Hands-on TensorBoard (TensorFlow Dev Summit 2017) Google Developers Indonesian English

Hands-on TensorBoard (TensorFlow Dev Summit 2017)YouTube↗ IA↗(long shot)
Science & Technology 305.4K views 3.2K likes 40 dislikes 15 Feb 2017

Google for Developers 2.7M subscribers United States YouTube↗

Click to Play Video

[MUSIC PLAYING]

Saya selalu melakukan demo langsung

DANDELION MANE: I always wind up doing these live demos,

jadi saya tak pernah sempat menjelaskan langsung dari <i>slide</i>

so I never get the privilege of just jumping straight

saya banyak berkutat dengan komputer.

into slides, and so there's some amount of computer wrangling.

Tapi [tidak terdengar]?

But [INAUDIBLE]?

Baik, nama saya Dandelion.

All right, so my name is Dandelion.

Setahu saya, cuma saya yang bertanda tangan emoji

I'm the only person I know who has an emoji signature.

yaitu Emoji dandelion dan emoji singa.

It's the dandelion emoji and then the lion emoji,

Tetapi saya tak memasukkanya ke <i>slide</i>.

although I didn't actually put this on the slide.

Saya di sini untuk membahas tentang TensorBoard.

And I'm here to talk about TensorBoard.

TensorBoard adalah alat yang sangat menarik.

So TensorBoard is a really exciting tool.

Ia dibuat oleh sejumlah teknisi di Google.

It's been built by a number of engineers at Google.

Ide dasar di balik TensorBoard, agar semua orang tahu

And the basic idea behind TensorBoard

bahwa jaringan neural terkadang bisa menjadi seperti kotak hitam.

is everyone knows that neural networks can sometimes be a bit of a black box.

Kami mau jadikan TensorBoard menjadi seperi lampu senter

And so we want to TensorBoard to be something

yang membawa Anda ke dunia TensorFlow yang membingungkan

of a flashlight, something that will let you take the confusing world of TensorFlow

dan mulai menyelaminya.

and start to dive into it.

Ia adalah alat yang cukup rumit.

Now it's a pretty complicated tool.

Banyak detailnya.

It's got a lot of bits and pieces.

Jadi cara saya menjelaskannya adalah dengan memberi contoh.

So the way that I'm going to explain it to you is by going through an example.

Kita ambil model MNIST yang sederhana.

So we're going to take a simple MNIST model.

Ada lapisan konvolusional dan <i>pooling</i> di atasnya.

It's going to have a couple convolutional and pooling layers at the top.

Ada dua lapisan yang tersambung sepenuhnya.

And then it's going to have two fully connected layers.

Lalu ia akan membuat prediksi.

And then it's going to spit out our prediction.

Kita akan membangun model ini bersama-sama dan mengenal kesalahan

And we're going to just build this model together and explore

dan cara pakai TensorBoard untuk lebih memahaminya.

what can go wrong and how we can use TensorBoard

Saya akan tunjukkan kumpulan kode.

to better understand it. So I'm about to show a bunch of code.

Saya akan membahasnya dengan cepat.

I'm going to go through it kind of quickly.

Kini tak perlu khawatir mengingat semua kode

You don't need to worry about getting all the code

karena akan diposting di GitHub saya sebagai sebuah <i>gist</i>

right now because it's posted on my GitHub as a gist,

juga kami akan tweet tautannya di akhir nanti.

and we'll also tweet out the link at the end.

Pertama, mari gabungkan beberapa lapisan di TensorFlow.

So to start, let's put together some layers in TensorFlow.

Ingin ada lapisan konvolusional.

So we wanted to have a convolutional layer.

Kita akan buat dua variabel.

So we're going to make two variables.

Variabel ini akan punya <i>weight</i> dan <i>bias</i>.

These will have the weights and the biases.

Kita akan terapkan operasi konvolusional

We're going to then apply the convolutional op,

kita akan terapkan reLU ke <i>convolution</i> dan <i>bias</i>

we're going to apply a reLU to the convolution

lalu mengembalikannya.

plus the biases, and then we return that.

Kita akan buat lapisan yang terhubung sepenuhnya.

Then we're going to make a fully connected layer.

Ini sangat sederhana.

This is even simpler.

Sekali lagi, kita punya <i>weight</i> dan <i>bias</i>.

Once again, we have weights and biases.

Kita akan lakukan perkalian dan pertambahan matriks reLu

We're going to do the reLU of the matrix multiplication

dan mengembalikannya.

and addition, and we return that.

Setelah ada beberapa lapisan, kita atur <i>feed-forward</i> dari jaringan kita.

Now that we have some layers, we can set up the feed-forward part of our network.

Kita punya penampung input yang kita ambil pada gambar dan label.

We have some input placeholders which we'll just take in our images and our labels.

Kita bentuk ulang gambarnya agar terlihat seperti 28/28 pixel

We'll reshape the image so that it actually looks like a 28

agar kita bisa menggabungnya.

by 28 pixels so we can convolve on it.

Mari mulai gabungkan lapisan kita.

Then we'll start putting together our layers.

Kita punya dua lapisan konvolusional.

We've got the two convolutional layers.

Kita ratakan, ada dua lapisan yang terhubung sepenuhnya.

We flatten it, we get the two fully connected layers.

Itu adalah keseluruhan jaringan kita.

And that's basically our whole network.

Akhirnya, kita harus menghitung <i>loss</i> dan mulai melatih.

And finally, we need to compute the loss and start to actually train things.

Kita akan gunakan entropi silang untuk fungsi <i>loss</i>.

So we're going to use cross entropy for our loss function.

Kita akan gunakan pengoptimal Adam untuk melakukan langkah pelatihan.

We're going to use an Adam optimizer to do the training

Agar praktis, kita akan rekam akurasi sehingga kita bisa mencetaknya ke konsol.

step. And just for convenience, we'll record the accuracy so that we can print that out to the console.

Akhirnya, perlu sedikit sentuhan untuk menghidupkan model ini.

So finally, we just need to do a little bit of work to turn this model on.

Kita akan memulai seluruh variabelnya.

We're going to iterate-- we're going to initialize all of the variables.

Lalu, kita akan melakukan 2.000 langkah latihan.

Then we're going to do 2,000 steps of training.

Tiap langkah, kita akan ambil [tidak terdengar] dari data latihan MNIST.

For each step, we're going to take an [INAUDIBLE] of the MNIST training data.

Terkadang, kita akan melaporkan akurasi agar kita tahu itu berfungsi.

Occasionally, we're going to report the accuracy

Akhirnya, kita akan melakukan langkah latihan.

so we can see if it's working. And then finally, we'll run the training step.

Kita sudah punya model.

So we have our model.

Let's turn it on.

Pada langkah nol, ia punya 100% akurasi.

At zero step, it has 10% accuracy.

Sepertinya, itu benar.

That seems about right.

Pada langkah 500, ia punya 12% akurasi.

At step 500, it has 12% accuracy.

