BINARY CLASSIFIER WITH TENSORFLOW - CS50 on Twitch、EP.7 (BINARY CLASSIFIER WITH TENSORFLOW - CS50 on Twitch, EP. 7)

字幕表動画を再生する

COLTON OGDEN: All right.
Hello world.
This is CS50 on Twitch, and today we're joined
by CS50's Nick Wong, who's been a teaching fellow here
for a couple of years.
And today he's going to teach-- oh, what are you going to teach us today?
What are we going to talk about?
NICK WONG: So we're going to be talking about binary classifiers
using TensorFlow and Keras.
And so TensorFlow was actually a project that was developed and built by Google.
And Google has their hands in everything,
but TensorFlow is really cool because it basically took
the whole linear algebra-ness of machine learning and made it very simple.
And then Keras is a wrapper on top of TensorFlow
that actually made even that simplification a little bit simpler.
COLTON OGDEN: Let's transition to your awesome little screensaver back here.
Yeah, so like an introduction to machine learning using some nice open source
libraries, some Google base-- is Keras a Google open source library as well?
NICK WONG: I believe so, yeah.
COLTON OGDEN: OK.
TensorFlow-- I know everyone's heard of TensorFlow, or at least a lot of us
have.
I know almost nothing about machine learning,
so this will be a learning experience for me as well,
but I'm super excited to talk about binary classifiers and what they are.
NICK WONG: Awesome.
So yeah, I'll talk a little bit about the theory first.
I won't talk too much on theory.
I know people aren't usually huge fans.
I like the theory, I think it's really cool,
but we're going to just talk about it a little bit
to motivate and contextualize what I'm actually going to start
saying and coding in front of you.
Before I do that, I just wanted to point out this screen saver
that I've got going on here.
I think it's super cool.
I'm a huge fan.
It's technically CMatrix.
That's the actual program that's running.
But I put that through another program called
lolcat, which is one of my favorite "messing with people"
or troll programs.
If you're ever trying to mess with someone,
you can display lolcat and stuff through a screen.
COLTON OGDEN: We'll have a separate stream on how to troll people,
how to hack into a--
NICK WONG: One of my favorite things.
COLTON OGDEN: And this screensaver is in a shell, right?
NICK WONG: Right.
So this is in a shell.
If I press any key, then I'm back in my shell,
and I can go back to coding and doing all sorts of cool things.
But I like having it there.
It's entertaining.
It just drains power, so you want to be plugged in.
COLTON OGDEN: We have a few comments in the chat there.
MKLOPPENBURG says, "Hey everybody."
BELLA_KIRS says, "Hello."
WHIPSTREAK, BELLA, and I can't read that last one.
NICK WONG: Yeah, it's hard to read.
COLTON OGDEN: It looks like ILYAS.
They all say, Hi, Nick.
NICK WONG: Oh, awesome.
Hello guys.
Appreciate it.
Very happy to be here.
COLTON OGDEN: We have a lively chat, so we'll definitely read off the messages
as we get them in the chat.
NICK WONG: Sweet.
All right.
So machine learning.
It's one of my favorite buzzwords that everyone throws around,
along with like blockchain, AI--
what are some other ones?
There are some really good ones out there.
Bitcoin people that are out there.
COLTON OGDEN: Yeah, blockchain.
They've been massively-- yeah, those last couple years.
NICK WONG: Everyone keeps throwing them out there, and they're all great.
And so I think one of the things that I heard this last summer where
I was working, someone said that machine learning is just fancy statistics
wrapped up in a computer.
And I was like, yeah, not really.
Like in concept, yes, but not necessarily in what
they actually intended.
What they were meaning is, like, you could just
use statistics to do the same thing.
And when I was walking through our science center today,
I was looking at the old punch card style computers.
And I was like, yeah, technically I can run statistics on that too,
but it's different.
The game has changed a little bit given our computing power, our memory,
and things like that.
COLTON OGDEN: Little bit apples and oranges type deal.
NICK WONG: Yeah.
It's just like, yeah, you're not wrong really, but you're pretty off.
And so machine learning is just a very broad field.
It really means just any way in which you
get a computer to figure things out.
And a lot of it's modeled after how humans learn.
So we-- well, in concept.
It's modeled after the way our brains are structured.
So a lot of things are called neurons.
So you hear a lot of just jargon that gets thrown out.
You'll hear, like, tensors, neurons, layers, models, things like that.
COLTON OGDEN: I've heard, like, neural nets, for example.
NICK WONG: Right, neural network.
And those sorts of things can be very intimidating,
but they generally have very well grounded meanings.
So for example, a neural net, that just means that, like our brain,
they took a bunch of individual nodes and they linked them all together,
and that's all that it really ends up meaning.
Now, there's some nuance to that there's some complexity,
but in concept, it's actually pretty straightforward.
So what we're going to do is we are technically
going to build a neural network today.
COLTON OGDEN: That's cool.
NICK WONG: It is what is required for-- or not required,
but it's what we're going to use for the binary classification.
And that brings us to what binary classifiers are,
which is if you have a bunch of data, can I
tell you if it's in one of two groups?
So it's either going to be group A or group B.
And that can mean all sorts of things for us.
COLTON OGDEN: So if we're looking at pictures of apples versus pictures
of oranges, what's an apple--
NICK WONG: Is it an apple or an orange?
COLTON OGDEN: --what's an orange?
By the way, DARKSLAYERX says, "How often do you stream?"
So we typically stream, as of the last couple of weeks,
this is a fairly new thing, but usually four days a week,
so Monday, Tuesday, Wednesday, Friday.
Thursday is a little bit of a busy day right now for CS50.
We're shooting CS50 seminars.
You're doing a seminar, right?
NICK WONG: Yep.
I will be also doing a seminar.
COLTON OGDEN: Do you know what it is on yet?
NICK WONG: It'll be on web development with Python and Django.
COLTON OGDEN: OK, awesome.
NICK WONG: It's a completely different--
COLTON OGDEN: Completely unrelated.
NICK WONG: They're not even related.
COLTON OGDEN: But yeah, pretty frequently.
Our goal is to do this roughly three to four times per week.
It looks like WHIPSTREAK also tossed in that's it on the CS50 Facebook page.
Yes, it is on the CS50 Facebook page most definitely.
NICK WONG: The human equivalent of check demand page.
COLTON OGDEN: Yeah, yeah.
Yeah, basically.
NICK WONG: One of my favorite responses.
That's awesome.
COLTON OGDEN: MKLOPPENBURG-- Django rocks.
NICK WONG: Yes, I agree.
And we use I guess Flask and-- what is it?
Flask, Jijnja2, and MySQL in CS50, and I'm a huge fan of Django
which kind of wraps that all together.
But today, we'll be just using Python and we're
going to talk about [INAUDIBLE] stuff.
COLTON OGDEN: Separate stream.
We'll do a separate Django stream for it.
NICK WONG: Yeah, completely different.
Yeah, exactly.
COLTON OGDEN: By the way, you will be doing another stream--
NICK WONG: Yes.
COLTON OGDEN: --coming up on Linux commands.
So Nick will do-- if anybody is curious on how
he got his awesome show operating this way,
we'll dive into the basics of how that works with the basic Linux commands.
NICK WONG: Yeah.
That's next Friday, right?
COLTON OGDEN: Yeah, next Friday.
NICK WONG: Sweet.
Yeah.
So feel free to tune into that if you think this is cool
or if you want to learn more about Linux.
All right.
So we have binary classifiers.
We're just separating things into two categories.
And that sounds like an intuitive thing.
If I show you a picture of a square and a picture of a triangle,
you'll know that one is not like the other.
But for machines, that can be really complex.
And actually, the theory underneath that seems very difficult.
And so machines, early on, that was one of the easiest problems
that they could try and tackle solving.
And it's still hard.
It's not an easy problem, but it is one of the easiest.
COLTON OGDEN: Identifying basically simple shapes, those sort of things?
NICK WONG: Right.
Can I just say, is this this or not?
So is it a shape or not?
Is it dark or not?
And so what we're going to actually start with--
and for you guys, I actually started a GitHub repository
that you guys are welcome to kind of pull code from as we go--
COLTON OGDEN: Oh, do you want to--
NICK WONG: --which we might want to bring that up.
COLTON OGDEN: --let me know what that is and I'll type it into chat?
NICK WONG: Yeah.
So it's at GitHub.com/powerhouseofthecell--
COLTON OGDEN: Let me get the--
Sorry, I can't read it too well.
Powerhouse.
Oh, you're right.
Yeah.
Literally--
NICK WONG: Literally powerhouse.
COLTON OGDEN: You changed your Twitch handle a little bit from last time.
NICK WONG: Yeah, I had to change that, unfortunately.
COLTON OGDEN: Is_it_real.
NICK WONG: Yeah, exactly.
So our eventual goal for today, what we will get to by the end,
is given pictures of cartoon people versus real pictures, like a portrait,
can we distinguish which one is real?
COLTON OGDEN: That's an interesting--
NICK WONG: Can I tell us-- is something cartoon or is it not?
And the reason I thought that was really cool
is there is no necessarily apparent programmatic way
to distinguish the two.
You can't tell based on color.
You can't necessarily tell based on orientation or type of shape.
Maybe for really bad approximations in cartoon structure,
you could maybe guess at that, but even then, that might not be true.
COLTON OGDEN: I was going to say, if I were to take an initial stab at this
completely naively, I would probably build the cartoon classifier and then
work my way up to real humans because I feel like that's--
it's almost a simplification or an abstraction of that.
NICK WONG: Right, exactly.
And so that's what my point was, is that even given very limited tools--
we have basically no theory that we've talked about.
We don't have to talk about likelihood maximization or any sort of summation
notation, none of that.
We can still do that.
We can get really high accuracies with pretty minimal tools.
And they're all open source.
They're all available anywhere you want to go.
So I think that's really cool.
And I think when I was beginning CS, that was super inspiring to me, so I
like sharing that with people.
COLTON OGDEN: Cool.
NICK WONG: We're also going to use a third library I forgot to mention.
It's called OpenCV or Open Computer Vision.
I think it was UC Irvine's project.
And I actually used that my freshman year,
but I've used it ever since because it is
one of the primary libraries for reading images,
for just pulling them in and doing something like that.
But the problem with OpenCV is they update their documentation possibly
less consistently than college kids go to sleep.
It's so all over the place.
It does say MIT license.
Sorry.
To respond with WHIPSTREAK's question, which is, is it an MIT license?
I'm in my GitHub.
Why?
Isn't this Harvard?
MIT license is actually kind of a broad license.
It's generally used for personal projects or projects
that you want to be able to share it with the world
but you do want them to acknowledge you or just reference the fact that it's
not necessarily theirs.
It's one of my favorite licenses to use on GitHub.
Another common one that you might see is the GNU license as well as
the Apache license.
Those two or those three are pretty well-known.
And then there's a bunch of other license options
that I don't understand super well.
COLTON OGDEN: And then FERNANDO also says, "Hi."
Hi, Fernando.
NICK WONG: Hey, Fernando.
COLTON OGDEN: Thanks for joining us today.
NICK WONG: Awesome.
So we're going to get going a little bit on our environment and what we're doing
and how we actually go.
So I'm in Bash.
This is my command prompt.
I just like the color blue a lot.
Ironically, I'm wearing red, but that's OK.
And so this is my Bash prompt.
If you're not super familiar with Bash or Linux shell, don't worry.
I might refer to with a couple of words like terminal, console, Bash, shell.
They all mean the same thing.
I just mean here.
I can type stuff ls, which has all sorts of stuff on there.
I'm a mess.
I can do ps to see what's actually running-- very little.
There's a bunch of other commands, and we'll go into them next Friday.
But what I'm going to do is I actually have a technology called
Virtualenvwrapper, and what that means is
I can basically create these separate little Python modules,
Python environments for myself.
And so that means that if I have some projects that uses Python 2.5 and then
my current project uses Python 3.7 because I'm not in the 20th century--
sorry, I like to jab at the old Pythons--
then I can separate everything out and not
worry about colliding dependencies or colliding version numbers.
I can keep everything very clean.
COLTON OGDEN: Or even two versions of the same library.
NICK WONG: Exactly, which would be-- that is the worst.
I have done that many times, and it's awful.
So we're going to call that, this version.
Is_it_real.
It'll match with my GitHub, and I'll try to be very consistent.
And you'll see that when I run this mkvirtualenv,
but I'm not really necessarily going to go into.
There are a lot of tutorials online on how to install Virtualenvwrapper.
That is what I'm using, though.
And it takes a little bit to go, but it creates all these ways for me
to actually interact with Python in an isolated place.
And you'll notice that, over here, it puts a little--
it prepends a part to my prompt that says
that I'm in that virtual environment at any one time.
So I can leave it by just saying, deactivate,
just for those of you that are gone, and then I
can work on it using this command.
Cool.
So I'm also going to create a folder for this.
I normally would have done some make directory or mkdir,
but I'm actually going to just go into the one that I--
oops.
It's called is_it_real.
My apologies.
I'm going to go into one that I already created just because already cloned
my GitHub repository.
If you cloned it, then you'll actually end up being in a place like this,
and you'll have these two files in there--
just the readme and the license.
We are literally creating this from scratch.
COLTON OGDEN: That's the best way to do it.
NICK WONG: So I apologize if we have some syntax errors.
You'll get to see some really live debugging.
COLTON OGDEN: We've had plenty of that on the ones that I've done, so--
NICK WONG: That's good to hear, because there's going to be plenty on my end.
All right, cool.
So we're in a Python virtual environment,
and we need to pick up a bunch of packages.
So we're going to do that live too, which never goes well.
I've never had this code correct on the first go.
So we're going to pick up some kind of maybe not conventional ones,
or they aren't the names of the packages that I said at the very beginning.
We're picking NumPy and SciPy.
And those are very common.
Machine learning, any sort of numbers or data
analysis-- you're going to usually see those.
NumPy or NumPy makes things very fast generally speaking,
and it's going to give you a bunch operations with numbers.
I'm going to actually let this run while I talk.
Oh good, they're cached.
Thank god.
They're kind of large, so they're kind of a pain.
And then SciPy is a bunch of science-like tools.
So if you're analyzing data, it's really useful.
COLTON OGDEN: It's literally In the name--
SciPy.
NICK WONG: Yeah.
It's always like Python is super intuitive.
It's meant to be user friendly.
The next thing that we're going to grab--
and I always forget what it's called.
I think it's opencv-python.
Their package name is not super intuitive because the package itself
when you import it-- thank god--
is called cv2.
However, we're using cv3.
And it's still called cv2, so there's that.
And you download it as opencv-python, which is absurd.
So I don't really like their naming scheme, but that's OK.
COLTON OGDEN: WHIPSTREAK says, "Nick and Colton are acting like best friends.
Are they?"
We're definitely good friends.
NICK WONG: We are very good friends.
COLTON OGDEN: We taught a games course together actually,
so we do have a little bit of history here.
NICK WONG: Yeah.
Colton's awesome.
I actually got to watch him full powerhouse last year
teaching his course that he went and did the whole nine yards on.
I was just there for the ride.
It was very cool.
COLTON OGDEN: We flipped the situation around a little bit,
so now you are teaching and I am observing.
NICK WONG: I appreciate it.
COLTON OGDEN: It's fun.
