Accessible Machine Learning with Tensor2Tensor Video Lecture | Coffee with a Googler - IT & Software

[Music]
so Lucas you work on tensor - tensor
which is like it's a set it's a library
data sets that's really designed to make
machine learning more accessible right
could you could you tell me all about it
yes it is
to me it's it's a follow up on the work
on tensor flow when a tensor flow is now
used by many people it's a great system
it lays the foundations for machine
learning we found that it's quite still
quite hard for people to get into
machine learning to start get their
first model gets some system working and
then you can start going from there
tweaking tuning there are a few
tutorials but then you get stuck on them
or there is a limit to what you can do
so in tensor to tensor it's library
where you can get the data get the
models train them get a working system
but then we know you can also go further
get more complicated systems we do
research on it so you can really try to
do difficult things with it so when you
mention get the data
so there's known datasets out there
they've packaged as part of this yes
we've we've packaged a lot of the
datasets that are used in academia image
classification datasets there is a data
set of digits that's like hello world of
of machine learning and it's called a
mist yeah and and there are many more
tour see far the larger data sets like
image net text data sets captioning
there are a lot of them so we package
them so you don't need to know where to
download them from it's there but you
also don't need to know how to
pre-process them because there are a lot
of tricks that people use for
pre-processing for adjusting it's
sometimes hard to find them you don't
know them in tensor - tensor they're all
in the code we explicitly point out here
is the pre-processing so you can read it
know how it's done
or you can just run it then you get the
data you have it you can run your own
models or you can run the models that
are there in tensor - that sort and
that's really useful to have because
sometimes you just want to get straight
to training and executing our models
right rather than needing to figure out
all the things that I need to do for a
specific data set so if that's already
done for us it yes on the other hand
sometimes you have your own data set
which you and you just want to run
existing models on it it it can
harder than people think intense
repressor you have the API you just make
your data set adhere to this very simple
API and all the same pre-processing all
the same models will be applied so you
should get reasonably good results yeah
just running the existing models common
concept and programmer is right don't
repeat yourself so if you've already
done the work for pre-processing
something if I put my data in that
format then I'm not repeating yes it's
an effort to put the machine learning
components and best practices on on a
single platform so that not everyone has
to redo it every time when I first
started looking into tensor 210 so one
of the things I noticed is that a lot of
the things intention to tensor are
described as problems and can you
describe what you mean actually by a
problem yes in the simplest case we say
amnesty is a problem okay it's a you
want to put in an image of a digit and
get out which digit that is and feminist
is a data set that has supervised
examples for of this kind but some
problems are unsupervised you just get
the image and the neural network should
figure out the digits some problems are
kind of reinforcement learning problems
you just get some signal this was good
this was bad but it's coming from an
external environment so we are trying to
make a single interface so you can
package these different things if you
put it as a problem in tensor two tensor
you're guaranteed that the models we
have will run on so if you're coming
with your data or your reinforcement
learning problem put it there and the
existing models the baselines that we
have can be run on cool yeah cuz for me
like as I've been doing it one of the
things that I've been doing is I've been
just trying to find raw data and then
I'm pre processing that data and then
I'm training a model but I don't have
any context like what are the best like
best ways to pre-process and best ways
to format the data do I normalize they
do I know you know all those kind of
things so from instead of thinking in
those terms what I'm able to do is just
shake my data according to tensor two
tensor and then you'll have yes
rewritten code that will work for me yes
it's certainly a great thing to do once
or twice to learn do everything and 20
on your own take a course do a tutorial
but what we found out in brain with
doing research on these topics for years
now after your tennis model you don't
want to redo everything from scratch and
maybe you want to compare with your
model from a year ago and it's really
great to have these things organized
tested right maintained working yeah and
I think this really makes machine
learning more accessible to a wider
group of people which is one of the
goals in Google right yeah exactly and
also it's like because it's open source
right so some of the things that you do
for pre-processing I can go and take a
look yes and I can learn from that so if
I want to write my own preprocessor I
can learn from what we we strive for a
little bit more than just open sourcing
we really try to keep the code readable
because you can open source code that's
very hard to read we're trying to keep
it organized so you'd know where to look
you know there is a pre process function
in every problem class and you know ok I