Itu bukan yang kita harapkan.

Not exactly what we were hoping for.

Bahkan terus menurun.

And then it even goes down.

Bisa dilihat, model ini tak mempelajari apapun.

And we can see really, this model didn't learn anything.

Ia hanya memperoleh <i>baseline</i> 10% dengan tebakan acak.

It's just getting baseline 10% by random guessing.

Apa yang akan kita lakukan dengan model rusak ini?

So what are we going to do about this broken model?

Saya katakan bahwa kita punya senter, ayo nyalakan.

Well, I just said we have a flashlight,

Yang pertama kita lakukan adalah

so let's turn it on. And the first thing we might want to do

menggambar struktur komputasi dari grafik TensorFlow.

is visualize the computational structure of our TensorFlow

Kita menulis semua kode untuk menyetelnya

graph. Because we spent all this code setting it up,

bagaimana jika keliru?

but what if it's miswired?

Bagaimana jika kita tak menghubungkannya

What if somehow we just didn't make the connections

atau menghitungnya sesuai harapan?

or the computation we expected?

Kita butuh kelas agar kita bisa menulis data dari TensorFlow ke disk

So we're going to need something called-- we're going to need a class that lets us write data

sehingga kita bisa membacanya di TensorBoard.

from TensorFlow to disk so that we can read it TensorBoard.

Kelas itu disebut FileWriter.

And that class is called the FileWriter.

Utamanya, ia hanya menulis informasi yang akan digambarkan di TensorBoard.

It basically just writes any information you're going to visualize in TensorBoard.

Buat grafik visualisasi sangat sederhana.

And to visualize the graph, it's very simple.

Kita atur FileWriter.

We set up the FileWriter.

Kita berikan direktori.

We give it a directory.

Lalu kita tambahkan grafik sesi ke <i>writer</i> itu.

And then we add to our sessions graph to that writer.

Akhirnya, kita hidupkan TensorBoard.

And now finally, we just turn on TensorBoard.

Ia sudah terpasang di paket PIP

It comes pre-installed in the PIP package,

bisa ditambahkan di <i>command line</i>

so you can just, at the command line,

tulis <i>TensorBoard</i> dan beri <i>logdir</i>.

write TensorBoard and pass it a logdir.

Itu menghantarkan kita ke demo pertama yang berada-- di sini.

And that brings us to our first demo, which is here.

Tak punya skala atau informasi, tak punya gambar.

So it has no scale or information, has no images.

Kita hanya perlu memberi grafik.

The only thing we gave it is a graph.

Kita bisa menuju ke tab grafik.

So we can go to the graph tab.

Kita menemukan sesuatu yang lumayan membingungkan.

And we get something that's really quite confusing.

Saya tahu hubungannya dengan grafik yang kita buat.

So I can kind of see how this corresponds

Mungkin karena di sisi kiri

to the graph we made.

ada <i>convolution</i>, beberapa tambahan, reLU, <i>pooling</i>.

Because maybe we see on the left we've got a convolution, we've got some addition, we've got a reLU, we've got pooling.

Tapi ia memiliki seluruh bagian ini yang saya tak tahu sumbernya

But it also has all of these parts that I really don't have any idea where they came from even though I just

walau saya sendiri yang membuat grafiknya.

made the graph myself.

Terlihat di sini bahwa grafik TensorFlow sangat rumit.

So what we're seeing here is that TensorFlow graphs

Untuk membuat alat ini memberikan kita wawasan bermanfaat

are really complicated. And the way that we can get this tool to actually give us

kita harus beri sedikit sentuhan pada struktur grafiknya.

useful insight, we need to do a little bit of work to structure the graph.

Terutama, kita harus bersihkan grafiknya.

So essentially, we need to clean the graph.

Anda mungkin tahu bahwa semua grafik TensorFlow

Now you may know that every TensorFlow graph is based

berdasarkan sistem penamaan ini di mana terdapat operasi dan punya nama.

on this naming system where it has these ops and the ops have names.

Bisa dikendalikan dengan memberi nodus tersendiri

We can control the names by giving individual nodes

nama spesifik dan dengan mengatur cakupan nama

specific names and by setting up names scopes, which

yang menyebabkan semua grup operasi terhubung memiliki

cause all groups of related ops to have the same naming

struktur penamaan yang sama.

Mari aplikasikan ke model yang baru saja kita buat.

So let's go and apply that to the model we just built.

Pertama, di definisi lapisan, kita

First, in our layer definition, we're

akan mengambil beberapa operasi penting,

going to take a couple ops that are important,

seperti variabel, dan beri nama yang jelas,

like the variables, and give them explicit names,

seperti W untuk <i>Weight</i> dan B untuk <i>Bias</i>.

like W for Weights and B for Biases.

Kita akan aplikasikan cakupan nama

And then we're going to apply name scopes

ke semua lapisan jadi semua operasi konvolusional

to the entire layers so that all of our convolutional ops

tetap mengelompok dan juga di lapisan yang terhubung sepenuhnya.

will stay grouped and likewise in the fully connected layer.

Lalu, saat kita mengatur grafik

Then, when we actually set up our graph,

kita beri nama pada <i>stakeholder</i>.

we pass a couple of more names, like giving the placeholders

Kita gunakan fungsi ini untuk memberi cakupan nama ke lapisan kita.

names. And we use this function to give name scopes to our layers.

Terakhir, tambahkan cakupan nama

And finally, we add a couple of names scopes

ke pelatihan dan komputasi <i>loss</i> dan seterusnya.

to the training and the loss computation and so forth.

Setelah kita melakukannya, kita bisa membuat ringkasan lain

So now that we've done that, we can

di mana FireWriter menunjuk ke direktori lain.

construct another summary FileWriter pointing

Pakailah direktori berbeda dengan

at a different directory. You always want to use a different directory

TensorBoard agar data tetap terpisah.

with TensorBoard to keep data separated.

Kita hidupkan TensorBoard yang menunjuk ke direktori itu.

And we can turn on TensorBoard pointing in that directory.

Kini kita punya grafik yang jauh lebih baik.

And now we get a much nicer graph.

Lihat di sini, ini

So you can see here that this actually

mulai berhubungan secara bermakna dengan grafik yang kita pikirkan.

starts to correspond pretty meaningfully to the graph

Ada blok pelatihan ini, yang terhubung dengan semuanya.

that we had in mind. We have this training block, which

Itu karena ketika Anda menghitung gradien,

is connected to everything. That's because when you are computing the gradients,

Anda harus berhubungan dengan tiap variabel di grafik.

you need to connect to every variable in the graph.

Untuk sekarang kita pindahkan ke sisi

But maybe for now we'll move it to the side

karena itu tidak menarik bagi kita.

because it's not as interesting to us.

Kita bisa jelajah dan katakan, kita

And now we can kind of explore and say, OK, we

punya <i>placeholder</i> x kita.

have our x placeholder.

Bentuk ulang datanya.

We reshaped the data.