It all comes around.
TWITCHHELLOWORLD-- "I don't see the Livestream video.
Is everyone seeing a Livestream video?"
The video should definitely be up.
I see it on my Twitch.
Definitely refresh the page and see if that fixes it.
NICK WONG: Very possible.
COLTON OGDEN: Astly, who's NUWANA333, says, "Hi everyone.
Yes TWITCHHELLOWORLD."
Good to see you, Astly.
But yeah, if you have any issues, definitely refresh the page.
I'm seeing it live, and It looks like everyone else
is seeing it live as well.
NICK WONG: Yeah, good luck.
Debugging browser stuff is one of my least favorite tasks.
It's why I work in Python.
All right.
So let's see.
We've picked up NumPy, SciPy, opencv.
We also need Pillow, which is not an intuitive package that you would want.
It's capitalized too, so just in case you encounter some weird errors there.
And Pillow, it's the-- what is it?
Python image library.
So that translates to PIL.
They call it Pillow and it's capitalized, but when you import it,
you don't.
You don't import Pillow anywhere really.
Some of these packages are kind of absurd.
So then, we're also going to install the stars
of our show, which are TensorFlow--
so shout out to Google for making that accessible to everybody--
and Keras, which is a wrapper on top of TensorFlow.
It makes it a little bit higher level.
It's a little bit easier to interact with.
And that way, you're not dealing with as much terminology,
so you're not sitting there being like, what's a tensor?
What are the dimensions?
And all these things.
Keras takes care of that.
COLTON OGDEN: Shout out to also how easy it is to install all these packages.
That's probably why Python is so successful.
NICK WONG: Gotta love Python.
I am a huge fan of Python.
I think Colton is too.
Python is just very well done.
You're seeing maybe a different output than you might see on your screen
because I have used these packages pretty extensively,
so they're all cached on my machine.
It just means that they've been saved before.
And you'll see some bars that go across, and that should be correct.
Cool.
So we've downloaded all of our packages.
If we wanted to test that out, we could go into the Python REPL
by just typing Python and say, like, "import tenserflow,"
and that should not fail.
That would suck if it did it.
And maybe "from keras.models import--" that might-- yeah,
I think that was probably right. "import Model"
I'm also going to assume a lot of Python syntax,
so if you have any questions, feel free to post it in the chat
and we'll be happy to answer it.
And so yeah, we have both terrace--
sorry-- TensorFlow and Keras.
And then the last one that we want to double check that we have is cv2.
COLTON OGDEN: It looks like Keras even has
a implicit dependency on TensorFlow--
NICK WONG: Yes.
COLTON OGDEN: --it looks like with that line there.
NICK WONG: Yeah, exactly.
So actually, Keras has this thing.
I think most people use TensorFlow now underneath it.
It's kind of the back end for it.
But it can also use Theano or Theano, which is a different machine learning
background or back end.
It's very cool, also very powerful.
I just happen to use TensorFlow.
So you can use either one.
We're going to just talk about TensorFlow though.
Cool.
So we can exit out of that with Control-D
or by typing in the word "exit" and parentheses.
And now we're going to actually start creating some code
or writing some code.
So what I'm going to do is, generally speaking,
I like to separate things out and be very modular and have good practices.
But when I first go through stuff, I just want to get it working.
I just want things to actually function correctly.
So we're going to write a monolithic style for this,
meaning that, basically, we're going to--
and actually, I'm going to just touch this.
Oops.
I can't spell.
I can never spell live.
And we're going to just touch run.pi.
I like calling it run.py.
It indicates what I'm going to do with it.
I'm going to run it.
And so that makes things a little bit easier on ourselves,
and it'll be a little bit easier to follow so
that we're not flipping between files all over the place.
However, I also don't exactly love typing in just Bash
and using nano for everything.
So we're going to use Visual Studio Code.
It's one of my favorite IDEs.
A completely different team was behind it.
I think mostly different team was behind it at Microsoft
then the team behind Visual Studio.
So if you have any kind of group with Visual Studio, one,
they've revamped it a lot, but two, it's very different from VSC.
And VSC is super customizable.
It has a lot of nice plugins.
COLTON OGDEN: Big fan of it myself.
NICK WONG: Yeah, which is nice.
COLTON OGDEN: We have a couple comments there.
Looks like WHIPSTREAK says, "What is the machine learning program's objective?"
NICK WONG: Ah.
OK, right.
I didn't actually it.
I mentioned it to Colton right beforehand and then didn't actually,
I think, say it.
So we're going to be going over--
basically, the end goal is to see if we can distinguish, given an image,
if it's a cartoon image or a real-life one.
And we're going to focus on images of people
because that makes it a little bit easier for us.
But that's going to be the end goal.
We're going to start with a little bit easier
of a task, which is distinguishing whether or not
an image is dark or light, because I think
that that conceptually is a little bit easier for us to understand.
There is a way for us to figure that out.
I think you could, maybe upon thinking about it for a little bit,
come up with a program that does it deterministically.
It takes the average--
images are just data.
And it maybe takes the average overall of the image's intensities and says,
if they're over some threshold, it's light,
and if they're under that threshold, it's dark.
And I think that that would be a very nice programmatic way to do it.
You need to adjust a mic?
DAN COFFEY: [INAUDIBLE]
NICK WONG: Ah, yeah.
My apologies.
DAN COFFEY: Sorry about that.
NICK WONG: Thank you.
No, no worries.
COLTON OGDEN: Dan Coffey, everybody.
Shout out to Dan Coffey.
NICK WONG: Dan Coffey makes all of this magic much more possible.
COLTON OGDEN: Oh, we got some--
NICK WONG: Awesome.
COLTON OGDEN: --other comments there too.
NICK WONG: Oh, yeah.
COLTON OGDEN: CHITSUTOTE-- "Hi there!
What's the topic today?
Machine learning?"
Yeah, machine learning.
Going to be talking about TensorFlow and cartoons, light and dark.
Oh, DAVIDJMALAN is in the chat.
NICK WONG: Wow.
That's awesome.
COLTON OGDEN: Everybody shout out to DAVIDJMALAN giving vc code link there.
NICK WONG: Yes.
COLTON OGDEN: He's coming in the clutch.
NICK WONG: Very helpful.
COLTON OGDEN: Lots of hellos for David there.
MAGGUS503.
NICK WONG: Nice.
Yeah, I love it.
Is that a unicorn?
COLTON OGDEN: I think so.
NICK WONG: Or maybe a pig.
COLTON OGDEN: I think it's a Brony maybe.
NICK WONG: Oh, a Brony.
Nice.
COLTON OGDEN: I can't tell.
It's hard to tell.
NICK WONG: It's something, and it's cool.
COLTON OGDEN: "If I login on a desktop," says TWITCHHELLOWORLD,
"then is it maybe an option to see the livestream video there?"
I would use Chrome.
Use Google Chrome because that's the only web browser that I really
tested this extensively on.
It should work and Firefox, should work and Safari,
should work on most major web browsers.
But the latest version of Chrome on Twitch.tv/CS50TV should work just fine,
so give that a try.
NICK WONG: And best of luck.
COLTON OGDEN: I'm not too familiar with any Twitch for desktop apps,
but presumably those would work as well.
But chrome is the only one that I'm personally familiar with.
NICK WONG: Makes sense, yeah.
Yeah, actually the same.
I usually don't watch too many Switch streams, actually.
COLTON OGDEN: Just star in them, really.
NICK WONG: Yeah.
I like to just host them.
All right, cool.
So we have this open.
If you're not super familiar with an IDE, don't worry.
On the left, you just have my file structure,
all the files that I have access to in this directory.
At the bottom, I have a console, so this is just Bash again.
You can just list things out and do normal Bash things.
And then, in the main bulk of the screen is
where we'll actually be writing code.
Hopefully, that's not too small for everyone.
It should be big enough that everyone can read it.
COLTON OGDEN: It's the size I usually program on.
NICK WONG: OK, perfect.
COLTON OGDEN: If anybody thinks it's too small,
definitely let us know in the chat.
NICK WONG: Yes.
And actually, I just moved my files off the screen--
give us a little bit more real estate.
So what we're going to do is we're going to setup all of our imports
and get this file built so that we can actually run it and execute
things that are going on.
I know just ahead of time I'm going to need os.
I'm going to use it for path joining and checking directories and things
because we're going to have pictures in a bunch of directories,
so I'm going to use os for that.
I also know that I'm probably going to want argparse for later design stuff.
argparse is just a really convenient library
for doing command line arguments so I have
to do, like, if the length of argc--
or if argc is greater than 2, than this, and then otherwise--
I just particularly like argparse.
It's built into Python.
And those are our two system libraries or the built-ins for Python.
We're also going to definitely need Keras.
But with Keras, you don't usually indirectly import Keras.
We don't actually need Keras on its own.
So we're actually going to say, "from keras.models import
Sequential," which is not exactly model.
It's actually, I think, a subclass of model.
And Sequential has a bunch of things preset for us
because we know that we're going to be building some form of model where
we just add layers to it one at a time.
And actually, we may not have known that.
I know that, and I'm going to tell you that,
that that is how it's going to work.
And so, yeah, Sequential model means that we're going
to just stack everything together.
It looks-- if you were to view it in real life,
it would literally be a stack of layers.
And each of those layers represents all sorts of different things,
and we'll talk about those when we get there.
COLTON OGDEN: That's like refinements possibly?
NICK WONG: Yeah, exactly.
They're ways of tinkering with the data as it comes through.
I like to think of it as, like, if you imagine them as a bunch of filters
and you're just pouring sand through it, and at the end
you get something meaningful.
COLTON OGDEN: A ruby or something, right?
NICK WONG: Hopefully sand you like.
Yeah, or a gem.
Cool.
So we're going to import Keras.
We're also going to import-- oops--
cv2.
And that is going to just let us open images and do things, basically all
the image manipulation through that.
And I don't think we necessarily need any other ones.
Oh, we actually do also want NumPy.
We're going to use it to reshape arrays and images and things like that.
And so NumPy actually usually gets imported as something usually.
The convention seems to be "import numpy as np."
And so a lot of tutorials, you'll see just "np."
blabbity-blah, and that assumes that we'll use np.
Cool.
And then, I like to follow the C paradigm of "if name--" oops.
Yeah, you're going to see me in mistype constantly today.
Then we're going to run stuff.
And I'm going to put "pass" there for now,
but we will generally put something there.
And I like to use this space up here to write out some functions
and help with debugging.
And then down here, we're going to just call those functions.
So very C script mashup is what we're seeing here.
However, I also like to use Control-C to exit out of functionality as we go,
so I'm going to wrap this and try an except KeyboardInterrupt,
and that'll let me just say, like, oh, the user deliberately aborted.
I like to put preface that.
And that'll just make it a little bit easier
for me to distinguish between me Control-Cing and actual errors that
happened while we were running.
So this is one of my favorite ways of setting up programs.
You're welcome to do your own.
Excuse me.
All right.
So before we actually even get to the keras.model's import Sequential,
I'm going to let us have access to command line arguments in case
we get to that later.
So I'm going to say, ap just stands for argparse, is argparse.
I was doing this without a editor earlier when I was testing things out,
and man, it's much harder.
This is great because I just have autocomplete,
and I don't have to remember the actual names of things.
COLTON OGDEN: IntelliSense.
NICK WONG: Right.
IntelliSense is-- what a beautiful, beautifully built thing.
And then I'm going to just say that this is ap.parse_args.
Sorry, I didn't really narrate any of that.
In between this line 4 and line 6, we can put it like ap.add_argument,
and those will be available to us on the command line.
So that's just a really convenient way of doing this.
And actually, even right now, there are certain commands
that are now available to us on the command line.
So we're going to look at that.
If I do run.py, it'll tell us that we're using TensorFlow
and nothing will happen.
That's what we expect.
But then, from here, I can actually do -h and we'll get a usage message,
and that's arg_parse doing its job.
And the reason that I do that before TensorFlow or before importing Keras
is because Keras is actually a very slow import,
so we're completely able to ignore that if someone
didn't pass a necessary command line argument
or they wanted to just get help.
It speeds things up just a little bit.
So maybe stylistically it looks a little strange,
but it's definitely not too problematic.
COLTON OGDEN: I think old WHIPSTREAK has a question for you.
NICK WONG: Oh, yes.
"In line 12 and line 13 necessary or-- sorry, 16 and 17."
I believe you're talking about this.
You can also use the actual code that's in those lines
because lines will change and we're a little bit delayed,
or you guys I guess are a little bit delayed from what we perceive.
So also telling us which lines of code you're talking about by name
COLTON OGDEN: I think he's asking or he or she is asking whether the
"if name is main" is necessary to run the script.
NICK WONG: Ah, OK.
I see.
No.
So if you're asking about the "if name equals main, then we do stuff,"
it's totally not necessary.
Python is a fully capable scripting language.
You can just run Python scripts as you go.
I just like to follow a C style paradigm where we have a main
and it actually gets called and things go.
This also means what also ends up happening
for us is that if I import this whole file as a module,
then that stuff is not going to get automatically run.
So I would have--
excuse me.
I would have access to the functions that I might build earlier on or later
in this file, but I'm not necessarily going
to have them all just run without me telling them to.
So it serves a couple of purposes.
I think it's very convenient.
Cool.
So now we're going to start building some stuff.
So you might say, OK, well, the first intuitive thing
is that we're going to want to be able to load data.
And that makes a lot of sense, except I don't
know what my data looks like, I don't know where we got it from,
there's no data that exists at the moment,
and I have no file structure for it.
So loading data is un-implementable at the moment.
And I like to walk through these and motivate different parts of the problem
as we go, and we've not motivated that we don't have any data.
To run a machine learning model without data is, I think, a nonsense idea.
So we're going to get some data.
You're going to get a nice view of just how messy everything is in my file
structure.
I'm going to use Finder to just create some files.
If you will trust me on faith, you can hopefully
believe that this is the same directory that we're looking at here
and it's the same directory we're in over here.
They're all the same.
And if you don't believe me, then I apologize.
Nothing I can do on that.
I'm going to call this images because our data type is images.
You can call it whatever you'd like.
And then, within img, I'm going to actually have three folders.
So the first two are train and test, or validate is another word
that you might use for that.
And this is pretty standard in any sort of machine
learning setup, at least for supervised learning
where we're trying to feed it answers and get it to learn some pattern,
is we're going to say, take my data and split it into two data sets.
I have a training data, and I have a testing data
set that validates whether or not that data's good.
And the reason that you would split something
is in machine learning models, they can sometimes over-learn things.
For example, humans do this all the time too,
where I might tell a child that after 1 comes 2, and after 2 comes 3,
and the child might then go and say, OK, well then
after 3 comes 4, and then 4 comes 5.
And that's very reasonable.
That's something that's totally what we might expect,
except I might have just fed you the first couple of numbers
and a Fibonacci sequence, in which case after 3 does not come 4 but rather 5,
and then 8 and then 13.
And so it's really important to us to be able to give the machine learning model
data that it doesn't incorporate into what it actually does,
and data that it can then just see how well it actually is doing.
Basically, it's called over-fitting, where a machine learning model learns
all of its training data by memorizing.
And so it just says, this is what I know.
This is everything that I'm going to do.
An analogy for that is a child saying, I know four colors
and I know them very well.