need to know pre-processing for this
kind of data I look they're some of the
coolest things that I've seen tensor to
tend to do is actually generation of
content right you've done some like
Wikipedia generation yes so we should I
say wiki generation rather than right
we have classification models but we
work on generating text a lot and one
example of generating text is
translation you take a sentence in one
language and then generate say a
sentence in English that's a translation
you generated usually word by word but
it's a lot of fun to just let it
generate without conditioning on
anything just say generate some
sentences and for a single sentence
might not be that fun to read but we
train this on whole Wikipedia articles
which is a great data set it's large it
has a lot of articles with nice
structure and we we see that it can
generate wonderfully coherent things
that are really hard to tell from real
articles but totally made-up so what
what do you write articles about we have
every time you generate because you
sample something different can come out
okay so we sampled and once to come out
an article about a Japanese music team
as a composer
has a name I think moto hero other
totally made-up we go doesn't seem to be
anyone like this but if you're watching
motor here oh thank you sounds
reasonable and it has sections when the
band formed who joined it has a section
on when they perform they performed with
Led Zeppelin so it clearly learned some
knowledge about the world like what our
band names what do they who do they can
perform with it has it has a period in
which it exists and the dates are like
make sense and and then the band is to
break apart with moto hero in tears so
it generates a long longest story
I mean tears - I never got to see them
yes but then you you run this again and
you get like some fake book or some fake
TV series
I like the fake TV series it had like
seasons and and names of episodes oh wow
it was like about the boy in the suburbs
as a summary oh this is great so I mean
if I'm a writer I can get ideas by
having you throw you know articles at me
so and another thing that I saw that
you've been working on there was a blog
post recently about generation of poetry
that was done together with Locke from
Google cloud and his technical writer
and he said well you're saying machine
learning is accessible with this library
let me be the guinea pig okay let me try
to do something I make up and see
whether it's really easy enough to do
and he said okay why not generate poetry
you're generating text I would like to
generate some poems he says there is the
Project Gutenberg which is a large
database open literature he wrote a
script that selected some poetry and
said let's train on it let me try can i
hook up this data into the system train
one of the models on it will it work and
he did it and it didn't work well he
trained but the model was not good it
was not tuned to this kind of data he
says well but we have automatic tuning
for the models let me see how I can do
that and he tried and tried and then it
worked and he wrote a whole blog post
how to I
I love this because it shows you if you
come with your own data and want to get
something how do I actually do this mmm
it goes through the whole experience and
at the end there is actually a service
which type Elena continues your poem I
think I've tried it out was fun and I've
put a link in the description below and
I wrote better poetry than I can so if I
wanted to get started to do this kind of
thing
maybe it's poetry maybe it's song lyrics
or whatever and then bring my own data
what advice would you give how should I
get started with it well go to cancer -
cancer on github okay or check out the
poetry blog post you can keep install it
and try to run from the command line we
also have Jupiter notebooks Python
notebooks where you can see step by step
what's happening also inspect the data
especially if you do image just actually
seeing what's coming on input what
what's supposed to come on output what's
actually coming on output this is a very
useful thing so there is a notebook
wordly step by step show you what to do
go to the website try it out and also
it's a thanks to a lot of my colleagues
it it's an it's a very active project we
work on it every day we work with it we
work on it so you can start with the
small things like poetry it's a very
small data set containing a few hours
amidst you can train in two minutes but
you can move on like the Wikipedia we
also we did just Wikipedia generation
but we also did Wikipedia conditioned on
like web searches so it doesn't make
things up it's actually factual it
generates things that the internet says
about people I see and this data set is
8 terabytes it takes a bit longer to
train but we we still open-source it
thanks to the common crawl which which
has a large part of websites online it's
possible to actually do this well at
home I don't know but at a university
you can get this really large data set
you can run state-of-the-art models so
it's a platform where you can start
small but really go and go big this time
cool so tensor - tensor and it's on
github as part of the tensorflow project
so cool well thank you so much Lucas
this has been so much fun and I've
learned so much and thank you everybody
for watching this episode of coffee with
a Googler if you've any questions for me
if you've any questions for Lucas please
leave them in the comments below and
also don't forget to take a look at the
description we've put links to
everything that we spoke about and also
we have a brand new tensorflow channel
on youtube so be sure to check it out
thank you