Lalu telusuri tiap lapisan konvolusional ini

Then we go through each of these convolutional layers

di mana ada beberapa variabel.

where we have some variables.

Kita punya <i>convolution</i> dan penambahan.

We have the convolution and the addition.

Kita bisa pindai grafik

And we can kind of scan through a graph

dan perhatikan semua yang dihubungkan sesuai harapan kita.

and see that everything is wired up the way that we expected.

Saya akan sebutkan fitur dari penggambaran grafik.

Now, I'll point out a few features

Masalahnya, jelas, Anda

of the graph visualizer.

bisa jelajahi lapisan berbeda dari tingkat abstraksi

One thing is, as you've noticed, you can explore different layers of levels of abstraction

dengan menyelam ke bagian-bagian tersendiri.

by diving into individual pieces.

Kita bisa menyelam lebih dalam.

And we can dive in further.

Di sini, kita peroleh mekanik

And here, we're getting the mechanics

dari komputasi entropi silang.

of computing cross entropy.

Rupanya semua yang terpotong ini

It turns out all of this sliced stuff

adalah untuk komputasi entropi silang.

was for computing cross entropy.

Anda akan lihat kotak yang berbeda

Also, you'll notice that different boxes

memiliki warna yang berbeda.

have different colors.

Cara penggambar grafik bekerja adalah ia

The way that the graph visualizer works is it

mengidentifikasi substruktur terulang dan serupa dari grafiknya

identifies repeated, similar substructures of the graph

dan memberinya warna yang sama.

and gives them the same color.

Bisa pastikan dua lapisan terhubung sepenuhnya ini punya

So we can be sure that these two fully connected layers have

struktur yang sama satu sama lain

the exact same structure as each other

karena mereka berdua berwarna hijau.

because they are both colored green.

Terdengar seperti itu bukan fitur yang begitu penting.

Now this sounds like it might not

Nyatanya, saat kita hubungkan

be such an important feature.

grafik insepsi ke TensorFlow

But actually, when we were porting the inception graph over to TensorFlow

satu setengah tahun lalu--

back a year and a half ago--

dan ini grafik yang amat rumit dengan puluhan ribu nodus

and this is a very complicated graph with tens of thousands

terhenti berminggu-minggu.

of nodes, we were stuck for weeks.

Dan hasilnya salah.

And we were getting the wrong results.

Kita bahkan berpikir mungkin TensorFlow rusak.

And we actually thought maybe TensorFlow was just broken.

Mungkin setahun untuk mengembangkan software baru ini

Maybe we had spent like a year developing this new software

dan itu takkan berhasil.

and it just wasn't going to train.

Rupanya masalah sebenarnya ialah salah satu submodul

And it turned out the real issue is that one of the inception

insepsi memiliki sedikit <i>bug</i>.

submodules had a slight bug.

Itu terhubung sedikit berbeda

So it was wired slightly differently,

dan itu merusak pelatihannya.

and that was breaking the training.

Ada pengaturan lain yang bisa gunakan di sini.

There are a couple of other settings you can use here.

Contohnya, kita bisa perhatikan pewarnaan berdasarkan perangkat.

So, for example, we can look at coloring by device.

Semua diletakkan di perangkat yang sama jadi semua abu-abu.

In this case, everything was put on the same device

Tapi jika pakai campuran CPU dan GPU,

so it's all gray. But if you were using a mix of CPU and GPU,

Anda bisa memakainya untuk tahu di mana meletakkannya.

you can use that to get an idea of where things were placed.

Kita juga bisa hidupkan input jejak,

We can also turn on trace inputs,

lalu bisa pilih potongan apa saja

and then you can choose any individual piece

dan lihat <i>dependency</i>-nya di dalam grafik.

and kind of see what all of its dependencies are in the graph.

Semua dipudarkan.

Everything else gets grayed out.

Melihat ini, setidaknya kita punya grafik yang kita harapkan.

So yeah. But from looking at this, it does look like we have, more or less, the graph we expected.

Tapi sebenarnya, itu bukanlah pelatihan.

But as we saw, it's not training at all.

Waktunya mendapatkan lebih banyak informasi

So it's time to get a little bit more information and time

dan waktu untuk menggunakan tab lainnya.

to start using these other tabs.

Untuk melakukannya, kita harus menulis

Now to do that, we need to actually write

data, data <i>runtime</i>, dari TensorFlow ke disk.

data, runtime data, from TensorFlow to disk.

Caranya adalah melalui sistem yang disebut <i>summary</i>.

And the way we do that is through a system called summaries.

<i>Summary</i> adalah jenis khusus dari operasi TensorFlow.

A summary is basically a special kind of TensorFlow op.

Ia mengambil <i>tensor</i> reguler dari grafik Anda

It will take in a regular tensor from your graph,

lalu itu akan <i>output</i> penyangga protokol yang bisa kita tulis ke disk.

and then it will output protocol buffers

Ada beberapa jenis perbedaan dari <i>summary</i>.

that we can write to disk. Now, there are a couple of different kinds of summaries.

Paling mudah ialah <i>summary</i> skalar yang akan

The simplest is the scalar summary which will just

menuliskan satu nilai, dan Anda

write down a single value, and you

gunakan itu untuk membuat grafik garis bagus ini.

use it to create these nice line charts.

Lalu ada <i>image summary</i>, yang akan kita buatkan gambar.

Then there are image summaries, which we'll write out an image.

Itu untuk menggambarkan, misalnya,

And you can use it to visualize, for example,

jika Anda punya model generatif gambar,

if you have a generative model of images,

atau ingin periksa <i>input</i> data

or if you just want to check that your input

apakah sudah diformat dengan benar.

data is formatted correctly.

Kita punya <i>summary</i> audio.

We have an audio summary.

Misalnya, tim magenta

So, for example, the magenta team,

akan dibahas nanti, sedang mengerjakan

which we'll hear from later and is working

generasi musik banyak yang betul-betul

on a lot of music generation actually

ditambahkan <i>summary</i> audio ke TensorBoard.

added the audio summary to TensorBoard.

Ada <i>summary</i> histogram, dan ini

There are histogram summaries, and these

membiarkan Anda perhatikan distribusi bentuk dari nilai berbeda.

let you look at the shape of distributions

Sangat bagus mengedepankan variabel Anda, seperti <i>weight</i>

of different values. So it can be very nice to be attach to your variables, like your weights.

Akhirnya, kita punya TensorSummary.

And finally, we have the TensorSummary.

TensorSummary itu <i>summary</i> terbaru.

The TensorSummary is the newest summary.

Idenya ialah ia bisa membuat jenis nilai apapun

The idea is that it can write out any kind of value

karena semua di TensorFlow adalah sebuah Tensor.

because everything in TensorFlow a Tensor.

Kita belum tambahkan dukungan sepenuhnya untuk itu.

But we haven't quite added support for it fully yet.

Saya akan sedikit bahas saat kita

And I'll discuss it a little bit when we get

bicarakan rencana TensorBoard mendatang.

to TensorBoard's future plans.