And then you hand them a new color, and they have no clue what it is.
And that's something that real humans could, in concept, do.
We tend to compensate for that, but it's definitely very possible,
and machines do it all the time.
So we split this into training data to give it something to actually learn on,
and then we see how well it's actually doing
kind of in the real world on data it's never seen before in the testing data.
And that data does not getting incorporated back
into how the machine actually learns.
It's just used, literally, to test it.
And then, the third directory that we're going to put in is a predict directory.
And I just like to structure things really cleanly.
And so this directory is just going to contain a bunch of images
that I want the machine to predict on.
I want the machine to tell me, what do those images actually have?
Are they light?
Are they dark?
Are they cartoon?
Are they real?
Things like that.
COLTON OGDEN: Cool.
NICK WONG: Cool.
So in train and test, we're going to want--
the way that Keras pulls images from a directory
as it uses the directory name as the label for that data.
So I'm going to create a new folder, and we're going to call that light.
And I'm going to create a new folder here and call that dark.
And so we'll want to replicate these over to test.
And so these two folders, basically what they do is inside of dark,
I'm going to put a bunch of dark images--
not dark in the metaphysical sense, but literally just darker images.
They're just blacker colored.
COLTON OGDEN: R-rated stream.
NICK WONG: Yeah.
We're going to keep everything very nice and PG I hope.
And so everything here will just be-- is it
literally just more black pixels or more light pixels?
And a lot of times, people are like, oh, I have to collect a bunch of data.
And I've actually been very much an advocate
of collecting just smarter data, not necessarily more.
So what I mean by that is, are you covering all the cases, for example?
And in our case, we don't really have that many.
There's just, like-- actually, I guess there's
a very large number of ways you can make a dark image versus a light image.
But what I mean by that is we can collect some things,
like maybe there's a just pure black screen.
Well, that makes sense.
And that's dark, for sure.
But what about if you had a black screen with some stars in it?
Well, then that might actually be slightly different.
Or for example, this picture.
I'm literally going to just drag and drop a bunch of pictures from Google.
This picture, I would classify it as dark.
It's pretty black.
There's a little candle in there.
And so a machine might not deterministically
be able to say, oh, well that's dark, but machine learning
can probably figure that out.
And so I'm going to copy a bunch of these images.
You'll notice a lot of them are not exactly all black.
Most of them have some sort of image in them.
And this one is pure black.
Love that.
This one is also quite dark, but not black
really, not as black as the other ones.
COLTON OGDEN: An interesting test case on that one, yeah.
NICK WONG: Right.
So I think that that's something that would be really worth putting
in front of our computer.
And I'm deliberately ignoring an image that I know is on the right
because it's a really good test example.
So I'm going to grab 10 of these.
Oh sorry, it's actually down here now.
So there's this image, and this image I would classify as dark,
but it has a very bright center to it.
So if we can get our machine learning model
to figure out whether or not that's light or dark,
that'd be really interesting for us.
So we're going to put that in our predict folder.
COLTON OGDEN: ZODI4KX says, "Hey guys, I'm Wesly."
ZODI4CKX.
NICK WONG: Oh.
COLTON OGDEN: Good to have you, Wesly.
Thanks for joining us.
And then BHAVIK_KNIGHT, who's a regular, is in the chat.
NICK WONG: Welcome back.
COLTON OGDEN: Says, "Hey guys.
Seems like I'm very late today."
No worries, BHAVIK, we're just getting started.
And the code is on GitHub.
I'll try to post that URL again.
NICK WONG: And at basically every major point, I will push all of the code
that we have so far to GitHub.
If you were around last week, they actually did a whole stream on GitHub
and how oh this very powerful tool--
COLTON OGDEN: Oh yeah, with Kareem.
Yeah, on Twitch and on YouTube we do have a Git and GitHub stream
with Kareem Zidane, Kareem the Dream.
NICK WONG: Kareem the Dream.
All right.
And then, I'm going to just switch over to our testing folder
and put some other dark images in there.
This will let us get a good gauge of just
whether or not we're accurately measuring things.
I think in the previous folder for training
I put about 10 images for testing data.
Oh, Black Mirror.
What a fitting, very appropriate thing to test on.
I put about 5.
There's not really a hard number on how many of anything you should have,
but I generally try to go for nice numbers that I think
are really cool, like 5, 10, and 42.
I think the more data you get, in general
you can do better, as long as your data brings up new patterns.
So then we're going to just bring in some more images.
Sorry, this part is not particularly interesting.
I'm not writing any code.
COLTON OGDEN: Hey, it's from scratch.
The whole thing is from scratch.
NICK WONG: It is from scratch.
So you could literally follow along click by click,
and you will technically do this.
COLTON OGDEN: And choose their own images even.
NICK WONG: Yeah, you could even pick your own images.
This is meant to be as generalizable as I hope possible.
And so we're just picking up very light colored--
oh, sorry.
I picked up a few too many.
I like to keep my data sets balanced.
So if I have 10 in dark, I'm going to try and keep 10 in light.
I am a little--
weirdly enough, you wouldn't believe this by looking at my desktop,
I do like to keep things organized when I code.
I think when I organize my life, not so much, but that's OK.
COLTON OGDEN: People have different definitions of organized.
It's OK.
NICK WONG: Yeah.
It's a flexible definition.
So yeah, I'm very organized when I write code.
I'm not necessarily so organized when I do the rest of my life.
Cool.
So then I'm going to rename these.
I just think it's a little bit cleaner to have this.
And then we're going to get some light testing data.
Oh.
Actually, I think this is just such a pretty picture.
COLTON OGDEN: Oh yeah, that's a nice picture.
NICK WONG: It's very cool.
Oh, I thought it was a unicorn.
I was super excited for a second.
I thought it was a unicorn.
COLTON OGDEN: Can explain again the difference between the test and the--
what was the other category?
NICK WONG: Yeah.
So we have test and training data, and there's also a predict.
COLTON OGDEN: Predict, that's--
test and predict, what's the difference between those two?
NICK WONG: All right.
So testing data is going to be said-- at every step
that we run our machine learning model, we'll see how it did on testing data.
That won't be incorporated into the model, but it will tell us,
like, how are we actually doing?
Because on its training data, it'll generally go up
as it starts to just memorize things.
So we want to see, on data it's never seen before,
is it actually getting better?
Are we actually seeing some improvement, or are we kind of at a loss here?
And so we'll see a little bit clearer what exactly I'm
talking about when we actually start to run some models.
But basically, the testing data shows us, how are we actually doing?
The training data we generally should improve at.
If you're staying at just this flat 50/50,
you're probably doing something wrong in your model.
In fact, there is an error I discovered earlier
that I think would be really good for me to show everyone
that I like to mess up.
Cool.
And then we have two images-- oh, and predict.
Sorry, I didn't answer that part.
The predicting data is, after we've trained our model
and we have it thinking and it thinks it knows what it's talking about,
then we're going to test it on data that it doesn't really
get thrown into any sort of framework.
This is really the use case of our model.
So if I wanted to use it for actively predicting things,
then we're going to say, hey, let's take a look at how we might do that.
So it might not be in the directory predict.
We might even pass it as a command line argument,
and it just goes and finds it.
But I like to keep it organized there for now
while we're going and developing.
COLTON OGDEN: Make sense.
NICK WONG: Cool.
So we have a bunch of images.
I guess for the sake of copyright, they're not mine.
I literally pulled them off of Google.
You watched me do that.
They are in no way my images, and I certainly do not take credit for them.
I only take credit for knowing how to Google--
or barely.
And so we just went found some random images.
And if you don't believe that those are the same thing,
here are all of my same directories that we just built.
I just use Finder because I think it's a slightly easier interface.
Cool.
So we have all these images now.
Let's see if we can go find them.
So when we say load data, that means that somehow, I
have to get the images that I just put everywhere
and return them to me as test and training data.
So we're going to go and get some of that data.
I'm probably going to be given an image directory or img_dir,
and I probably want to go and maybe enumerate
what's in that image directory.
I happen to know that the structure of this directory
is in a test and train split.
So what we might do is I might say, OK, then my test--
or we'll start with train.
"train_images."
And we'll say "paths."
And so I will probably switch these to abbreviated versions
because I think those are conventional.
So this is our training data, and these are the image paths that we get.
I'm going to use a list comprehension because it's convenient,
but if you are not comfortable with those,
then I think it's somewhat easy to flip this into a for loop--
hopefully not a problem.
If you have any questions on syntax and things, then feel free to let me know.
If you spot any bugs, maybe just chuckle to yourself
and you'll see me find it later when I try to run it.
Cool.
So what we want to do is we're going to do "os.path.join,"
which is a cool command.
It lets you just actually join paths together
without really having to worry about like slashes and things
and double slashes.
And I know that this is going to be in my training directory.
And then I'm going to just do "img_path."
And this is "for img_path in os."--
OK, this is possibly very poorly styled code.
It's going to be a very long line.
We could've split this up.
"in os.listdir(os.path.join)."
Oh man, that maybe was not the best planning.
That's OK.
COLTON OGDEN: The beauty of list comprehensions right there.
NICK WONG: Gotta love list comprehensions.
COLTON OGDEN: At least you can enter in the middle of it, though, to break up,
which is nice.
NICK WONG: That is true.
Well actually, that's a very good point.
We might do that.
So technically what we are doing is we're
taking the variable train_imgs_paths, we're
setting it list equal to this list.
I've realized I'm starting to skip words.
That's kind of cool.
And so what we're doing is we're saying join the paths of the image director we
were handed, the training directory which we know will exist--
we're hardcoding that for now--
and then the image path.
And the image path is given to us by iterating through and assigning
as we go through image path to each of the things or items in this list.
And this list is a list of the directory which is given by, again,
joining our image directory and train.
So what that basically ends up meaning is that we're going to go and collect
all the paths of things that we went and put into our image training directory.
And then-- you really should never copy and paste code. #CS51.
We're going to copy and paste code, and we're going to change some variables.
Gotta love that.
And we're going to say, OK, here's my test img_path, [INAUDIBLE] path,
and this is also in test.
Yeah, live coding will really test your ability to be confident when you do it.
All right.
So I generally like to iteratively debug things.
I don't like doing huge monolithic swaths of what's going on.
So we're going to just run this in our script.
So we're going to have "load_data," and we're
going to say img is the data directory.
We might pass that in as a command line argument later,
and that might trigger some bells.
But now we should be able to run this, and it
should print out to me both of the things that contain our image data.
So then, the thing that we have in here is you'll notice that there's this .DS
Store.
Actually, that causes some bugs sometimes.
It won't cause bugs in this case because Keras ignores it when
it's flowing things from directory.
But if you try to load individual images, that does mess with you.
So just keep that in the back, very far back of your mind.
It is something to keep in mind.
All right.
So now we're going to get into--
oops-- the meat of Keras.
And unfortunately, I don't remember the exact syntax off the top of my head,
but I'll tell you what we're going to go do.
And then we're going to do what every cs person should be very good at.
We're going to look at some documentation
and we're going to remember what that name is.
So I have these paths now.
I know exactly where all of the images are named.
I actually don't necessarily need each of their paths.
I really only need this os.path.join of here.
That's the actual directory.
We're going to use that later.
But I like to have the image paths for displaying which images we're actually
looking at later.
So it's just a future thought focused thing.
But we're going to do what Keras has beautifully built,
which is called a flow from directory.
And the reason that we're going to do that
is because we don't necessarily have a lot of data.
And our data is certainly incomplete.
There's plenty ways make another dark image.
But the reason that I think a flow from directory is brilliant
is because it allows us to take our small data set
and make it into a very large one.
Because an image that looks like this, if I
were to flip it around, to the computer, that's a different image.
But to us, for our purposes, it's identical.
It's still dark.
It didn't change anything about it.
So what's beautiful about Keras's flow from directory
is it randomly does that.
It'll start flipping images.
It might cut them in half.
It can flip them around in all sorts of directions.
It can shrink them.
It can resize them however we like.
And it allows us to create data from just our existing data set.
We don't have to go and find anymore.
COLTON OGDEN: That's a cool [INAUDIBLE].
NICK WONG: So it's a very cool feature.
COLTON OGDEN: BHAVIK_KNIGHT says, "We don't have a main, have we?"
NICK WONG: We kind of do.
So what Python actually ends up doing, in a nutshell,
is if the Python script is run as a script from the command line,
then it is going to have main.
But if it's run as a module, it's imported as a module,
then its name, this variable that we never really instantiated,
is not going to be main.
I actually don't know what it is if you run it as a module.
But yeah, no worries there.
Let's see.
"Am I supposed to log in?"
TWITCHHELLOWORLD says, "Am I supposed to log in to GitHub
to see the code at the URL?"
You don't need to log in it.
Is a public repository.
There is no code there currently.
After we load all of our data in, I'll push that to GitHub
so that everyone can see what's going on there.
Cool.
So we have Keras.
The documentation got updated I say recently.
I mean within the past five months or so,
which I'm really happy about because for a little bit there
it was a little confusing.
And I know that down at the bottom they do this whole thing talking
about image pre-processing.
And that's just a bunch of stuff that I happen to know by reading through this.
I know that I'm going to want to do some sort of actual generator.
But this is not the exact one I want.
This one is.
So this little bit of code that is in the Keras documentation--
I take no credit for writing it--
is basically what we're looking for.
This gives us the ability to do our data generators, which
are that whole applying random operations to our data set.
And then we do these actual generators, which are
going to be the flowing from directory.
It's going to pull all our data in to some sort of directory--
or from the directories that we listed, and it's going to do stuff with them.
We're going to do stuff with it.
So I'm going to literally copy and paste this code.
Not necessarily the best model, I think, for doing things, but--
COLTON OGDEN: This is true "from scratch" programming today.
NICK WONG: This really is "from scratch" programming.
I did actually walk through this with myself a little bit last night,
but I deleted all of that code so that you
would get a legitimate version of this and you
would see that there is a lot of effort and thought
and messiness that goes into this.
So we'll notice that in VSC, Image Data Generator is
highlighted because it doesn't exist.
We'll deal with that in a second.
It's just an import that we need.
I'm not a huge fan of their style.
I prefer this one.
But that's OK.
So in this train_datagen, we know that rescale basically--
sorry.
I keep saying "we know."
I apologize.
I mean just rescale is going to allow us to reshape the image's just size.
It's literally taking the image and just making it--
trimming it this way or that way.
Sheer range is not something we're going to actually deal with,
though it would be totally valid.
And that takes an image and just--
you can think of it as--
just cuts off parts of it.
And that's definitely a useful operation for us, but not necessarily
in this case.
I don't necessarily need zoom range.
I think there's one called vertical flip.
And so vertical and horizontal flips do exactly what they sound like.
They flip the data.
And we're going to use this for now.
So it allows you to rescale stuff to grayscale
for this, allows you to do all sorts of rescaling here as well as rescaling
the image size.
I should actually double check on what rescale does.
That's OK.
We'll look at it in a second.
But then, horizontal flipping and vertical flipping
allows us to create a bunch of new data-- a little bit of new data,
where you can do a horizontal flip, you can do a vertical flip,
or you can do both.
And those will create new images for the computer.
And then this test_datagen is going to just be our testing data.
So it's the same kind of thing, but for the test
data that's going to be used to validate our model.
And this one has only the rescaling factor.
