tensorflow as opposed to other DL libraries. I guess they have focused on tensorflow 2.0 where eager execution will be set by default. That will be amazing but I guess it has a lot of work on the way to be done.

[@user-nq6xr6gb4u]

Good job!

[@leitroy2010]

awesome tool!

[@gavinfang4176]

Amazing！！

[@inigoreiriz1299]

super useful!

[@JohirulIslam1270]

Is there any update on the integration of TF Debugger with Tensor Board?

[@tomatosauce9561]

Hey, how would I add in the test accuracy functionality? Thanks in advance :)

[@timrault]

11:52 Could someone explain why the gradient of the loss with respect to the parameters is always 0 when the parameters are 0 ? I'm not sure to understand why setting the weights and biases to 0 is actually an error. Thanks ! :)

[@nectismanta8080]

It's the "dead neurons", when a neuron/node has 0 weight and gradient is 0 too, it has no impact on the network and has zero response to the gradient descent learning process, so it's weight value cannot change and stays zero, it's necessarily "dead". It doesn't happen to all 0 weight neurons, but it happens to neuron/nodes with some certain settings -- neuron structure and activation function (frequent issue with ReLU ). Machine Learning textbooks should contain more detail.

[@citiblocsMaster]

at 2:17, where does variable "pooled" come from? (line 8)

[@chrisdahms9682]

on that line, "pooled" should be "pool1"

[@michaelmcgrath9653]

I see the next talk is on layers, is it possible to combine TensorBoard and layers?

[@TheGodSaw]

Tensorboard is amazing I use it all the time. But to be honest. It fits much better with a dynamic language like Pytorch. I can very much recommend to check out TensorboardX which is usable with Pytorch

[@mauricioalouan1323]

Hi, Great talk but I am getting very frustrated trying to follow through basically because i believe I am missing some imports (or some other part of the code). Can you point out all imports the are needed please. thanks

[@sudip2818]

Sudip Das 1 second ago Thank you for your tutorial, it is really great, But I am facing a problem, suppose I want to write "CONV (5x5)" which will not support by name_scope() method, is there has any other method that can help me to write??

[@snehotoshbanerjee1938]

Great talk!.The below link is not valid.Any new link to get the slide and code?

[@robertjarvis6145]

Very cool :)

[@liweidai4474]

Where is the slides? Only found the code, please add a link to the slides. Great presentation.

[@rosi.j3485]

dear Dan, I am having a problem with my tensorboard i am doing all the steps you have mentioned but the scalar summary is being inactive along with all the other summaries!!! do you have any ideas on why this is happening i searched everywhere and tried wverything and it is not working

[@hossein7031]

awesome

[@faizanshaikh5326]

Are slides available for this talk?

[@leopardsaga]

Anyone knows the website in the Chrome tab `sigma/danmane` at 17:35 ?

[@aditya8404]

He has "Nail-Polish" ... But the talk was so interesting that I noticed it when I rewatched this :P

[@equitytruthandjustice8309]

Indeed. Disgusting. When the world makes an identity from an abomination.

[@17999msl]

13:00 It was Joy Division ;P

[@lflee]

Dance, dance, dance, dance, dance, to the tensorflow 🐱‍🐉

[@cyrilanderson]

Was going to say the same thing. Ha.

[@bonesfelipe93]

Hahahahahahaha

[@JunkerJames]

exceptional

[@wolfgangplato8593]

Is there anyone here than can help me with implementing tensorboard on a mac to visualize my automatic punctuation system's output?

[@user-sv1vt6yn6e]

i tried the githubwebsites and find the png can't be downloaded......so sad

[@rahulkumarjha2404]

I know it's too late but that slide link is not working so if anyone had saved the slide on the drive or somewhere else then please send the link.

[@martinboelle2658]

13:37 explaining the distribution charts, he states that one line stands for the 90th percentile and another one stands for the 60th percentile. How to know which percentiles the individual lines stand for?

[@elliotwaite]

The lines are: [maximum, 93%, 84%, 69%, 50%, 31%, 16%, 7%, minimum]. These percentiles correspond with the percentiles you would get by taking half standard deviation steps away from the mean on a normal distribution: [maximum, μ+1.5σ, μ+σ, μ+0.5σ, μ, μ-0.5σ, μ-σ, μ-1.5σ, minimum]. So they give you a sense of the variance and shape. If the lines on your chart are evenly spaced (aside from the maximum and minimum), your distribution is probably close to a normal distribution. And if they are not evenly spaced, you can get a sense for how your data's distribution differs from a normal distribution. Source: github.com/tensorflow/tensorboard/blob/master/README.md#distribution-dashboard

[@martinboelle2658]

Elliot Waite Awesome, thank you!