Kita akan tambahkan <i>summary</i>.

So now we'll add a couple of summaries.

Ada dua <i>summary</i> skalar ini,

We had these two scaler summaries,

karena lihat bagaimana entropi silang dan akurasi

because it would be nice to look at how

berubah sepanjang waktu.

the cross entropy and the accuracy are changing over time.

Baik juga untuk lihat gambar <i>input</i> kita,

It would also be nice to look at our input images,

pastikan itu digit MNIST, seperti yang diharapkan.

just make sure that they're MNIST digits, like we expect.

Akhirnya, di tiap lapisan

And then, finally, in each of our layers

akan kita tambahkan histogram

we'll add a couple histograms that

yang akan menunjukkan <i>weight</i>, <i>bias</i>, dan aktivasi.

will show us the weights, the biases, and the activations.

Setelah itu, kita harus pikirkan,

So with that all said, we need to think, OK,

bagaimana cara menjalankan <i>summary</i> ini?

how are we actually going to run these summaries?

Jawabannya ialah menjalankan tiap <i>summary op</i>.

And the answer is that we run each summary op.

Memberi kita <i>buffer</i> protokol,

It gives us the protocol buffers,

lalu kita bisa oper ke FileWriter untuk menempatkannya di disk.

and then we can pass them to our FileWriter to get them to disk.

Akan lama jika kita harus

It would be really tedious if we needed

perhatikan 6 atau 7 <i>summary op</i> ini dan jalankan sendiri-sendiri

to keep track of those six or seven summary ops and run them all individually, especially if we'd

terutama jika kita ke grafik seperti insepsi yang punya ratusan

go to a graph like Inception that might have hundreds

atau ribuan <i>summary</i>.

or thousands of summaries.

Alih-alih, kita punya<i> Merge All summary</i> ini

So instead, we have this Merge All summary

yang membuat target tunggal.

that creates a single target.

Kita jalankan, dan dapatkan tiap<i> summary</i> di grafik.

We run it, and we get every summary in the whole graph.

Kita bisa membuat FileWriter lainnya.

So we can go and create another FileWriter.

Kita bisa ubah kode pelatihan

We can go and modify our training code

untuk jalankan <i>summary</i> yang tergabung dan menulisnya ke disk.

so that it will run the merged summary and write it to disk.

Akhirnya, hidupkan saja TensorBoard

And then, finally, we just need to turn on TensorBoard

yang menunjukkan direktori baru kita.

pointing out our new directory.

Kini, kita bisa mulai jelajahi yang sebenarnya terjadi di model ini.

So now we can start to explore what's really happening in this model.

Seperti yang terlihat, akurasi di angka 10%.

So as we saw, the accuracy bounces around 10%.

Di <i>backdrop</i>, Anda bisa lihat data asli sangatlah berantakan

In the backdrop, you can see the original data is very noisy

karena adanya sampel acak yang berpindah-pindah.

because it's just got random sampling moving around.

Bisa kita perhalus untuk melihat baris tren lebih jelas.

So we can smooth it to see the trend line more clearly.

Kita juga punya entropi silang.

And then we also have the cross entropy.

Ini lumayan menarik karena entropi silang

And this is pretty interesting in that the cross entropy is

tetap konstan sepanjang proses pelatihan,

staying exactly constant throughout the entire training

yang menunjukkan ada yang salah dengan modelnya.

run, which is just something is seriously wrong with the model.

Kita juga bisa lihat tab <i>Images</i>.

We can also take a look at the Images tab.

Di sini kita temukan, gambar kita sebetulnya

And here we find OK, our images actually

terlihat seperti gambar MNIST yang indah.

do look like fine MNIST images.

Kita bisa lihat ukuran pixel sesungguhnya

And we can look at a true pixel size

lalu diingatkan bahwa--

and then be reminded that--

ada yang salah.

something went wrong with that.

Tapi Anda juga memperluasnya.

But you also expand it.

Kita punya beberapa kapabilitas di sini.

So we have a couple of different capabilities here.

Hal yang kita anggap amat menarik

And what we'll see really interesting

ialah saat kita lihat distribusi dan histogram.

is when we look at the distributions and histograms.

Sebenarnya ini 2 cara berbeda

So this is actually two different ways

untuk menggambarkan informasi histogram yang sama.

of visualizing the same histogram information.

Distribusi grafik memberitahukan

The distribution chart is telling us

bahwa semua nilai terhenti di nol.

that all of the values were stuck at zero.

Grafik histogram akan memberitahukan hal yang persis sama.

And the histogram chart is going to tell us

Semua aktivasi ada di angka nol.

the exact same thing. All of our activations were one giant spike at zero.

Semua <i>bias</i> dan <i>weight</i> tertahan di angka nol.

And also, all of our biases and weights were stuck at zero.

Ini memberi tahu bahwa mungkin ada

So this gives us an intuition that something was probably

yang salah saat kita mengatur lapisan.

wrong when we set up our layers.

Karena saat <i>bias</i> dan <i>weight</i> semuanya nol,

Because when the biases and weights are all zero,

kita akan punya gradien nol melewati

we're going to have zero gradients passing

grafik dan tidak ada latihan sama sekali.

through the graph and nothing will train at all.

Mari lihat kode kita dimana

So let's take a look at our code for where

kita atur lapisan ini.

we set up these layers.

Saya akan lihat lapisan, misalnya konvolusional.

I'm going to look at the convolutional layer,

Bisa dilihat, sebenarnya kita mengatur <i>bias</i> dan <i>weight</i>

for example. And you can see that we actually set the biases and weights

untuk memulai dari nol, sebenarnya itu salah.

to initialize at zeros, which was a mistake.

Kita ubah agar punya inisialisasi yang lebih baik,

So we can change that to have a more sensible initialization,

seperti normal terpotong untuk <i>weight</i>

like a truncated normal for the weights

dan nilai konstan untuk <i>bias</i>.

and a constant value for the biases.

Kita bisa jalankan ulang model kita

And now we can just rerun our model

dengan pengaturan yang sudah berubah.

with the changed settings.

Mulai dapat hasil lebih baik.

And we start to get much better results.

Kita lihat di sini bahwa entropi silang secara cepat

We can see here that our cross entropy rapidly

berkurang ke nol.

decreases towards zero.

Akurasi kita secara cepat bertambah ke satu,

And our accuracy rapidly increases towards one,

atau akurasi yang sempurna.

or perfect accuracy.

Jika ingin lihat dua grafik bersamaan,

If we want to see these two charts at the same time,

kita punya kemampuan untuk menciptakan kelompok grafik

we have the ability to create groups of charts

menggunakan ekspresi reguler.

using regular expressions.

Kita bisa menaruh semua di satu grafik, di satu grup.

So we could put everything in one chart, in one group.

Kita bisa lihat bahwa jika kita cek ke histogram,

Similarly, we can now see that if we look at the histograms,

misalnya, kita sebenarnya punya distribusi yang baik sepanjang waktu.

for example, we actually have nice distributions over time.