Cool.
So then we get into the generator.
And so this is where you see this flow from directory.
It's one of my--
I think my favorite command that Keras has because it allows
you to do this extra data creation.
We are actually-- this happens to be for a binary classifier.
Cool.
So it's still going to be a binary classifier.
And you'll notice that there's this thing called target_size,
and then there's this data directory.
So we are going to change those two things.
So I'm going to just copy a little bit of what I had above.
And that's my training directory, is just
os.path.join this, which is our image directory that we
supplied via the argument, and the testing
directory, which is just called test.
And we now have two directories which hopefully have two folders within them
that have a bunch of images in them.
That's what Keras expects.
And then, this target size is the size of the image.
It assumes that they're of any dimensionality.
We rescaled them so that they are only two-dimensional images.
Color images are actually in three dimensions.
We have our width and our height, and then we have our colors.
We have R, G, and B. If you've been doing CS50 stuff,
you actually know exactly how that works probably very well.
And so I'm going to change our target size a little bit to be 500 by 500.
They don't have to be squares.
I just think it's kind of cool.
And it won't really matter for us.
A dark image that has been reshaped and distorted
but is still all the same pixels is identical to us for these purposes
as the image that's in its original size.
And then, our batch size--
and this actually has a few too few parameters.
I'm going to add a few.
Our batch size let's say is just a nice, clean, round number.
And then we're-- so batch size, meaning how many images does it pull for any
given step within the process?
And then, I don't think we need anything else from this except knowing
where exactly that came from.
And if you look on their documentation, they
do list all of the possible ways in which you could
manipulate the data before it comes in.
There are some really cool ones.
Fill mode I think is actually really nifty.
And that basically just inputs some data for you
if you need it to fill out images that aren't all the same size.
It allows you to--
instead of shrinking them or expanding them,
it lets you just fill in the extra space to the maximum image size.
This is the actual split between the two,
so it reserves some fraction of things for validation.
So if you just had training data, then you could split some of them
to be only validation data.
And there's some other stuff that goes into this.
But we have this ImageDataGenerator class.
And you'll notice that it seems to come from here.
That's the full module path.
So what we're going to do is I'm going to copy that,
and we're going to paste that here.
"from keras.preprocessing.igame import ImageDataGenerator."
And that's how I would convert those full pads to, hopefully,
what actually does its job.
We'll see if that's true.
That's the intuition, at least.
And you'll notice this actually went away because IntelliSense thinks
that I've at least done it correctly.
Hopefully, it's right.
And then the rest of this we're going to leave on its own.
So we have our validation-- or our train generator and our validation generator.
And those are going to be what our model actually ends up getting fitted to
and trained on.
Trained on and fitted to are actually pretty much synonyms in this sense.
So we have our train generator, and I'm also
going to return the validation generator.
Cool.
So we've loaded our data, and hopefully we end up actually getting data.
I'm also going to now excepts those arguments.
I'm going to call it train_data and test_data to be consistent with myself
from the way that I labeled my directories.
You're welcome to change those variable names.
I think that would count as some sort of consistency points.
And we're going to run this and make sure that nothing happens.
That's literally what we want.
We want it to tell us the images it found and do nothing else.
And that's exactly what it did.
And so I'm very glad that that's what it did,
and also that is exactly what we were expecting, where it found
20 images belonging to two classes.
So you might intuit--
oh, that was our training data, which is good
because that was the one we ran first, so I'm glad it came out first.
And then it found 10 images also belonging to two classes.
We're still doing binary.
We've got two classes and 10 images.
We know that that was our testing data.
So everything seems to check out.
And hey, we're still using TensorFlow.
So I'm going to git at add.
I am in the habit of saying git add dot.
Not a great habit.
I'll git add img and I'll git add run.py.
COLTON OGDEN: I'm in the same bad habit as well.
NICK WONG: It's a terrible habit.
Someone in our CS61 class actually committed the entire subversion tree
of--
it was some Unix command.
They accidentally pulled an entire Unix subversion tree into it.
COLTON OGDEN: Probably not what they intended to do.
NICK WONG: Was certainly not what they intended to do.
They actually crashed our grading server--
COLTON OGDEN: Oh, nice.
NICK WONG: --which was really kind of cute.
Everyone was freaking out.
They were all like, oh no, I can't get my grades in.
I'm going to fail.
The professors are like, we'll just extend the deadline.
It's fine.
All right.
Cool.
So that has now been pitched to the GitHub,
so if you would like to grab it, you're welcome to pull that up.
Can anyone-- sorry.
Is that NEOKRES--
COLTON OGDEN: NEOKRES-- it's hard to pronounce that name.
I'm not sure.
NICK WONG: I do not know how to pronounce your name.
I'm so sorry.
But I'm going to call you NEO for now, and if that offends you,
please let me know.
So NEO asked, "Can anyone tell me, is there a link for C++ and JavaScript
Harvard video lectures/sections if there are any?"
I don't necessarily know from my other classes.
I know CS50 has some great videos on JavaScript.
I don't believe we have any on C++.
COLTON OGDEN: Yeah.
CS50 doesn't do any-- hasn't done any C++.
We have done some JavaScript videos in the past,
and Brian Yu's web course, I believe, goes into some JavaScript
if you want to go to that URL for some JavaScript stuff.
But yeah, no C++ yet.
We'll keep it in mind for maybe some future stream videos.
NICK WONG: Yeah.
I think that'd be very fun to go through.
C++ is a great language.
It does all sorts of cool things.
JavaScript I have a lot of beef with, but it is technically
a very useful language.
I think it does a lot of very strange things.
And actually, speaking of JavaScript, TensorFlow,
which is the back end of what we're doing here,
was recent-- well, recently implemented in JavaScript by Google.
And so they released that, and they have all sorts of really cool demonstrations
on using machine learning on the web browser.
So that's actually what my winter break will probably consist of,
and I'm super excited for.
COLTON OGDEN: That and the web assembly.
NICK WONG: Yes.
COLTON OGDEN: That's be huge.
NICK WONG: Web assembly I think is a huge just advancement of browser
technology.
I'm super excited for that too.
COLTON OGDEN: It feels like those two will go hand in hand.
NICK WONG: So many go.
I think so.
I think that that will actually vastly improve
the abilities of what we see browsers doing, which given
Spectre and Meltdown makes me a little bit nervous.
I don't know if I want my browser to be able to do a lot more than it already
can.
But I think it'll be cool as we work out kinks.
All right.
So we have built our loading data, and I'm
reasonably confident that we have loaded it correctly.
You'll notice that when I load the data here,
I actually don't use either of these two.
I'm realizing that in my head I had thought a little bit too far.
So these two are not necessarily useful.
I'll comment them out for us, and we might revisit them later.
Hopefully, we will revisit them later.
All right.
That's actually one of my favorite features of Visual Studio Code,
is being able to just collapse the code of a function.
And there are a lot of IDEs that do that.
I think a CS50 ID does not, though, and it's really annoying.
Cool.
So we're going to-- oops.
I was going to say we're going to fit the model, but we don't have a model.
My next thing would be, oh, OK, let's train our model on things,
but we don't actually have a model to train.
So let's build a model.
And this is, I think, one of the coolest parts of what we get to do.
So building a model.
I'm going to create a kind of "global" variable.
It is global, actually.
I don't know why I put quotes around that.
And I'm going to call that Sequential.
Or sorry, I'm going to call it m, and it is a Sequential model.
And that's going to be global because all of our functions
are going to access it.
And you might be thinking to yourself, oh, this
might fit really well in a class, and you'd be absolutely right.
This is actually really well implemented as a class that has certain methods.
But we're building it as a script, and here we are.
So we're going to use build_model to actually construct our model
and see what layers we can add to it and things like that.
So the first thing that I might want to do is m.add.
And that's going to be the way that we add layers to our model.
They're going to be sequential so that if we add a layer,
that's the last layer.
The last added layer is the last layer that we are going to go through.
All of our data we'll go through that layer at the end.
So the first layer that we're going through
is the first one that we add, and we're going to have that be-- oops, "Desne".
And Dense is just your standard, good old-fashioned neural network layer.
It basically has a bunch of nodes in it, and those nodes
are going to look like that's the number of nodes,
roughly, that it has inside of it.
And that's pretty much all you necessarily
need to know about what it does as far as technical specs go.
But we're going to add a bunch of parameters
to it because this is our first model.
Or sorry, our first layer.
And in our first layer, we need to tell it
the input size that it should expect.
So this sort of thing is always very annoying.
Dense requires a bunch of parameters.
I know input size is one of them, so we're going to give it the input size.
And the input size is going to be something like--
sorry, we specified 500, 500--
3?
Question mark?
That 3 goes either there or on the other side.
I can't always remember.
It's close to that.
We're going to give it an activation.
And with binary classifiers, they use a sigmoid activation.
And you might be wondering, what is an activation?
Great question.
COLTON OGDEN: What's a sigmoid?
NICK WONG: What's a sigmoid, also a great question.
And so basically, what that means is activations
are the initializations of our layers.
So our layers, in reality, they're just a bunch of numbers.
It's really just a weighting of numbers.
And then those numbers are going to be modified as we go,
and then that weighting will change to hopefully give us some answer.
And so those numbers, they need an initial value.
You might default to 0, but that actually, based on the math behind it,
doesn't necessarily give us a whole lot to work with.
Sometimes that can actually cause you to get stuck in one particularly
version of your model.
And so this activation is an initialization to it.
It's a initial set of values.
And so sigmoid, you can think of it as, like, if you were to plot it out,
it has that classic curve structure to it.
And that's going to be what you actually would imagine
if you were to plot out all of the activation
or all of the individual nodes' individual values
from an arbitrary left to right.
It's a little bit hard to describe without the theory underneath,
but that's roughly what's going on.
We're setting a bunch of numbers equal to a pattern, roughly speaking.
And all that means for us is that a sigmoid response is
really good for binary classifiers because binary classifiers have
one of two answers.
It's either up or down.
It's either 1 or 0.
And so a sigmoid response, what that might look like,
if you know anything about signals processing or filtering
of images or data, or even music, then a sigmoid response
makes it really, really hard to land in the middle.
It's very difficult for me to be halfway between.
It's really easy for me to be 1 or 0.
And so that is basically what we're looking for.
That's the behavior we're looking for in this sort of model.
So we're going to stick with sigmoid activations for most things.
We will see a bug that I introduced to myself last night, where
if you choose the wrong activation, we can actually
get totally random behavior from our model, and that's kind of cool.
But we're going to stick with sigmoids for now.
COLTON OGDEN: METAL_EAGLE has a question for you
if you want to pick that one up.
NICK WONG: Oh, awesome.
So a METAL_EAGLE asked, "What is the problem of 'git add.'?
Should you have a habit to run git status beforehand
to make sure that running git add will not break
anything you do not want to commit?"
I have broken all sorts of things that I did not want to commit.
I have committed API keys.
I've committed my only personal keys to things,
depending on what you're building.
So yes, you should definitely make sure.
Running git status is another really good habit
that I do not have to make sure what you're actually committing.
"git add."
is problematic, I think, because let's say
that you have something in your directory
that is not part of your project.
It just happened to be in there because you were reusing some old code
or you wanted to look at something for an example.
And I've certainly done that.
You probably don't want to add that to your git project.
You don't want to put it in the git repository history.
Because of the way git works-- it's a feature, not a bug--
you have access to all the previous things.
So-- excuse me.
If you were to add some sort of API key, for example, then using "git add."
means that that is now permanently part of your git history.
GitHub has a really cool tool.
It's GitGuardians, actually.
They monitor perpetually for API keys, and they'll email you
and they'll block your repository for a little bit
until you get that sorted out.
And that's really cool.
COLTON OGDEN: That's a cool feature.
NICK WONG: They've emailed me a couple times.
I occasionally am just really trying to go, and I'm a little careless.
COLTON OGDEN: Do as Nick says, not as he does.
NICK WONG: Right, exactly.
I've picked up some bad habits along the way.
TWITCHHELLOWORLD says, "I couldn't see the Livestream in the app
so I missed the start.
Are you in CS50 sandbox or somewhere in TensorFlow?
NICK WONG: Yes.
So we're not-- sorry, we're not in the CS50 sandbox.
We are using TensorFlow.
We are in Visual Studio Code on my personal-- on my local environment.
So I'm just using this on my laptop, and I will be pushing code in chunks
as we go to GitHub.
So basically, whenever I get a chunk of code working,
then we'll actually push that to you guys,
and you're welcome to copy it down and see how it works on your own machine.
You will need TensorFlow.
You will need Keras.
You also need Pillow.
You will also need opencv.
Actually, I can make that really easy because Python has the freeze feature.
So I can freeze, which basically means take all of the current Python packages
you have and output them.
So if I just run "pip freeze," then it'll
output all the packages that I've downloaded and I use.
And there's a few more than you might think.
I only downloaded Keras, but Keras, for example, downloads three other--
or like three packages.
So this is actually a really good way to transfer what
Python things you're using.
And in this command, I just piped all of that output into a requirements.txt.
It's a very common feature on Python packages and modules and things.
And if you wanted to build from that, you
could do pip install -r of requirements.txt,
and that'll pull all those requirements in the correct versions,
or at least the version I'm using, to you.
So I'll also push that to-- oops--
to GitHub.
So we'll git add-- oops.
And git commit.
COLTON OGDEN: Thanks also to MKLOPPENBURG
for pitching in the React course that Jordan Hayashi taught,
which did cover JavaScript, and for mentioning
that we are using the VS code.
NICK WONG: Awesome.
Thank you.
All right.
So you should have a requirements.txt file in the repository now,
and that should help if you're trying to get everything set up.
Cool.
So we now want to continue adding layers to our model,
and we can keep things actually very simple.
So we add this dense layer.
I'm positive I'm missing a parameter.
I think it goes in the beginning there.
And we're going to figure that out when it doesn't compile.
But this dense parameter is just your input.
That's the first thing that's available to us
whenever our data goes through our model.
The next thing that we want to be able to add is--
our models are images.
Images in general are displayed as arrays of some sort of data.
And in our code, they're also arrays.
They're NumPy arrays, but they are still arrays.
And we're going to need to modify that a little bit.
We need to tinker around with what exactly we're seeing.
So actually, we're going to change this beginning
layer to be a convolutional layer.
Ah, that's what I was forgetting.
OK.
We're to use a convolutional layer at the top.
It still uses basically all the same parameters,
except it has an additional parameter which is its filter size.
And we're going to use three by three.
That might be a name parameter.
I'm not entirely sure.
And so convolution is a operation that you can do on signals data
from anything, from visual signals to auditory signals
to electronic ones in general.
And a convolution, mathematically speaking, is you take a signal
and you take a response, and you loop the response
and pass it over the signal.
And if that sounds complex and difficult to understand, it is.
I don't necessarily fully understand it either.
COLTON OGDEN: Is it like feedback almost?
NICK WONG: Yeah, exactly.
It's like your response to what's going on in the signal or the way
that two signals might interact.
I would think of it maybe like that.
COLTON OGDEN: That makes sense.
NICK WONG: Yeah.
COLTON OGDEN: Almost like the messing with the image data that we had too.
Feels similar.
NICK WONG: Right.
Yeah, exactly.
It's just a way for us to manipulate the image data.