[@isaacbamidele9899]

did he had is finger painted?

[@KhuyenVyLog]

Great talk! I can't go to link at 23:10. Someone help me!

[@dqvinh87]

There is no information about x-axis and y-axis in TensorBoard (what do they represent, what are the value units, ...) Is there some available ways to display these information? Thanks

[@damienstanton]

You don't need axis labels because each visualization always represents values over time

[@dqvinh87]

Hello Damien It's true, but I am not sure if I understand you correctly. Indeed the graph represents values over time, but at least we should know the unit of time (second or minute ...) and details information about y-axis also.

[@sheingng1495]

It does display the axis units, you have to click into the graph to display it. Try it out on tensorboard ;-)

[@martinboelle2658]

I agree on that it's not always obvious what the charts represent. Especially the distribution charts (13:37) could use a legend. @Stanley Ng: Clicking in the graphs doesn't make axis labels appear for me, it just increases the resolution.

[@frankwitte5170]

20:58 what's that shortcut? the go thingy.

[@elliotwaite]

It's a short link only available on Google's internal network. Employees can register "go/..." URLs that when navigated to on the internal network, they act as redirects to another URL. They make it easier for employees to recall and share URLs. More info about them can be found on this GitHub page that offers a similar service: github.com/kellegous/go

[@JadePX]

The jacket though

[@anyonetherealready]

sidebar histogram visualizations -> good old windows media player's visualizations

[@deepschoolai]

github repo for this?

[@russelldicken9930]

github.com/decentralion/tf-dev-summit-tensorboard-tutorial

[@ralphrazor788]

TensorBoard is a godsend

[@LengErrong]

at 2:17 in this video, a little mistake. conv2 = conv_layer(pooled, 32, 64) should be conv2 = conv_layer(pool1, 32, 64)

[@saeidghafouri8501]

link not working

[@madhuriramachandran7582]

OMG!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!

[@brandoblues]

joe macmillian, i see such a resemblance between the speaker and the fictional character.

[@AvtarSingh-ml3yx]

He has already loaded the graph . But the thing is HOW to load the graph Which file to run amongst all those created files after the retraining process. No clue regarding that ! How man !

[@derfer12345]

You need to log the directory using "tensorboard logdir=[path to directory]". If you have correctly set your FileWriter, all your summaries should be stored in that directory and Tensorboard will display them correctly.

[@eilaarich-landkof7044]

Hello, can someone help me view tensorboard on my machine while the graph and execution is happening on a GCE. I tried the following with no success: on the local machine, i run: gcloud auth login on the console: gcloud compute firewall-rules create tensorboard-port --allow tcp:6006 On the Mac, port forward configutaion: gcloud compute ssh datalabvm-pgp --project project_id --zone us-east1-b --ssh-flag="-R" --ssh-flag="6006:localhost:6006" on the datalab machine: !tensorboard --logdir="./logs" --debug on the local machine (MAC) IP:6006 the result: This site can’t be reached 35.190.179.233 took too long to respond. Search Google for 190 179 233 6006 ERR_CONNECTION_TIMED_OUT what am i missing? Many thanks!!

[@richarddecal9215]

Here's a direct link to the tensorboard code from his slides gist.github.com/dandelionmane/4f02ab8f1451e276fea1f165a20336f1#file-mnist-py

[@HoraceKLai]

I can't get tensorboard to open up >tensorboard --logdir /graphs 'tensorboard' is not recognized as an internal or external command, operable program or batch file.

[@dzhiurgis]

Hey its that the guy from Suraj videos? Also he has red fingernails!

[@Suldok]

Yep, looks like it!

[@TienTaioan]

Which video you're talking about? I'm curious.. LOL

[@robinranabhat3125]

channel "Siraj Raval" run by Siraj Raval .Turns out , the presentor is actually roommate of siraj

[@TienTaioan]

Thank you! Well perhaps I should start looking for a roommate

[@henryjulianmurillogutierre3421]

is still too complicated to use tensorboard, this example is the one in tensorflow web (perhaps), but extreme dense to follow, maybe could be something like google charts documentation for ease of understanding, the source does not work as expected without a few adjustments, i know this was a showcase of tensorboard and they need to break peoples eyes (amaze) with its tools and functionalities, but come on!! regular (not red nail devs) people need a dummies version, its funny that almost every video regarding tensorboard has the same example, so apparently no one has ace this tool dan dandelion and suraj

[@ottomondo1368]

It was Joy Division not Joy Wave... Hoodie and watchcap do not a cool kid make...

[@WeJustDrive]

yo has red nails

[@fuu812]

those nails