<i>Sidebar</i> tentang penggambaran histogram ini.

Sidebar about these histogram visualizations.

Mereka dibuat oleh Shawn Carter yang

They were made by the really talented Shawn

amat bertalenta, dulu bekerja di "New York Times."

Carter, who used to work at the "New York Times."

ia terinspirasi jika pernah

and he was inspired if you've ever

lihat cover album Joywave, dimana itu

seen this album cover called Joywave, where it's

seperti bentuk gelombang bunyi meluas sepanjang waktu,

like an audio wave forms spread out over time,

ia tipe pria yang lihat cover album dan berpikir,

he's the kind of guy who sees an album cover and thinks,

itu akan membuat visualisasi yang bagus.

ah, that would make a great visualization.

Kita bisa lakukan hal-hal keren

So we can do a couple of cool things

di sini, seperti kita bisa mengubahnya dari <i>offset</i> ke <i>overlay</i>.

here, like we can switch them from the offset to overlay.

Jika Anda ingin lihat bagaimana berubah sepanjang

So sort of if you want to see how it evolved over time,

waktu <i>offset</i>-nya itu berguna.

the offset is useful.

Jika ingin lihat bentuk dalam absolut dan relatif,

If you want to see how it looked in absolute and relative terms,

<i>overlay</i> cukup berguna.

the overlay is quite useful.

Ada informasi yang persis sama

We have the exact same information

di grafik distribusi, dimana sekarang itu

in the distribution charts, where now it's

menunjukkan tampilan berbeda, kita bisa lihat,

showing us a different view, where we can see,

misalnya, untuk baris sangat samar ini ialah nilai maksimumnya.

for example, this very pale line is the maximum value.

Ada nilai maksimum sangat tinggi untuk aktivasi.

So we can see that there was a very high maximum value for the activations.

Tapi disini, kita lihat perseratus ke-90

But here, we're looking at like the 90th percentile

dan perseratus ke-60.

and the 60th percentile.

Itu sebenarnya lumayan ramai.

It was actually fairly closely clustered.

Yang penting ialah model kita akhirnya berfungsi

But the important thing here is that our model is finally

Kita bisa lihat itu berlatih dengan lebih baik.

working. And we can see that it's training a lot better.

Ada model yang baik, kita mungkin bilang,

So now that we have a good model, we might say,

oke, kita memilih tingkat pembelajaran secara arbitrer,

OK, well, we chose a learning rate kind of arbitrarily,

kita pilih model arsitektur kita secara arbitrer.

and we chose our model architecture kind

Bagaimana jika dilakukan secara berbeda?

of arbitrarily. But what if we had done things differently?

Jika coba tingkat pembelajaran berbeda

What if we tried a couple of different learning rates

dan bagaimana jika kita coba model arsitektur berbeda?

and what if we had tried a couple of different model

TensorBoard bisa sangat berguna untuk ini juga

architectures? TensorBoard can be really useful for this as well

karena memberi Anda cara bagus untuk bandingkan

because it gives you a good way to compare

versi model kita yang berbeda

how different versions of your models

menumpuk pada semua variabel yang berbeda ini.

stack up on all of these different variables.

Kita bisa buat perubahan ke kode kita.

So we can make a couple of changes to our code.

Kita ulang tiga tingkat pembelajaran berbeda.

We're going to iterate over three different learning rates.

Lalu kita akan ulang menggunakan

And then we'll iterate over using

satu atau dua lapisan terhubung

one or two fully connected layers

dan satu atau dua lapisan konvolusional.

and one or two convolutional layers.

Kita akan buat rangkaian unik untuk

And we're going to make a unique string

tiap pengaturan <i>hyperparameter</i> yang menjelaskannya.

for each of these hyperparameter settings describing it.

Misalnya, tingkat pembelajaran, garis bawah, 1e negatif 3,

For example, learning rate, underscore, 1e negative 3,

terhubung sepenuhnya = 2, lalu konvolusional = 2.

then fully connected equals 2, and then convolutional equals 2.

Kita akan buat penulis <i>summary</i> baru

And we'll create a new summary writer

yang menunjuk ke direktori yang punya rangkaian itu di namanya.

that's pointing to a directory that has that string in the name.

Sekarang kita bisa menjalankan modelnya

And now we can actually run the model.

Kita akan menunjuk TensorBoard di direktori tingkat awal.

And we're going to point TensorBoard at the parent level

Tadi, kita menunjuk

TensorBoard di direktori untuk<i> run</i> spesifik.

So in the past, we've been pointing TensorBoard at the directories for specific runs.

Sekarang kita ingin lihat awal <i>run</i>

And now we want to look at the parent of runs

sehingga bisa temukan semua data dan membandingkannya.

so that it can find all of this data and compare it for us.

Sekarang kita bisa ke demo selanjutnya.

So now we can go to our next demo.

Di sini, kita punya situs<i> run</i> ini di sisi kiri

And here, we have this run site on the left

punya semua pengaturan <i>hyperparameter</i> ini yang sudah kita coba.

has got all of these different hyperparameter settings

Kita bisa lihat, misalnya entropi silang kita

that we experimented with. And now we can look at, for example, our cross entropy

dan lihat bagaimana itu membandingkan

and start to see how they compare it

sepanjang <i>run</i> yang berbeda.

across the different runs.

Kita punya akurasi di sini.

Similarly, we have accuracy up here.

Bisa lihat, ada banyak variasi.

So we can see, actually, there was a lot of variance.

Beberapa tidak belajar banyak sama sekali,

Some of our models barely learned anything at all,

dan beberapa darinya sangat berhasil.

and some of them were extremely successful.

Jika lihat entropi silang, kita

If we look at cross entropy, maybe we

ingin membuatnya sedikit lebih jelas.

want to make things a little bit clearer.

Seperti pertamanya semua mula-mula tinggi,

Like the beginning everything starts high,

jadi itu menghitung ulang tampilan.

so it rescales the display.

Kita bisa perbesar, dan mendapat gambaran lebih jelas.

So we can zoom in, and we get things a little bit clearer.

Kita juga bisa <i>toggle</i> di skala logaritma, yang

We can also toggle on a logarithmic scale, which

membuat perbedaan sedikit lebih jelas dan terkadang berguna.

makes the differences a little bit more pronounced and sometimes very useful.

Bisa mulai jelajah dengan<i> tips</i> alat kita

And then we can start exploring with our tool tips

disini, dan lihat,<i> run</i> ini yang terburuk

here and seeing, OK, this run that was the worst

tingkat pembelajaran negatif 3, dan keduanya konvolusional

was learning rate negative 3, and then both convolutional

dan terhubung sepenuhnya.

and both fully connected.

Yang ini yang terbaik adalah sama

And this one that was the best was the same thing

tapi dengan tingkat pembelajaran lebih kecil.

but with a smaller learning rate.