What convolution will do for us is it--
I believe in this case, if I'm not losing my marbles,
it is convolving the image with this three by three array.
And what that's going to do for us is make our data little bit more--
not centralized, because that's actually what our next layer is going to do--
but it makes our data more readable and friendly to the machine.
So it gives it some features that make our machine a little bit happier.
COLTON OGDEN: Massaging the data.
NICK WONG: Exactly.
That's exactly what it does.
All right.
COLTON OGDEN: BHAVIK_KNIGHT says, "Why not pip3?"
Why are you using pip instead of pip3?
NICK WONG: Ah.
So actually, if you look at which pip or pip --version, this actually is pip3.
I happen to use pip because it's alias to the same command.
So pip3 would also work.
There's not necessarily a difference between the two.
However, there is a difference if you're on a Mac
and you use the installed version of pip versus the system installed
version of pip.
And that is one of the most annoying bugs you will ever
encounter, almost guaranteed.
COLTON OGDEN: CYLVTWITCH says, "Will the videos be on YouTube
so I can watch them later?"
Yes, they will be.
If you just go to YouTube and type CS50 on Twitch,
you should be able to see our playlist.
And it's also see CS50's YouTube channel where David posts his lectures
and a bunch of other course videos.
NICK WONG: Awesome.
COLTON OGDEN: GHASSEN says, "Hello, are there any interns open
in Silicon Valley?"
I would Google and see.
I'm sure there are internships available,
but none that necessarily CS50 is directly affiliated with
or knows of at this moment.
But definitely do some Googling and see, and I'm
sure there are links on those websites that you could reach out
to get some more information on them.
NICK WONG: And BHAVIK_KNIGHT actually pointed out
that convolution is output going into the input again.
That is actually-- yeah, exactly.
That is what this convolution does.
So it's not the general definition of convolution,
which is what I was I guess stumbling through,
but it is the convolution that happens here.
It convolves the image with itself.
Thank you.
That was something that was very important
that I could not for the life of me remember how to talk about.
Cool.
So now we're also--
COLTON OGDEN: You could say it's a little convoluted.
NICK WONG: It's a little convoluted.
That's awesome.
COLTON OGDEN: Zing.
NICK WONG: Man, I love puns.
Puns are the light of my life.
All right.
So we can also add this MaxPooling2D layer, which sounds very--
sorry.
I was reading the comments, and it's kind of funny.
Sometimes you guys have just these entertaining comments OK.
So we're adding a MaxPooling2D layer, and that sounds kind of ridiculous,
but it's not.
Max pooling just means that we're looking at the maximum of the values
in a given range or frame.
We're pooling all those together, and that's
going to be representative of that actual area.
So you could imagine it as taking a 5 by 5 array, and the maximum of that
is now representative of all of that five by five array.
So it could be thought of as like feature--
man, the word is escaping me.
Combining features or feature combination.
And it basically just means that some of the data in our image will get ignored,
but we're assuming that the maximum part of our data is the important part.
There's also min pooling and there's average pooling.
There's a bunch of other ways that you can pool data.
There's all sorts of things you can do with it.
And then, 2D means two dimensional.
You're using two-dimensional data.
So the parameters to MaxPooling2D are something like filter_size.
And this is just that, like how much of a--
like what's the size of the pool we're using.
Activation is definitely one of those parameters, I think.
OK.
I say "definitely," and then I said, "I think."
That's self-contradicting.
I don't know if it takes anything else.
We'll find out.
This is generally something where I don't necessarily
memorize it because I can just go look at documentation.
And then, we're still working with an image which is not super clean.
Or sorry, it's not super easy for our machine to then put out a 1 or a 0.
So what we might do is something like this--
model that add of a Flatten layer.
And that pretty much does exactly as its name.
It just flattens the image into a one-dimensional stream.
And so if you did CS50's resize recover and you played around
with images a lot, you could think of them as all having been flattened.
So they're no longer represented in two-dimensional arrays.
They're represented like a stream of data.
And there is some breakpoint that's logical, but to a computer,
those breakpoints don't matter.
They're kind of irrelevant.
So we can actually then flatten our images.
And now that we've done this max pooling and convolution of the image,
we can flatten everything and then just feed that into standard dense neural
network models.
And that should be pretty reasonable for us.
And we're actually going to do exactly that.
We're going to add a dense layer to do some kind of minor computation.
I like to use powers of two.
That is certainly not a--
it's not a requirement, but I like to use them.
And that's just the number of nodes within our layer, roughly speaking.
It's not exactly correct, but it's close enough.
And then we're going to add our final layer.
And this one has a specific set of parameters.
It has to have the number of nodes within it be actually one.
And the reason for that is we're using a binary classifier,
so we want it to output only using one node or one number.
We don't want it to output a bunch of nodes because then each of those nodes
might be weighted to something else.
They might represent non-existent or imaginary classes.
So we don't want that.
Cool.
And then we're going to give that one a sigmoid activation as well.
So now, running through this in concept would work,
but you'll notice that in IntelliSense, it actually
highlights all of these things or all of the layers
that we just added because they haven't been imported yet.
So what we're going to do is import them.
And I believe it's "from keras.layers," hopefully, "import Dense,"
can't remember the name, "Conv2D," and we'll say MaxPooling.
Oh, I love IntelliSense.
It pulls things that I don't even have up here yet.
"Flatten."
And I think that's it.
So those are all the layers that we tried to use.
Oops, that's not what I want.
And now that we have everything going--
oh, I killed my virtual environment.
That's a bummer.
We should be able to get this model built.
Now, the problem is that we haven't necessarily linked them yet.
But I also think I'm missing a parameter here.
I think I might have not named it correctly.
But that's OK.
We'll be told that very shortly by our Python run.py--
well, if we call build_model.
So build_model is a little bit strange, and it
deviates from the standard paradigm that I
like to use in that it has a side effect and doesn't actually return anything
to us.
So build_model is kind of like print, where
you just run it and things happen, and you don't necessarily
know what's going on underneath.
I'm not a huge fan of that.
I prefer a functional style model where you get a return given some input,
but that's OK.
We're using this more as a weird hybrid between scripting and modular building.
So it's kind of halfway.
So we can go ahead and run this, and it should
run without problems, minus the fact that I think I have-- oh OK.
Well, I also can't spell.
No surprise there.
Now that I spelled that correctly, this should run mostly correct.
Yes.
input_size size is not a keyword.
We're going to go and look at some documentation.
You would think I would know that part, though.
I believe it is.
Well, OK.
I'm obviously wrong.
It's not input_size.
But I know there is something in there.
Let's look at-- oh, right.
Sorry.
Convolutional layers.
That's what I was trying to look at.
Conv1D.
Conv2D.
There's all these things.
So we know that we need a filter.
This is the number of filters that the convolution's going to have.
The kernel size.
So those are two positional arguments.
And I actually might have-- yeah, passed those correctly.
Cool.
I just forgot how to name it input or tell it what kind of input it takes.
Ah, input_shape.
input_shape.
There we go.
So that's correct.
So what this actually does--
and you might be wondering-- oh, it's probably filter shape as well.
Or size filter.
There's all sorts of variations on these.
So what this does is-- we didn't actually
have to pass an input shape to all of the other filters
because they can intuit them from the previous one
or from the previous layer that was being used.
And so that is a cool feature, and it's certainly one of the benefits of Keras,
is you don't necessarily need to keep track
of all of the individual things that are going on in between.
You can just stick to adding layers in a sequential order.
All right.
I apparently don't know the name of that.
One of these days.
Let's see.
Where do they put their--
they might have called that Merge, or Merge is their own layer.
I haven't seen that one before.
Add.
Subtract.
Concatenate.
Oh, well we can also use MaxPooling.
And this will bring up--
wow.
Or it won't.
Oh wait, there we go.
Cool.
MaxPooling2D.
It seems to take a second.
Ah, OK.
Here.
We'll look at that.
MaxPooling2D takes not-- oh, it's pool_size.
That makes sense.
That's kind of intuitive.
All right.
Cool.
So we have corrected our syntax, and I believe that fixes
all the rest of our syntax errors.
And that should be able to--
or not.
Did I misspell that?
Yeah.
Doo-doo-doo.
Ah, it might not take an activation.
You're really seeing this live.
You quite actually--
COLTON OGDEN: It's part of the fun, part of the charm.
NICK WONG: It is part of the actual fun to this.
There we go.
It's literally just tinkering around with what exactly I remember and what I
don't.
Cool.
So we have now added correctly a convolutional two-dimensional layer,
a max pooling two-dimensional layer, a flattening layer, and then
two dense layers, the last of which is our output one.
And that's going to return to us a number that is either 0 or 1.
Well, kind of.
It will return to us a number that ranges from 0 to 1.
It won't necessarily be 0.
COLTON OGDEN: We have a couple of comments
here if you want to read some of these.
NICK WONG: Cool, yeah.
COLTON OGDEN: OK.
So GHASSEN says, "Here in Tunisia we study many, many technologies,
and access to Silicon Valley is restricted.
I want a guaranteed choice because that would be very useful."
I'm definitely not an expert on the programs that Harvard offers.
I would probably just say Google for "international transfer programs."
I know we definitely have Harvard students from Tunisia who have
come here, but I can't speak on that.
I'm not an authority on that. and I don't know anything
about anything related to Silicon Valley directly
because just familiar with what goes on here in Cambridge
at Harvard University, I'm afraid.
But definitely do dig around a little bit,
see what the research or the transfer programs look like.
Look up Silicon Valley companies.
See if they have any programs related to international work.
I know that some companies probably do have something related to that,
although I'm not going to say--
I can't speak on behalf of international work related laws, so I'm not sure.
But again, I would defer to Google on that, probably.
So afraid I can't give any better information than that,
but I'm not an expert on that.
42FORCE says, "Thanks for doing this."
NICK WONG: Oh, thank you.
Yeah, thank you for being here and watching.
We appreciate it.
"What structure is a Sequential?"
So Sequential is actually its own class.
It's given to us by Keras's model's submodule.
And it does work like a set, and very nice intuition on using--
it has the .add method.
But that actually is unrelated to the facts of how a set works.
I believe they do use an underlying structure.
It's not a set, but I believe there is an underlying list
to keep track of which layers you added when.
But that add method has a bunch of other stuff
that goes on underneath to verify your layer
and translate it accordingly TensorFlow-wise.
So very nice intuition, but it actually is Keras specific.
And thank you, NUWANDA, for pointing out I did miss an s in Keras.
Same thing at BELLA_KIRS.
I did miss the s.
I did miss that.
COLTON OGDEN: Part of the fun of the live coding.
NICK WONG: It is one of the great parts.
COLTON OGDEN: I feel 100%.
NICK WONG: Yeah.
Awesome.
Are there any other comments that were Going on here
COLTON OGDEN: I think we grabbed all the questions there.
NICK WONG: Awesome.
All right.
So I just pushed the build_model part, so you should have
access to building the actual model.
We are going to revisit the way this model works because it probably
won't work super well.
It's not very complex, but it does work in concept.
So now we get to the exciting part of fitting our model to our data.
We have a model.
It's been built. We now just have to actually fit it or train.
So I'm actually going to use the word train because that's
what I called our data sub directories.
But we're also going to use just this--
it's actually basically a one-liner.
It doesn't really return anything to us.
It just takes the model and says, you know what?
Now here's a bunch of data.
Tell us what happens when we shove it through you.
And we give you the answers along the way.
So again, syntactically this is going to be kind of fun.
Go figure.
Even after years of doing CS, and years of working with the same modules too,
I'm still looking up syntax occasionally.
It happens.
It's OK.
So we're going to-- in our train model, we're
going to want to have both the training data and the testing data.
We're going to use the testing data at each step to validate what's going on.
The training data is going to do exactly that.
It's going help us train our model.
So we're going to use m.fit_generator.
And that's the actual parameter.
It's very similar to m.fit, except I'm not supplying manual data,
I'm supplying a generator-- it's very similar in concept to Python
generators--
that just generates data as we go.
It'll never run out, which is really cool.
And I remember approximately zero of the syntax for this,
or at least the names of them.
They do change from time to time.
This actually used to be nb_epochs.
And epochs or epochs is just the number of times
that we run through the training set of steps.
So it's how many times do we actually just try to train the model.
And so what this means is that, if we run like 10 epochs--
epochs?
Whatever it is.
You'll know what I mean.
If it sounds funny, that's OK.
You can make fun of me.
So as we go through for 10, on the first time, the model's just totally fresh
and not trained on anything.
But on the second, third, and fourth times, and every time after,
it's going to have picked up knowledge from all the previous times.
So it's basically it practicing.
I would think of it as just practice rounds for the model.
So we're going to start with just 10.
And then we're going to say batch_size equals 20--
COLTON OGDEN: Does it save these iterations
of data collection, data gathering?
NICK WONG: No, it does not.
So it doesn't really save them anywhere on file.
So when you get to the end, it actually just tosses the model,
unless you save it yourself.
So we're actually not going to save them until later when we actually have
some models that we are satisfied with.
But no, it doesn't actually save them on each epoch.
You can-- Keras has these beautifully built structures for things.
They have these callbacks which, if you're familiar
JavaScript callbacks-- come from the same terminology.
But what that lets you do is at the end of every epoch,
you could actually save that version of the model.
And if you're running like 1,000 epochs and you have just these enormous batch
sizes and you really want to make sure that at every stage
you could tell us something about the model at that stage,
then you would want to do that.
And I don't think I'll probably cover that here.
It is a little bit more advanced and a little bit more--
I don't necessarily know it off the top of my head.
But super cool, and Keras has all sorts of powerful tools.
And as I said earlier at the very beginning,
they actually just updated some of the documentation,
and they have all sorts of cool things that I haven't even seen yet.
So I'm really excited about that.
WHIPSTREAK also just asked, "Are we only using Python?"
Yes.
I happen to particularly love Python, although you could also code this
in C++.
I believe they have libraries for it.
I don't know if Java does.
I don't believe it does.
But C++ would also probably work.
It would be a little bit faster, probably.
But we're going to stick to the Python implementation of Keras and TensorFlow.
COLTON OGDEN: It's clear that this is quite a rabbit hole to go into.
NICK WONG: Yeah, it really just fans out.
There's all sorts of things you could do.
There's just so many ways in which you can change what's going on here.
We're going to switch back to get on our fit parameters.
What do we actually need as far as that goes?
COLTON OGDEN: BLUEBOOGER asks, "Could you use R?"
And his name is actually blue in our chat
too, which is all the more appropriate.
NICK WONG: That's awesome.
Yes, I believe R has a library.
Actually, BHAVIK_KNIGHT just answered that.
I believe there is a library for it.
I am not super familiar with it off the top of my head.
COLTON OGDEN: I wonder if TensorFlow-- does TensorFlow have--
NICK WONG: Yeah, I believe TensorFlow for R.
COLTON OGDEN: Interesting.
NICK WONG: Oh yeah, look at that.
COLTON OGDEN: Oh, that's cool.
Yeah, TensorFlow does have an R binding, which is nice.
NICK WONG: Yeah, look at that.
COLTON OGDEN: Cool.
NICK WONG: Yeah, so TensorFlow is really beautifully well done.
It's a Google project, and I'm a little bit biased,
but I think they do a really good job.
It's very, very cool what they've been able to put together for us.
Let's see if I can find where fit generator is.
It might be under the Sequential model.
Oh, actually they throw it in image pre-processing.
I do remember that.
So let's go see what they actually do.
We're going to just copy all these parameters because we're original.
And I'm going to paste them all here.
They did a 50--
COLTON OGDEN: Copying and pasting is a real developer technique, though.
NICK WONG: It really is.
It is a very, very valid developer technique.
COLTON OGDEN: Stack overflow. #StackOverflow.
NICK WONG: Exactly.
And so we're going to just change these to our variable names.
Steps per epoch.
2,000 is quite a few.
Generally speaking, you would want to actually run many, many steps on these,
but our data's pretty small, and we can actually get really good results
with pretty low numbers of steps.
COLTON OGDEN: And oftentimes, big companies
are using massive server farms for this, aren't they?
NICK WONG: Yeah, they use enormous--
COLTON OGDEN: Not just like a MacBook--
NICK WONG: Just computational power.
COLTON OGDEN: --or MacBook Pro.
NICK WONG: Yeah.
Yeah, we're running this on my local laptop.
This is my just personal student laptop.
But yeah, a lot of big companies are running
these on just enormous clusters of paralleled networks and things.
They have all sorts of network capable computational sharing.
They have some just supercomputers running it.
They have some really just-- they have a lot of money going into this,
and there's a lot of just people really trying
to get these things to work for all sorts of crazy things.
I know there's a project at Lifts that does something really
cool with simulating a bunch of just real life data in real time,
and they just test random machine learning models on it.
Or not random, but their own personal projects on it.
And that's super cool.
I know Google has, I'm sure, just millions of products
that we don't know about.
There's all sorts of really cool things going on.
All right.
So we have done this train model, which does call the actual fit generator
method.
And what that means is I'm going to then call that in my--
the actual scripted part of this and call train_data and test_data.
And in concept, this should actually run.
This should get us pretty close.
Oh.
OK, I tried to git push.
I made the mistake of going up and pressing Enter without even looking.
That was great.
Gotta love the habit.
COLTON OGDEN: It's a hard habit to break.
NICK WONG: It really is.
Ah.
You must compile your model before using it.
That's entirely reasonable, and I completely forgot about it.
Technically, build_model is incomplete.
I forgot to compile the model.
So compiling it means you're telling TensorFlow--
oh man, I cannot type--
that you are done adding layers to the model.
You want it to be all put together, assembled
into tensors in this beautiful pipeline.
And so that's what we basically end up having going on there,
and that's all we're going to run here.
It also takes a couple of arguments.
For example, it takes an optimizer, which is, mathematically, a tool
that makes things a little bit better as far as how each layer
or how everything takes feedback.
So when it's told what the actual answers were,
it optimizes for making sure that we don't interpret that incorrectly.
I would think of it like that.
That's not exactly what it does, but it is the right idea.
I'm going to use SGD, which is the--
man, I am forgetting names left and right.
That is the name for something.
Actually, we can switch this to Adam, which is-- again,
I don't remember the acronym, but it is a version of an optimizer.
They just these slightly-- well, some actually majorly different things.
And you'll notice some different results, depending on what you're doing
and which optimizer you use.
We're going to also add metrics.
And this takes a list.
We're going to do binary_crossentropy, which sounds really fancy,
and great party term if you're at a party of nerds.
It's a really fun term to throw out there.
There's also categorical_crossentropy.
I think that one's even cooler.
It's even longer.
And so what that basically just means is every time we do--
I got distracted by a comment.
That's really funny.
COLTON OGDEN: Let me go back up to WHIPSTREAKS there.
He says--
NICK WONG: Right, sorry.
Ah, OK.
So, "I was away, only came back now.
Do summary."
Yes.
I don't know exactly where you left, but we are currently trying
to get our model learning on data.
And so what that means for us is that we want
to be able to actually run this train of the model or fit the model to the data.
And what we-- or what I forgot about in the interim
was that I actually have to compile the model, which
means giving it an optimizer and what metrics to use for how well it's doing.
And that's, I think, what will catch you up.
COLTON OGDEN: And the video will be on Twitch as a VOD
to look at later and also on YouTube tonight.
You'll be able to look back on it if you missed anything.
NICK WONG: Awesome.
And then-- ah, we're checking in on who BHAVIK_KNIGHT is.
That's kind of cool.
COLTON OGDEN: Part of the Handmade Hero stream.
I've checked that out a couple of times, actually.
NICK WONG: That's sweet.
COLTON OGDEN: It's nice.
It's a guy who makes a complete game from scratch 100% in C, which
is pretty cool.
NICK WONG: Wow, that's crazy.
COLTON OGDEN: And he's got like 1,000 videos up right now.
It's pretty ridiculous.
NICK WONG: Damn.
COLTON OGDEN: But I guess BHAVIK_KNIGHT is not the same person that's
on the Handmade Hero stream.
NICK WONG: Wow, that's crazy.
COLTON OGDEN: Just happens to be shared by-- another person
happens to have that handle.
NICK WONG: I didn't even know that was possible.
COLTON OGDEN: And then GHASSEN says, "I recommend
to have a live session of parallel programming.
It's just a hard thing to understand."
NICK WONG: Yeah, we can do that.
COLTON OGDEN: We could take a look into that at some point, yeah.
NICK WONG: Actually, it'll come up a little bit here too.
COLTON OGDEN: Yeah, OK.
NICK WONG: We do you use it a little bit.
But yeah, that'd actually be a really fun one.
We could even parallel code it and then see if it runs that way,
motivate it that way.
And actually, that's--
I can't read that name, but--
COLTON OGDEN: It's BLUEBOOGER.
NICK WONG: Oh, BLUEBOOGER.
Ah, my fave.
All right, so BLUEBOOGER then pointed out that we are parallel programming
right now.
COLTON OGDEN: Except I'm not actually programming.
NICK WONG: This is true.
COLTON OGDEN: I'm a bystander in this right now.
NICK WONG: You're a waiting process.
COLTON OGDEN: I am.
NICK WONG: You are the parallel equivalent of a waiting process.
COLTON OGDEN: I am.
NICK WONG: That's awesome.
COLTON OGDEN: An infinitely waiting process.
NICK WONG: Yeah.
The "while true" process.
COLTON OGDEN: A demon thread.
NICK WONG: Yeah.
Oh, man.
I think if you don't appreciate puns and jokes in life,
then it's just kind of sad.
It's just really depressing.
All right.
So we have an optimizer.
We have metrics.
There is a third thing that I am missing,
and I can't remember what that is.
So we're going to go find it in documentation.
COLTON OGDEN: This is part of the live coding thrill, man.
NICK WONG: It really is.
We're looking left and right at documentation.
COLTON OGDEN: It's a good documentation consulting tutorial.
NICK WONG: It really is.
Keras, your documentation, it is a lot better now.
I really appreciate what you guys did on that.
Great job.
It's making my life a lot easier.
It used to be, I think, a little bit of a mess,
but there have been a lot of clean cleanups on it.
Things have been updated.
It's great.
I'm a huge fan of up to date documentation.
If you can read documentation and be like, wow, things work--
oh right, "loss."
There we go.
COLTON OGDEN: 42FORCE says, "What you said,
Nick, is true about puns and such."
NICK WONG: I really appreciate that.
Yes.
That's one of my life philosophies.
Especially when you're sitting up late at night
and just trying not to stay up until the sun comes back up, puns keep you going.
All right.
So I actually misspoke a little bit earlier.
I said the binary_crossentropy was a way of metricizing what's going on.
Apparently, I'm losing my marbles.
That is actually a categorization of loss.
It is, in a way, a way of looking at how things are going wrong,
but loss is basically--
there's a whole lot of theory behind that,
but you could think of it as the way that--
kind of the loss in accuracy that the model had.
So on that run, how poorly was it doing?
The higher it is, the worst it goes.
The metrics are what we're actually using
to display in a human readable format.
Loss doesn't necessarily mean anything to me.
It could be, like, a 42.
That might be meaningful, maybe.
They're more useful as a relative metric.
But metrics on their own, like accuracy, that's super meaningful to me.
It means what it says.
It's an accuracy, usually given as a percentage.
I think mine in this case will be displayed not as a percentage but as
a proportion of 1.
So you'll get like 0.55.
Not great.
0.5 means we're literally guessing.
0.45 means we're doing worse than guessing, so we're anti-guessing.
We got it right, but then flipped our logic.
And then 1 is ideal.
"Ideal" in concept.
Cool.
COLTON OGDEN: Astly-- NUWANDA3333, who's Astley [INAUDIBLE],, says, "Agreed.
Hope you do more sessions, Nick."
NICK WONG: Awesome.
Thank you.
I really appreciate it.
COLTON OGDEN: Tune in next Friday--
NICK WONG: Yeah, we'll be here next Friday.
COLTON OGDEN: --for a LInux command tutorial.
NICK WONG: I very much enjoy that.
I appreciate it a lot.
And now we're going to see if we can get this to run.
That's one of my favorite--
I say that to myself when I'm coding things.
Hey look, it worked.
I'm always-- I'm not actually surprised it worked.
I'm a little bit surprised it worked.
Cool.
So I'll tell you what's going on here while it runs.
It's running exquisitely slowly, which is awesome.
It wasn't running so slowly last night, but we're Livestreaming.
COLTON OGDEN: Exquisitely slowly.
NICK WONG: So slowly.
Everything you Livestream, something that ran in maybe one second
will take four hours.
If you had a download that took a second, it's four hours now.
There's just no getting around that.
COLTON OGDEN: It feels like four hours.
NICK WONG: Yes, especially when you're sitting there just waiting.
But there's a lot for me to explain here,
so we're going to go through what's going on.
So you see this bar increasing on the left.
That's telling you your progress through the number of--
I'll see if I can find it in the code because that
would contextualize what's going on.
The number of steps.
So each one of the-- oh man, that was awful.
Each one of these little bars that goes through on the bottom
here is going to be one step.
Each line is one-- or sorry, is one epoch or epoch.
And that's going to be just-- we're going to run through 10 of them,
as per our parameters.
But then, our steps per epoch is how many steps do we actually
go through trying to train our data.
And that's going to be 20 in this case, and it's
going to count through each one that we walk through.
And then it tells us an ETA.
Those are generally fairly accurate.
I've found them to be pretty reliable.
This is the loss that we mentioned earlier-- not particularly meaningful.
You generally want it to go down.
That's pretty much all we really know.
COLTON OGDEN: Also, your computer is whirring right now.
NICK WONG: It's really funny.
If you run these in class and you're supposed
to be paying attention to lecture, then people
will start to look at you as your computer just
starts to get closer and closer to just exploding.
And I think it freaks people out.
They're like, oh god, why is your computer panicking?
And then we get to this accuracy metric, which is what we mentioned before.
And you'll notice there's actually two of them.
There's accuracy here, and then there's val accuracy.
What that stands for is this is the accuracy on the training data.
So that generally will go up just regardless
as it starts to memorize things.
But the validation accuracy--
excuse me-- should give us some reasonable metric
for how well we're actually doing.
And what we can do is, if this validation accuracy
is substantially higher than the training accuracy,
then you're doing a pretty good job.
You're literally predicting data that you've never seen before better
than the data you have seen before.
And that's not something that happens very often, if ever.
It's something that would be pretty rare.
If the validation accuracy was at 100 and our training accuracy was at,
like, 1/2, then--
oh, sorry.
I forgot that we were on the screen.
Right.
Thank you, Colton.
COLTON OGDEN: There it is.
NICK WONG: I appreciate it.
You'll see it.
I can actually probably move it up a little bit, and that might help.
COLTON OGDEN: Duck down just a touch.
NICK WONG: There we go.
Oh, well that-- it's close enough.
So yeah, that'll show up again when this epoch finishes.
And so validation accuracy versus accuracy are really good to compare.
If validation accuracy is substantially lower than accuracy--
and unfortunately, that happens kind of often--
that means you've probably over-fitted your data.
Because when you're given new data, you're just like,
oops, I don't know what's happening, but given data that you've seen before,
you're pretty good at it.
However, what you might notice-- and it takes a little bit at the end
here-- you'll notice it freezes on that last step in the epoch.
That's because it's running all the validation steps,
and they take a little bit too.
You'll notice that our accuracy is at 0.5,
and it's staying uniformly there, which is not great.
That basically means we're randomly guessing,
and we're randomly guessing really well.
We're just really random.
And what I mean by that is even if I were to flip a coin,
I would probably hover around 50.
I'd have a mean of 50 and an expectation of 50%, 50% heads or 50% tails,
but I wouldn't get exactly 50%, at least not usually.
So the fact that we are sticking at 50% is a result of the fact
that our data's pretty well balanced, but also we're just really random.
We're true randomness here.
We're literally guessing.
And so that's not ideal.
Probably not what we were looking for.
So that means that we now have to debug the actual model.
And I guess we deliberately left this here, where I made some of the mistakes
that I made last night, but I made them again for everyone else to see.
And that's because they're really common mistakes that show up
all the time in machine learning.
So it's really easy, usually, if you're given some sort of Python code or C
code-- how do you debug it?
How do I go about fixing that code?
And the problem with machine learning style things
is we don't necessarily know where to go debugging-wise.
There's no real article on how to debug a machine learning model.
There's a bunch of really technical ones,
but there's none that are really as approachable as this code seems
to imply it should be.
So we have a bunch of things to look at.
How do we build our model?
How are we getting our data?
How are we manipulating our data?
And what kind of optimizer are we using?
And what kind of activations do we have within building our model?
So I'm going to let this keep running, and we'll stay at 50, but that's OK.
It's cool to watch it go.
We can also make this a lot faster.
So there are some other things going on here that
can make our lives a little bit better.
Let's go about making it faster first.
So this is use_multipleprocessing, which is something that I think is correct.
And we're going to say false.
But I'm going to say the number of workers
or the number of independent threads you can spin up is 4.
And what this means is a lot of machine learning stuff
can actually be done in parallel.
I don't have to do it all in one thread on one processor.
I can spread it across as many processors as I want,
limited by the number I have.
And so what I'm going to do is I'm going to say, you know what?
You can actually use 4 workers or 4 threads.
And I turned off multi-processing.
If you use both at the same time, Keras will give you some warning and be like,
yeah, don't do that.
And the reasoning for that being, basically, you're
attempting to do two versions of the same thing.
I say that, and I'm sure someone somewhere just cringed.
They're not actually the same thing, but it would
accomplish a lot of the same goals.
So we're going to just use one instead of the other.
And I'm going to rerun this now.
It should run a little bit faster, which would be ideal for us.
COLTON OGDEN: Machine learning takes a long time.
NICK WONG: Machine learning can be very problematic as far as timing goes.
[INAUDIBLE]
COLTON OGDEN: Learning in real life takes a long time, so it makes sense.
NICK WONG: Yeah, exactly.
And we have one of the most powerful computers in the known universe,
actually, which is crazy.
So it is going very marginally faster, but not necessarily a whole lot.
All right.
So now we need to look at, well, what are we
doing that might be breaking our machine learning model?
Why is it not learning in the way we want it to?
So some things-- and this is where we get into a guess and check
sort of thing, and the programmatic part is
where you're looking, where you're guessing and checking.
So the first place I might look is in the way I built my model.