Jika kita ingin lakukan perbandingan lebih jauh,

And if we want to do a more in-depth comparison,

kita bisa mulai menyaring dengan pengaturan disini.

we can start to filter by settings here.

Jika kita ingin lihat semua hal

So if we want to look at all the things

yang hanya punya satu lapisan konvolusional,

that had only one convolutional layer,

ketik <i>conv</i> sama dengan 1, dan kini

we could type conv equals 1, and now

lihat dengan satu lapisan konvolusional,

we can see just the ones that had one convolutional layer, so

sehingga membuatnya lebih mudah untuk menyelaminya.

that makes it a little bit easier to dive it.

Atau kita bisa mencari hal-hal yang hanya

Or similarly, we could search for only the things that

punya tingkat pembelajaran e ke negatif 4,

had the learning rate of e to the negative 4,

dan itu termasuk <i>run</i> terbaik kita.

and that includes our best run.

Beberapa fitur lain yang akan dibahas.

A couple other features I'll point out.

Anda juga bisa <i>toggle</i> satu<i> run</i> dengan klik disini.

You can also just toggle one run at of time by clicking here.

Kita juga bisa merubah tampilannya

And we can also change the display

untuk menunjukkan cara berbeda dalam menggambarkan waktu.

to show us different ways of visualizing time.

Sekarang, kita orientasikan sumbu-x dengan langkah.

Right now, we're orienting the x-axis by step.

Kita bisa lakukannya dengan, misal, waktu dinding.

But we could do it by, for example, wall time.

Kita bisa lihat, sebenarnya, sementara semua

And then we can see, actually, that while they're all

ada di langkah, dijalankan dalam waktu berbeda.

aligned in steps, I ran them at different times.

Kita bisa pakai waktu relatif,

We can also use relative time, so this

ini akan menunjukkan menit dari awal.

will show the number of minutes from start.

Kita lihat, sebenarnya, beberapa model menghabiskan

And then we can see, actually, that some models took a lot

lebih banyak waktu <i>run</i>.

longer to run than others.

Walau langkahnya sama,

So even though they took the same number of steps,

yang lebih terhubung sebelumya

the models with more fully connected

dan lebih konvolusional berjalan dua kali lebih lambat.

and more convolutional layers ran about twice as slowly.

Kita juga bisa menggambarkan-- semua tampilan lainnya

And then we can also visualize-- all of our other displays

juga beri lebih banyak informasi.

also give us more information.

Jika kita ingin lihat seluruh <i>bias</i>,

So if we wanted to look at all the biases,

kita lihat bagaimana mereka bervariasi untuk model berbeda.

we could see how they varied for different models.

Mari lihat, <i>conv</i> 1, garis miring<i> bias</i>, mungkin.

And let's see, conv 1, slash bias, maybe.

Itu cukup menarik bahwa model berbeda sebenarnya

And so that's kind of interesting

punya pola berbeda dalam hal<i> bias</i> berubah sepanjang waktu.

that we can see the different models actually had very different patterns for how the biases evolved over time.

Mungkin dengan menyelidikinya,

And probably by digging into this,

kita bisa tahu model mana yang sebenarnya

we could learn some ways in which the models are actually

berbeda secara struktur.

structurally different.

Di penggambar grafik,

Similarly, in the graph visualizer,

kita bisa menjelajahi <i>run</i> berbeda.

we can explore the different runs.

Kita bisa lihat sekarang kita

So we can see here that right now we

punya model dengan satu lapisan konvolusional

have one of the models that had only one

dan yang terhubung sepenuhnya.

convolutional layer and only one fully connected layer.

Tapi kita bisa membandingkannya dengan model berbeda.

But we could go and compare it to different models instead.

So the final thing--

fitur utama TensorBoard yang akan saya jelaskan

major feature from TensorBoard that I'm going to describe

adalah <i>embedding visualizer</i>.

is the embedding visualizer.

Itu mungkin hal yang terkeren dari TensorBoard.

And it's maybe the coolest piece of TensorBoard.

Yang pada dasarnya <i>embedding visualizer</i> lakukan ialah

What the embedding visualizer basically does

membiarkan Anda mengambil data dimensi tinggi

is it lets you take high dimensional data

dan memproyeksikannya ke dalam tiga dimensi.

and then project it down into three dimensions.

Cara yang mungkin menarik untuk melakukannya

Now the way in which we might be really interested in doing

ialah saat kita mengambil seperangkat data input kita

this is when we take our input data set

dan memetakannya sepanjang jaringan neural,

and we map it through the neural network, to say,

lapisan final, dimana <i>embedding</i> sebenarnya ialah

the final layer, that embedding is actually

representasi yang dipelajari dari cara jaringan neural memroses informasi.

the learned representation of how our neural network is processing the information.

Saya akan menulis kode di layar,

So I'm going to throw a bunch of code on the screen,

saya takkan bahas secara mendalam.

and I'm not going to explain in great depth.

Tapi saya akan beritahu tutorial bagus

But I will say there's a wonderful tutorial

mengenai <i>embedding visualizer</i> di halaman web.

on the embedding visualizer on the website.

Intinya, yang kita akan lakukan

But basically, what we're going to do

adalah kita akan buat variabel yang

is we're going to make a variable that

akan menahan <i>embedding</i> tiap seperangkat gambar tes kita.

will hold the embedding of each of our test set images.

Kita akan menciptakan objek konfigurasi yang

And then we're going to create a configuration object that

punya sedikit informasi, dimanakah kita bisa

has a little bit of information like, where can we

temukan <i>sprite</i> PNG dari semua gambar MNIST

find a sprite PNG of all the MNIST images

jadi kita bisa tampilkan<i> thumbnail</i> di penjelajah?

so we can display thumbnails in the browser?

Dimanakah kita temukan label

And where are we finding the labels

sehingga kita bisa mengorganisir contoh

so that we can organize the examples

label manakah mereka?

by which label they are?

Kita tulis semua ini ke disk menggunakan FileWriter.

And then we write this all to disk using the FileWriter.

Sekarang, setiap 500 langkah, kita akan

And now, every 500 steps, we're going

menyimpan <i>checkpoint</i> model.

to save a model checkpoint.

<i>Checkpoint</i> model ini berisi

And this model checkpoint contains

seluruh variabel di model kita,

all of the variables in our model,

termasuk variabel <i>embedding</i> itu.

including that embedding variable.

Akhirnya, kita hidupkan TensorBoard sekali lagi,

And then, finally, we turn on TensorBoard once again,

menunjuk ke direktori ini.

pointing to this directory.

Dan kita pun punya <i>embedding</i> kita.

And here we have our embedding.

Ini 1.024 poin dari MNIST yang saya putuskan untuk

This is 1,024 points from MNIST that I decided

pakai dari seperangkat tes.

to use from the test set.

Apa yang sedang kita lihat

And what we're looking at right now

ialah analisis komponen utama dari <i>embedding</i>.

is the principal components analysis of the embedding.