Is there something inherently that jumps out
that tells me, yeah, this is incorrect, like something here just doesn't work?
So I noticed that my activation here is sigmoid, which is right.
That's what I want for a binary classifier because of the same reasons
that we mentioned earlier.
My activation in all intermediate steps is also sigmoid,
but I could change that.
I could make it relu.
I don't actually know what that stands for.
I did at one point.
I don't anymore.
Or something like tanh.
And both would be reasonable.
There's no real reason to pick--
there is a real reason to pick one versus the other,
but we don't necessarily need to care.
But I'm going to leave that actually as sigmoid.
I'm going to look at the original activation, which is also sigmoid.
And that might be kind of problematic.
It might not necessarily be the source of our bug exactly,
but it might also be something to reconsider,
in that we said that a sigmoid activation creates it
so that you're in a binary state.
It's either down or up.
And we don't necessarily-- thank you.
MAGGUS503 just pointed out that relu is rectified linear unit.
Thank you.
COLTON OGDEN: And GHASSEN said, "Can I work with you guys?
I need a Skype interview with a company in America."
So CS50, to the best of my knowledge, is not currently hiring.
When we do hire, we accept all applications.
Normally, they're posted on our website, cs50.harvard.edu.
But like I said, I don't believe we have any job openings currently available.
In terms of getting a Skype interview with a company in America, that's
also something that we probably won't be able to assist with, but definitely
reach out to companies you're interested in working with.
Most companies have links on their website for applications.
Definitely build a portfolio of projects on GitHub or the like
so that you have more code to share with people.
And keep the search going, but I'm sorry that we're
unable to provide any more than that currently.
NICK WONG: But best of luck.
COLTON OGDEN: Yeah, definitely best of luck--
NICK WONG: Very good luck.
COLTON OGDEN: --in finding something.
There's tons of openings, I'm sure.
NICK WONG: Yeah.
All sorts of companies want people who know anything about coding
or are looking for just people to, I guess, put in effort and work.
I think-- we'll go on a short aside, like for four
seconds, that I think that don't ever lose motivation
on these sorts of things.
As someone-- I sent out I think like 104 job applications last year.
COLTON OGDEN: That's dedication.
NICK WONG: I got rejected from most of them.
You're really trying to get a job, and for me it's
like, well, I want to be doing something with my life over the summer.
I want to actually be contributing.
So just don't give up.
Just keep going, keep pushing through.
Rejection-- people very clichedly will say, like, rejection makes you grow,
or rejection teaches you stuff.
But I learned a lot from my rejections, and I think
that they actually do help you a lot.
So [INAUDIBLE].
COLTON OGDEN: Yeah, there's a phrase that I like that's--
who came up with it?
"Be so good they can't ignore you."
NICK WONG: Oh, right.
COLTON OGDEN: Steve Martin, I believe.
NICK WONG: Yeah, I think so.
COLTON OGDEN: That's a good phrase.
Keep working hard, and if you're really, really, really good at what you do,
you just won't help but be able to get something somewhere. that you want.
NICK WONG: I think Lang Lang, the pianist, has a phrase or a story
in his book where there's someone who hates him and is a judge for his piano
competition, and even they can't deny how good he is.
I think that that's really just super cool.
I think that's super inspiring.
All right.
So we did change some stuff about what was going on in our model.
I rectified, so to speak, the activation of our input layer.
I changed it from sigmoid to rectified linear unit or relu.
Relu.
And that actually caused our model to start doing some interesting things.
It's still very slow, and we added some extra workers,
but it's not necessarily going that much faster.
But if we look down at our metrics here--
and I believe Colton might be in one of the metrics,
but we can just raise it up.
COLTON OGDEN: Oh, I'm sorry.
NICK WONG: Oh no, you're all good.
We can just do this.
COLTON OGDEN: Cut off there.
NICK WONG: I think it's above you now.
COLTON OGDEN: I'm trying to go this way.
NICK WONG: There we go.
I think you're good.
I think you're good.
COLTON OGDEN: Cool.
NICK WONG: All right.
So what we can look at here is in this first line,
in that first epoch, what we see is our loss is--
whatever.
It's some set of numbers.
But our accuracies, our accuracy and validation accuracy,
are not 50-- or 0.5.
Sorry.
I'll probably refer to them as percentages,
but they are technically fractions between 0 and 1.
So they're actually 70 on our training data and 100% on our testing data,
on our validation done.
And if you recall or maybe you're currently watching in the future-- huh,
weird--
then you might recall that I said it's pretty rare that you
have your validation accuracy go higher than your actual accuracy
or your accuracy and your training data because you generally
learn you're training data lot better than your validation data.
But in this case, we are using something that's fairly simple,
and we are technically solving a problem that isn't too terrible.
So we actually see that our validation accuracy,
in now two epochs out of three so far, has been higher
than our training accuracy.
And we do happen to have just more variance within our training data set.
It's bigger.
It's double the size.
So it makes sense that we might see something
like this for a pretty simple problem.
So we're going to let that keep running.
It's still very slow.
And something that-- the reason I keep pointing that out
is to motivate the rest of-- or the next part that I'm going to talk about.
But it would be really cool if I could speed that up and not lose accuracy
because that would be-- that's the ideal, if I can make machine
learning model run really quickly, really cheaply, and not lose accuracy.
COLTON OGDEN: Millionaire.
Future millionaire.
NICK WONG: That'd be awesome.
You could you make a start right there on that.
And that's a gross oversimplification of the problem
and certainly not something that we can do,
but there is something that we can do to fix what's going on here.
Because I've mentioned earlier that these images,
the size doesn't really matter that much as long
as we keep most of the important data or most of the features.
We're using some pretty large images.
They're sized 500 by 500.
And so if we're running through, what, 20, 40 images, and we have 500
by 500 for each, well, 500 squared, we're on the order of 10,000,
or I guess 25,000 actual pixels or data points, times 40.
So we're at, at the very least, what is that, 2 million or 200,000?
I'm not very good at math.
And that's OK.
I think we're on the order of a couple 100,000 things
and calculations that have to go through all the time, and that's terrible.
It's definitely not what we're looking for and I guess not
what we really need.
So I would go through and change all of these 500s
to then maybe something like 150, a very common version of that.
The problem is, I wrote it in a bunch of places,
and it's going to just change in one way.
I'm going to use pretty much square images forever.
So this is a pretty good place to switch to some variable.
So im_dim or image dimensionality--
oh actually, that's not a great variable name.
It means something else.
We'll say image-- we'll call this one width.
It actually doesn't matter which one we define as width
and which one we define as height as long as we're self-consistent.
So I'm going to say image height and image width are in that order,
and I'm going to replace all of our 500s with those,
which should, if we don't screw up missing one of them,
should help us out a little bit.
And then, similar to how we defined model up here,
we're going to also define each of those.
And we're going to bring that down to about 150 by 150.
Let's make sure that I didn't leave any spare 500s.
Cool.
And now we know that we can control the image height and width by just
changing a configuration variable.
And in my previous project or in the talk that I gave last year,
the code that I used for that actually has just a configuration module,
and you can just change config variables and it will change everything
about how it works.
And so I think that, generally, moving those out of the actual code
so that you can think of the code in terms of abstractly what it's doing
can clean up a lot of your actual debugging and processing.
So we're going to stop this, the currently running training session.
And it says, "User aborted!"
And that's great.
We're going to run it again but with our smaller sizes,
and we're going to see if that changes the speed at which we go.
And I say that like I didn't know.
COLTON OGDEN: Visibly it looks like it's definitely moving a lot faster.
NICK WONG: Yeah, it's so much faster.
And you'll notice our accuracy has not gone down.
We're still interpreting these images pretty accurately.
And you can think of this in a human context is if I handed you
a really small image versus an enormous poster,
you could tell me pretty quickly whether either one is dark or light.
You could look at the smaller version and be like, yeah, that's dark.
And you look have at the big poster board.
Yeah, that's light.
And that's pretty immediately doable for us.
So this is basically what we're seeing.
We're seeing the machine equivalent of that.
And right now, you could argue that we've over-fit our data.
That's probably true.
But our validation accuracy is still also true.
The problem is that that might lead you to think, oh, well then we
can now use this model and predict on our prediction data, which
is definitely our next step.
But that's not entirely the case.
You'll notice that loss parameter actually has
continued to go down from row to row.
And that tells us something about what we're doing to our data.
It's one of the only times that I'm going to point at loss as something
that we really want to consider.
As that continues going down, it means we're getting closer and closer
to just getting it right.
But that's problematic on our training data
because it means that we're over-fitting.
We're going to just hyper-fit ourselves to this data set,
and we're going to put on our blinders and say, this is all I know,
and this is all I can ever deal with.
That's really problematic for images, where even one pixel difference means
that you have a different image technically.
So we want to prevent that.
And luckily, right before we go into prediction,
we actually have a layer that does that.
So we're going to add that layer in between, sandwiched between two Denses.
And that layer is called Dropout.
And I know we are running a little bit low on time, but that's OK.
I think we are actually going to run perfectly into--
COLTON OGDEN: We can go--
NICK WONG: Finish this.
COLTON OGDEN: --a few over if you need to take a few minutes to--
NICK WONG: All right.
COLTON OGDEN: Fairly flexible format on Twitch.
NICK WONG: Right.
That is a good point.
That's actually one of the beautiful aspects of this.
All right.
So what we've done here is we've added Dropout, which means everyone once
in a while, we're going to just randomly drop out data.
We're going to just ignore it for that training session.
And that's really cool.
It means that we can actually counterbalance the over-fitting
that we might do.
I said 0.2, but I'm going to make it really aggressive.
The term "aggressive dropout" is something that I think is really funny.
It makes me think of a college kid who just drops out of college
and is extremely aggressive about it.
But yeah, we have this aggressive dropout,
which means that we're dropping out a pretty high number of things or data
points at each point.
And that's going to help us prevent ourselves from over-fitting.
I'm also going to, now that we have this whole thing working on up to training,
I'm going to add that to GitHub.
oops.
"Added model training."
And then we can git push.
And so now what we have left to do is--
can we actually try and predict on the data
that we have in our prediction directory?
So that's going to be our last step, and we're going to say that our predict--
our predict on, and we'll say image as well because that's
where our prediction data lives.
So then, in here I'm going to say define predict,
and it's going to take an img_dir, and all
we have to do is let's recall what the actual structure of predict is.
So under Image, under Predict, it's just images.
And what I might say is, OK, then let's actually just take those images,
and we're going to save their paths.
I need to be able to save their actual paths.
And that's actually what I had intended to do over here.
I just conflated the two in my head.
And we now have this lovely line of code.
There is no actual need to do any of this os path joining.
If you'll recall, we're just using listdir
to tell us what images are in that.
And then we're going to img_dir predict of what's going on in there.
And this should give us access to all of the images
that we actually want to use.
Let me actually think through that before I tell you that that's the case.
Ah. os.listdir.
I actually do need the os.path.join.
Oop, not "jin." "join."
That's one of the features I don't like of VSC.
And "predict."
So now we have access to all of the images
that we want to have prediction on.
And I'm just going to call these im paths or img_paths.
And the reason that I'm going to keep track of those
is so that I can tell which image I'm actually making a prediction on later.
And so I have to now load in these images.
And this is where we actually use cv2.
So let's say images is equal to--
and I'm going to use, again, list comprehensions. cv2.imread
of img for img in img_paths.
And this will allow us to actually read in each of the images.
And then I'm going to again do images is cv2.reshape, I believe, of the img.
And we're going to reshape it to our height.
Oops, I called that im_height.
im_height, im_width-- god I love the auto-complete-- by 3.
And this is, again, for img in images.
Cool.
And what this basically does-- and I'm abbreviating the steps
and cutting down a little bit on my narration-- but cv2.imread means
read the image into what we're dealing with.
It reads it in as a NumPy array.
And then this cv2.reshape of the img actually puts it
into the right shape for us, I believe.
I think that's a necessary step.
I might be conflating it with the NumPy version.
And there's also-- yeah, I have a feeling I mixed these two together.
We're going to actually comment that one out for now.
I think I'm talking about the NumBy one.
We have numpy.reshape, and that sounds much more correct,
so I'm actually going to just copy the line I had above.
And this basically just takes the images, like dimensionality,
and reformats them a little bit so that we can actually feed them
into our machine learning model.
Our pre-processing step before up at the top when we load in the data
actually did that for us.
We just didn't really realize it.
Cool.
So this should be all of our actual images.
And then we're going to say predictions--
actually, we can just return from there.
We can return again a list comprehension of m.predict_classes.
I don't know if it's plural.
We'll find out.
Gotta love the just random guessing sometimes.
On img for img in images.
And this will tell us, what does it think each thing is?
However, that's going to return to us something that's a little weird.
And we can actually printed it.
Printed it out?
Who knows?
We can look at what those predictions might look like.
And since we're debugging this and not actually the number of epochs--
or sorry, not the actual model, we just want
to see what it would look like later--
we can actually reduce this to two steps per epoch and one epoch.
And that's going to be really fast.
It's going to just skim through.
Oops.
I just did git push again.
There we go.
Cool.
I have invalid syntax, and that makes sense.
I forgot the word "in" on line 81.
Doo-doo-doo-doo.
That's one of those things where if you're watching someone code,
you see that immediately, and you're like, no!
How did you forget that?
But then when you're coding it live, you're just like, eh.
All right.
Cool.
Cannot reshape an array of size 150804 into shape 150, 150, 3.
I'm going to actually cat a piece of code
I wrote a long time ago to see what that was just
so I can verify what that resize was.
Ah, cv2.resize.
OK, so there was actually a step there.
Sorry.
I'm peeking at some code I wrote earlier.
It's code from a different project, but it does do the right thing.
COLTON OGDEN: Code reuse is an important thing.
NICK WONG: Code reuse.
So OK, we have cv2 resizing the image to the right dimensionality for us,
and then we have NumPy reshaping it into the proper array for what
we're actually looking for.
And that should actually be correct.
The rest of this is roughly true.
And I'm going to stop looking at my previous code now.
Cool.
So then, let's go ahead and print these out
and see what that looks like for us.
42FORCE points out, "Code reuse works all the time."
Yes.
I am generally a big fan.
All right, so we get these just random looking arrays.
They don't look particularly meaningful.
This is a NumPy data structure where you have an array
and it has a super listed in number.
It's 1.
The other one could be 0, but it's actually they're both 1.
And then our dtype is a 32-bit integer.
So that's not super meaningful to us, but if we actually just access
the 0th element of the 0th element of each of those arrays,
we'll just get the numbers back.
And then what we can do is--
I'm actually going to also print out which
image path that's associated with.
So I'm going to say that img_paths at i, i in enumerate of images.
COLTON OGDEN: MKLOPPENBURG says, "Will have to finish this later on.
It's been a joy.
Thanks Colton, Nick for doing this."
NICK WONG: Awesome.
COLTON OGDEN: It's been a good time.
NICK WONG: Thank you.
We appreciate it.
COLTON OGDEN: GHASSEN says, "Can I guys send you my CV so that you take a look?
I built a great resume."
Again, I don't believe CS50 is currently hiring or looking at CVs or resumes.
Normally, when we do, we send out--
we have an apply link setup which is currently not accepting.
But particularly if you're interested in working for a Silicon Valley company,
I would do research into companies in that region.