Jadi kita ambil tiga komponen utama

So we've taken the top three principal components

dan meletakkan tiga poin ini dalam manipol tiga dimensi.

and laid out these points in three dimensional manifold.

Jika kita warnai sesuai label, kita bisa

So if we color by label, we can already

lihat itu mempelajari separasi yang berguna, seperti semua

see that it's learned a pretty useful separation, like all

yang telah berpisah dalam salah satu dari komponen utama.

of the ones that have been separated on one

Ada satu komponen utama

of the principal components. And then there's kind of one principal component

punya 9s dan 7s, yang masuk akal

has the 9s and the 7s, which makes sense

karena mereka agak serupa.

since they're kind of similar.

Di sisi lain dari komponen utama itu,

And then on the opposite end of that principle component,

Anda lebih banyak punya 2s, 3s, dan 8s.

you have more like 2s and 3s and 8s.

Tapi kita bahkan bisa lakukan lebih baik.

But we can do even better than this.

Kita bisa gunakan fitur favorit saya disini,

We can use my personal favorite feature here,

yakni [tidak terdengar].

which is [INAUDIBLE].

Ini akan mengorganisir

So this is going to go and reorganize

seperangkat data sehingga itu menjaga kesamaan lokal.

the data set so that it preserves local similarity.

Lalu kita mendapatkan visualisasi keren ini

And then we get this really cool visualization

yang menunjukkan kita potongan data mana

that shows us which pieces of data

yang terhubung secara lokal di dalam manipol, di <i>embedding</i>.

were locally connected in the manifold, in the embedding.

Jadi kita bisa lihat, misal, 1s punya

So we can see, for example, that the 1s got really

susunan rapi sendiri, yang masuk akal,

nicely clustered on their own, which makes sense,

karena 1s adalah angka yang unik.

since 1s are very distinct digit.

Kita bisa lihat itu punya sedikit masalah,

And we can see that it has a little bit more trouble,

misal, dengan 3s, 8s, dan 5s,

for example, with the 3s and the 8s and the 5s,

karena mereka sedikit lebih dekat dan sedikit lebih membaur.

because those are a little bit closer and a little bit more

Kita bahkan bisa mulai menyelami contoh tersendiri

intermingled. And we can even start to dive into individual examples

dimana modelnya membuat prediksi buruk.

where the model made a bad prediction.

Jika kita lihat, misal, ke 7,

So if we look, for example, at the 7,

itu dikelompokkan dengan 1s, sehingga itu

it's clustered with the 1s, so that's almost certainly

kesalahan uji, itu secara tidak benar

a test error, whatever it incorrectly

melabeli 7 sebagai 1.

labels the 7 as a 1.

Tapi saat kita perbesar ke 7 yang khusus ini,

But as we zoom in on this particular 7,

itu memang agak mirip seperti 1.

it really does kind of look like a 1.

Itu mungkin 7 yang paling mirip 1 yang pernah saya lihat.

It's probably the oneiest 7 that I've seen.

Masuk akal jika modelnya membuat kesalahan itu.

So it makes sense that the model's making that mistake.

Jika kita lihat ke 9 ini, benar-benar mirip 1.

And similarly, if we look at this 9,

4 ini, menurut saya, memang 4.

it really looks like a 1. This 4, I think, really is a 4.

Mungkin modelnya punya masalah disini.

So maybe the model just has an issue there.

<i>Embedding visualizer</i> juga bisa digunakan untuk banyak hal selain gambar.

The embedding visualizer can also be used for a lot more than just images.

Itu sangat berguna untuk penggambaran kosa kata juga.

So it's really useful for visualizing vocabularies

Saya akan memperlihatkan sebuah contoh kepada Anda.

as well. And I'll actually just show you an example.

Ini seperangkat data balasan pintar.

This is the smart reply data set.

Tiap kali aplikasi kotak masuk Anda

So every time that's your inbox app

sarankan berterima kasih ke seseorang,

suggests saying thank you to someone,

itu memakai seperangkat data ini.

it's using this dataset.

Kita bisa lihat bahwa kita punya pernyataan kalimat seru di sini.

And we can see here that we have exclamatory statements

Tolong konfirmasi sekarang, konfirmasi jika Anda sudah dapat.

are over here. Like please confirm now, please confirm you got it.

Lalu akses ini punya pernyataan yang tidak terlalu berseru.

And then this access has less exclamatory stuff.

Kita bahkan bisa lakukan hal seperti membuat sumbu kita sendiri.

And we can even do stuff like creating our own axes.

Jika Anda tahu contoh raja ratu,

So if you know the king queen example,

Anda ambil kata [tidak terdengar] <i>embedding</i>.

where you take a word to [INAUDIBLE] embedding.

Anda atur satu sumbu yakni raja tanpa ratu,

And you set one axis is king minus queen,

lalu Anda tahu itu belajar banyak

and then you find it has learned a lot

konsep tentang gender, Anda bisa secara interaktif

of gender-related concepts, you can interactively

menjelajahi hubungan seperti itu menggunakan <i>embedding visualizer</i>

explore relationships like that using the embedding visualizer

dengan mendefinisikan sumbu Anda menjadi perbedaan

by defining your axes to be the differences

kata-kata yang berbeda.

between different words.

Waktu hampir habis, saya

So I'm almost out of time, but I'll

akan bahas sedikit mengenai rencana TensorBoard mendatang.

talk a little bit about the future of TensorBoard.

Kita akan, sekitar kuartal berikutnya,

We're going to be, in the next quarter or so,

mengintegrasi <i>debugger</i> TensorFlow baru ke TensorBoard.

integrating the new TensorFlow debugger into TensorBoard.

Itu artinya Anda akan mulai bisa pakai penggambar

So that means you'll start to be able to use the graph

grafik untuk lakukan hal seperti mendapatkan tampilan segera

visualizer to do things like get an immediate view of where

dimana ada NaN atau infinit di grafik kita.

there are any NaNs or infinities anywhere in your graph.

Jika demikian, kapan mereka muncul?

And if so, when did they appear?

Anda bisa pakai penggambar

Or you'll be able to use the graph

grafik untuk klik tensor tersendiri dan menggambarkannya.

visualizer to click on individual tensors

Kita juga lakukan pengembangan <i>plug-in</i>.

and visualize them. We're also doing a bunch of work on plug-ins.

Saya bilang kami menambah <i>summary op</i>

If you remember, I said we're adding a new summary

baru yang bisa menulis data apapun di TensorFlow.

op that can write out any kind of data in TensorFlow.

Idenya ialah kita bisa

And the idea there is that we can

mulai mendorong orang-orang mengkontribusikan visualisasi.

start to have people contribute visualizations.

Mereka akan fokus pada masalah domain yang khusus.

They'll be really specific to a particular problem domains.

Jika Anda ke <i>distill.pub</i>, Anda akan lihat Chris Olah dan rekan telah

So if you go to distill.pub, you'll

membuat visualisasi menakjubkan untuk menginspeksi model tulis tangan

see that Chris Olah and company have made amazing visualizations for inspecting handwriting models

atau untuk menginspeksi model ujaran ke teks.

or for inspecting speech to text models.

Kita ingin bisa memasukkan visualisasi khusus domain ini ke dalam TensorBoard,

And we want to be able to incorporate these really domain-specific visualizations into TensorBoard,

dan kita akan lakukan itu dengan mempermudah penambahan <i>plugin</i>.

and we're going to do that by making it easy to add plugins.

Akhirnya, kita membuat sesuatu

And finally, we're creating something

yang saya sebut skala organisasi TensorBoard.

I'll call organization scale TensorBoard.

Sekarang, TensorBoard

Right now, as you saw, TensorBoard

ialah hal yang orang secara

is something that one person locally

lokal hidupkan, di mesin mereka sendiri.

turns on, on their own machine.

Kita akan buat contoh besar dan berukuran produksi

And we're going to create a scalable, production-sized

sehingga keseluruhan organisasi Anda

instance so that your entire organization

atau keseluruhan grup riset bisa

or your entire research group could

berbagi TensorBoard tunggal lalu

share a single TensorBoard and then

berbagi hasil dengan mudah dan menjaga histori sepanjang waktu.

share results easily and keep a history over time.

Itulah pembahasan saya.

So that's it for my talk.

Kode dan <i>slide</i> ada di tautan ini.

The codes and slides are on this link.

Itu juga ada di GitHub saya, Dandelion Mane,

It's also on my GitHub, Dandelion Mane,

dan kita akan <i>tweet</i> tautan ini.

and I think we'll tweet this link out.

Selanjutnya saya akan memperkenalkan Martin Wicke.

And so next I'm going to introduce Martin Wicke.

Tak diragukan lagi Martin orang paling tampan,

Martin is undoubtedly the most handsome, best dressed,

berbusana rapi dan paling pintar di tim.

and most intelligent person on the team.

Dia juga bos saya.

And he's also my boss. [LAUGHTER]

Dia punya seluruh <i>open source</i> TensorFlow

And he owns all of TensorFlow open sourcing

dan dia akan beritahukan kita sedikit tentang API baru

and he's going to tell us a bit about new APIs

yang memudahkan menulis kode MNIST ini

that make doing all of this MNIST code much easier and much

dan lebih sederhana daripada di contoh saya.

less complicated than in my example.

Terima kasih banyak.

Thank you very much. [APPLAUSE] [MUSIC PLAYING]

Similar videos:

Integrating Keras & TensorFlow: The Keras workflow, expanded (TensorFlow Dev Summit 2017)
Google for Developers
67.4K views 770 likes 16 dislikes 15 Feb 201785% match

TensorFlow: Machine Learning for Everyone
Google for Developers
537.8K views 2.8K likes 59 dislikes 15 Feb 201775% match

Highlights from the 2017 TensorFlow Dev Summit
Google for Developers
29.1K views 199 likes 7 dislikes 24 Feb 201771% match

Skin Cancer Image Classification (TensorFlow Dev Summit 2017)
Google for Developers
42.8K views 579 likes 22 dislikes 15 Feb 201771% match

Case Study: TensorFlow in Medicine - Retinal Imaging (TensorFlow Dev Summit 2017)
Google for Developers
38.9K views 584 likes 4 dislikes 15 Feb 201770% match

Serving Models in Production with TensorFlow Serving (TensorFlow Dev Summit 2017)
Google for Developers
32.8K views 311 likes 2 dislikes 15 Feb 201768% match

Mobile and Embedded TensorFlow (TensorFlow Dev Summit 2017)
Google for Developers
28.6K views 290 likes 10 dislikes 15 Feb 201767% match

Wide & Deep Learning: Memorization + Generalization with TensorFlow (TensorFlow Dev Summit 2017)
Google for Developers
24.7K views 348 likes 11 dislikes 15 Feb 201767% match

Sequence Models and the RNN API (TensorFlow Dev Summit 2017)
Google for Developers
43K views 426 likes 18 dislikes 15 Feb 201765% match

XLA: TensorFlow, Compiled! (TensorFlow Dev Summit 2017)
Google for Developers
33.5K views 241 likes 9 dislikes 15 Feb 201763% match

Distributed TensorFlow (TensorFlow Dev Summit 2017)
Google for Developers
42.7K views 341 likes 4 dislikes 15 Feb 201763% match

TensorFlow at DeepMind (TensorFlow Dev Summit 2017)
Google for Developers
30.7K views 325 likes 12 dislikes 15 Feb 201763% match

TensorFlow High-Level APIs: Models in a Box (TensorFlow Dev Summit 2017)
Google for Developers
41.8K views 296 likes 15 dislikes 15 Feb 201762% match

Keynote (TensorFlow Dev Summit 2017)
Google for Developers
116.3K views 1.1K likes 25 dislikes 15 Feb 201758% match

TensorFlow Dev Summit 2017 - Livestream
Google for Developers
91.7K views 1.5K likes 52 dislikes 06 Dec 201657% match

Machine learning is not the future - Google I/O 2016
Google for Developers
175.5K views 928 likes 525 dislikes 06 Jun 201632% match

Breakthroughs in Machine Learning - Google I/O 2016
Google for Developers
61.6K views 536 likes 12 dislikes 13 Jun 201627% match

Effective TensorFlow for Non-Experts (Google I/O '17)
Google for Developers
307.6K views 3K likes 167 dislikes 18 May 201726% match

Open Source TensorFlow Models (Google I/O '17)
Google for Developers
59.2K views 1.2K likes 15 dislikes 18 May 201726% match

Chrome Dev Summit 2014 Highlights
Google for Developers
766.5K views 67 likes 21 dislikes 14 May 201526% match

Machine Learning: Google's Vision - Google I/O 2016
Google for Developers
151.2K views 1K likes 36 dislikes 20 May 201626% match

Getting Started with TensorFlow 2.0 (Google I/O'19)
TensorFlow
140K views 2.1K likes 40 dislikes 08 May 201925% match

VR & Cinema - Google I/O 2016
Google for Developers
76.6K views 683 likes 47 dislikes 19 May 201625% match

VR at Google - Google I/O 2016
Google for Developers
73K views 484 likes 16 dislikes 19 May 201625% match

Google Developers Channel - Trailer
Google for Developers
181.5K views 116 likes 21 dislikes 03 Sep 201524% match

Intro to Feature Engineering with TensorFlow - Machine Learning Recipes #9
Google for Developers
108.3K views 1.7K likes 25 dislikes 30 Oct 201723% match

Coming to a Chromebook near you - Google I/O 2016
Google for Developers
156K views 836 likes 27 dislikes 03 Jun 201623% match

What's New with Project Tango - Google I/O 2016
Google for Developers
113.5K views 757 likes 15 dislikes 19 May 201623% match

Machine Learning over Coffee with Josh Gordon - Coffee with a Googler
Google for Developers
85.1K views 1.9K likes 20 dislikes 30 Mar 201622% match

Settings Privacy About Contact Donate