Check out their website.
Check out their job applications.
And then submit your CV there.
You'll probably have much better luck doing so.
But thank you for the-- thank you for offering to us.
And if we are hiring in the future, definitely submit it.
And then BHAKIV_KNIGHT says, "0, 0 in brackets would do as well for NumPy."
NICK WONG: Yes, that is true.
I think it also works for Python.
I'm just generally a fan of the old-fashioned way.
I'm not that old, but I am certainly an old-fashioned user.
All right.
So I am still copying things to speed things along a little bit.
And then, I will tell us what's going on.
All right.
So we can actually then go from here.
All right.
So we want to be able to map between the numbers and the actual classes
that we're talking about, like light, dark, things like that.
So what we're going to say--
nice.
All right.
Colton got called out for his great people skills.
That's true.
He does have fantastic people skills.
I wanted to make sure that he didn't end up reading his own compliment.
COLTON OGDEN: Thanks, Astley.
Appreciate that.
NICK WONG: He does a great job with people.
All right.
So we're going to define invert mapping.
It's something that I think is just useful.
Given a dictionary, which I'm going to call d,
we're going to have inverted, which is the inverted version
of that dictionary.
And I can say, "for key in d."
And actually, you could say "key, value in d.items."
This might make it a little bit easier on us.
Then we can say inverted at the value is equal to the key.
And this assumes a lot about your dictionary and how it works,
but that's OK.
We can actually prove that those assumptions are valid
because of how Keras defines the class indices dictionary.
But that basically allows us to switch it from--
it gives you the index name or the label name with the number,
and I want the number with the label name mapped to it.
So I just invert the dictionary for us.
There might actually even be a Python function that already exists for that.
I just don't happen to know it.
COLTON OGDEN: BHAVIK_KNIGHT has a dictionary comprehension just there
in the chat.
NICK WONG: Ah.
Yeah, that's perfect.
If you look in the chat, BHAVIK_KNIGHT's version of the dictionary inversion
is exactly what I just wrote out, but in one line and certainly very readable.
So very nice.
Cool.
So now that we have this kind of mapping, we can then just go through.
And for our prediction---
or actually, we're going to say the I think value is what I called it,
value and im_name in predictions.
We can actually now print out what each of those things was.
So "We think that im_name is val."
Oops.
Sorry, not val but rather mapping at val.
Cool.
And if we run this through-- oh, that's not what I wanted.
That's what I wanted.
If we run this through really quickly, then this will not actually
build our model correctly.
It's certainly not the right version of the model.
It's not going to get things right.
But what we are going to have is at the very end,
it'll print predictions on any image that's in our predict directory.
And that's what we see at the very bottom here, is that it says,
we think that the dark_1.jpeg is light, and we think that light_1.jpeg is also
light.
So it got one of them right and not the other one,
which is not super great for us.
COLTON OGDEN: 50%
NICK WONG: 50%.
But we know that if we train things a little bit longer,
we can actually do a lot better.
COLTON OGDEN: FORSUNLIGHT says, "Nick, old is gold."
NICK WONG: I'm a huge fan.
I also love that the old way of labeling global variables
used to be by prepending a g to them.
COLTON OGDEN: Oh yeah.
NICK WONG: So that's actually--
I don't know if you intended that, but it is certainly
the greatest pun on that.
All right.
So what I promised at the beginning--
I can, I guess-- hopefully you'll take on faith
that if we increase the number of times that we train this
and allow it to run its full course, we could actually get this to a point
where it is correctly predicting on light and dark images.
However, I promised that we could do cartoon versus real images,
and so let's actually set up that data.
And we'll find that it's actually really easy.
The way that we set up this structure, all we have to do
is replace the image directory and then re-run our script, and that's it.
And so what we're going to do is I'm going to call this img.old.
And then I'm going to just copy it and create a new one.
Oh, I can't paste it into it's own location.
Bummer.
That's OK.
And we're going to now create this new image which
still has predict, test, and train.
It's going to have two different jpegs in there.
And in test, instead of having dark and light,
it actually has cartoon and real.
COLTON OGDEN: GHASSEN says, "Like swift 3.
Very interesting.
We spoke about SQL last session.
I hope that we can build a common platform
and start talking programming."
Yeah, Swift is a good idea.
I'll maybe see around if anybody knows how to program in Swift
and wants to maybe make a simple application on stream.
If you want to maybe read TWITCH's last comment there.
NICK WONG: Ah.
All right.
So TWITCHHELLOWORLD says, "Thank you.
This is so interesting.
I heard that the later AI that triumphed over the initial program, which
won over the human at go, had not been trained by humans,
rather simply given the rule's games, then left to train itself."
Yes.
That's reinforcement learning at its best.
"Is that a type of coding that is called AI as opposed to machine learning?
How is it similar or different in terms of the coding?
Thanks."
So yeah, fantastic question, and something that
is one of the greater confusions with how AI and machine learning
are related.
A lot of times, AI utilizes machine learning
practices in order to generate what looks like intelligence.
And the goal is actually to simulate human intelligence.
That is the benchmark, give or take.
And so what TWITCHHELLOWORLD points out there,
which is that the AI, or AlphaGo, that beat a human being at go
was not actually trained by humans and was just handed the rules
and figured it out.
That's really crazy.
It's one of the most powerful versions of machine learning,
or artificial intelligence I think is actually much more close
to what you're actually going at there.
And it's super cool that it was able to do that.
That is a completely different.
COLTON OGDEN: Just played a ton of games against itself, basically--
NICK WONG: Pretty much.
COLTON OGDEN: --and figured out what led to a victory?
NICK WONG: It kind of simulated it in its head and figured it out,
and that's nuts.
That's insane.
Actually, there's this really cool version of that type of learning.
It's similar.
It's the reinforcement unsupervised learning
challenge bu I think Microsoft.
And they use Minecraft for it.
So you actually build a Minecraft agent and have it compete with other agents
to survive.
And that's something that is also on my list of projects for winter break,
is to build my own Minecraft agent and see how I can do on that.
COLTON OGDEN: See if you beat out everyone else.
Battle royale.
NICK WONG: Battle royale.
I'm a huge fan of building agents for online video games
like PUBG, things like that.
Also, we're going to just start dragging cartoon portraits of people--
COLTON OGDEN: Oh, nice.
NICK WONG: --onto our data list.
Now, generally speaking, you want to be really careful
when you're picking up human data because there are a lot of racist AIs
out there.
And that basically just means they were trained
on data sets that weren't complete.
These data sets did not have representations of all kinds of people,
and that's really important.
It's something I actually feel very strongly about--
making sure that people realize that you're just doing bad science if you
don't have a complete data set.
So even in the worst case, where I can't, I guess,
help you from being racist, you should certainly
be able to recognize that it's bad science
to not have a complete data set.
And in the best case, you realize that being racist is not ideal,
and so you are also trying to obtain the real goal of being
representative and inclusive of everyone that could possibly
be using your device.
COLTON OGDEN: Well said.
PUPPYLIONHD says, "Halfway in this CS50 course,
should I be able to understand all of that?
Because I don't, and it's kind of frustrating."
NICK WONG: I would argue certainly not.
I think even if you were working on--
if you were all the way through CS50, I would
argue that you shouldn't necessarily be able to just understand
this right off the bat.
And actually, yeah, I think it should be a little bit frustrating.
It certainly was for me when I began.
And I know for a lot of people it is constantly very frustrating.
So I think what I would advise, or my general advice for that
is, yes, it's very frustrating at the beginning and throughout,
but looking back at what you can do now is a really good way
to alleviate that frustration.
So for me, when I went into CS50, that was my first time I'd ever coded.
That was two years ago.
We're I guess getting on three.
COLTON OGDEN: Are you a junior now?
NICK WONG: I am a junior, yeah.
And so I actually took CS50 as my first ever CS class.
I had never really coded before then.
I had built computers and things, and so there
were some we'll say routes for doing some computer stuff,
but I had never actually written a line of code.
And so I was very lost.
And I give a lot of credit to my TF at the time, actually,
for helping me with how CS could be a big part of my life.
COLTON OGDEN: Who was your TF?
NICK WONG: His name was Vik Jayaram.
COLTON OGDEN: Oh, yes.
NICK WONG: So if you're watching, then--
COLTON OGDEN: Shout out to Vik.
NICK WONG: Yeah, shout out to Vik.
COLTON OGDEN: I know Jik--
Vic personally.
I can't talk.
NICK WONG: He's a great guy.
COLTON OGDEN: Yeah, he's super cool.
NICK WONG: I still remember the sections that I was in with him.
And he also was super willing to be like, yeah it's hard,
but you can do it.
He was very encouraging in that sense.
And I give him a lot of credit for making
me feel like I could actually go through and accomplish a lot in CS.
I've been teaching the class that he taught for two years now.
So that's very cool.
And I was really excited about that.
And I like to tell it to a lot of my students,
is that I started out not that far from where you are now.
I really wasn't all that different.
And I think that that can be really helpful,
especially if you're sitting there and you're like,
wow, this seems very complex or it seems absurd,
it seems basically impossible for me to accomplish.
That is certainly not the case.
And it's not the case that I hope any CS people or any CS staff
ever propagate that opinion.
I think that it is very important for people
to realize that CS can be for anyone that wants to be in it.
I think that it rewards people who have natural talent.
That's true.
Almost any field does.
But it particularly will really reward people for putting in effort.
And a lot of fields are like that too, where
you will actually get a lot of rewards out of it for putting in real effort.
COLTON OGDEN: Yeah, and all these streams, they take a long time,
and it's just a small fraction of the amount of time
that it actually does take to do all of this stuff.
NICK WONG: That's true.
COLTON OGDEN: It's time consuming, but if you
enjoy it and you put in the work, it's very rewarding.
NICK WONG: Exactly.
COLTON OGDEN: And it's fun.
It can be fun.
NICK WONG: I could not agree more.
COLTON OGDEN: "Nick is not racist and loves puns.
Great addition to the channel."
Yes, I very much agree.
NICK WONG: Thank you.
I appreciate it.
COLTON OGDEN: "What is the difference between React and Angular?"
NICK WONG: Oh, that's a good question.
COLTON OGDEN: So unrelated to machine learning, but in a nutshell,
they are both front end frameworks for web development.
They abstract what it means to do for an end development
and make it a bit more modular and are more high level.
Tomorrow, we're actually going be talking about React with Brian Yu--
NICK WONG: Hey, Brian.
COLTON OGDEN: --where we can maybe do a deeper dive into the differences
and nuances between the two.
But React is certainly the more popular presently,
although Angular has been around for longer
and has also had a lot of fame associated with it.
"Is it safe to study the same sort of coding and courses
as machine learning to eventually execute unsupervised learning?"
NICK WONG: So let's see.
Yes and no.
Yes in the sense that you need the basic background and other information--
oh, sorry.
Ignore my desktop, it's a mess.
You do need a lot of the same background and basic information and practices
that you see in this supervised rigorous learning that we're doing right now.
But you also would be exposed to a lot of new techniques
that people realize they could use when they
didn't have to actually watch or give answers to their AI or their machine.
There's also a lot of different best practices in play,
like I'm not really necessarily collecting data.
I might be, as was pointed out earlier, I might be pointing out rules.
I might be looking for patterns that are a little bit broader
than the actual data that we're looking at.
And that can be really important, especially
if you're creating these agents.
Then you give them a set of rules, and then you see what they do.
Or you might give them a set of rules, see what they do for a little bit,
feed them some information as a response to what they did,
and then let them go again.
So there might be interactions.
Actually, I think a lot of the time there
are interactions between this unsupervised learning
and the supervised learning.
So yes, you definitely would want--
I think this is easier to start with.
It's a little bit easier for us to contextualize.
Excuse me.
But you will learn a lot more in that field as well.
There is a lot to that on its own.
I think this is a good place to start, though.
There are a bunch of questions.
COLTON OGDEN: "What is the best technology that fits
for certain and guaranteed hiring?"
Honestly, that's a hard question to answer just because every industry is
a little different.
Every niche in every industry is a little bit different.
I would say figure out what you want to do in computer science
or whatever field you're interested in.
Do a deep dive on it.
Learn a lot.
Spend a lot of time figuring out whether you enjoy it or not.
And make some projects.
Flesh out your GitHub a bit.
Reach out to people.
Get some smaller jobs.
See if their work experience is fulfilling.
And build upon it from there.
But it's hard to frame it as there being any sort of magic bullet or magic trick
because at the end of the day, it's going to be a lot of hard work
ultimately, because if it was easy and free, everybody would do it
and everybody would have a very well-paying job in tech.
NICK WONG: That's true.
COLTON OGDEN: So the reality is you're going to have to spend a lot of time
figuring that out for yourself and doing research and getting
good at various technologies, and ultimately probably finding
a specialization or a given strength in a particular field.
But good question.
NICK WONG: Very, very good question.
COLTON OGDEN: FORSUNLIGHT says, "This is the real power of teaching.
Connection fuels commitment.
I guess daily live streams will increase the commitment of students."
I think so too, and I honestly hope that this channel helps inspire other people
to dig into more projects of their own and we
get to build some really cool stuff on stream.
I know today's has been really cool.
It's been very unique, all the ones we've had so far, and very technical,
which is nice.
The deeper we can go into this stuff, I think it shows a little bit more
of the reality of actually getting into the nuts
and bolts of this kind of stuff.
NICK WONG: Yeah.
I'm super excited that Colton asked me to come and join and be
able to present with all you guys.
But also, I think Colton's idea for just having
a Twitch Livestream, that's fantastic.
I think that's super awesome for what you guys get to do.
And you get to have just new topics all the time.
It's so cool
COLTON OGDEN: I get to come in and not even have to prepare anything.
It's great.
I can mess up live on stream and everyone sees
what it's like to be a real programmer.
How about that?
NICK WONG: Actually, that's a really important part of this.
And Colton told me about it last week when
I was thinking about what to do for this, is that being able to mess
up live on camera, that's great.
I think it shows a lot of people just the ridiculousness of the belief
that people just are monolithic coders, that they just come in
and they just go boom, boom, boom.
Those coders, I guess, do exist, but they aren't necessarily--
they're usually a product of having done it a lot.
They've sat down and really fleshed this out.
So having I guess the live version of that,
going online and having that in front of you, that's really cool,
and it's something that I wish you saw more in other fields too.
COLTON OGDEN: Yeah.
A big trend is making an app in 10 or 15 minutes.
But the reality is it's very unlikely that any actual app or project you make
is going to take 15 minutes.
And those videos are cool for introducing
at least the concepts behind a lot of those ideas things,
like building the games and such.
GHASSEN says, "Thank you, Colton.
Hope to be friends on Facebook."
Yep.
Send me a friend request.
I'll accept it.
"Shout out to Colton.
Livestream is cool.
I like to see people get frustrated and break code
like I do in psets, practice, et cetera."
Yeah, that's going to happen.
NICK WONG: It certainly happens.
COLTON OGDEN: That's absolutely going to happen.
BHAVIK, thank you for that.
NICK WONG: All right.
So as promised, a little bit behind schedule, but definitely as promised,
I've edited a little bit of our code.
I added a command line argument so that we could modify things as we go.
But we have just changed the entirety of what we're doing.
We're no longer classifying between light and dark.
We are classifying between real and cartoon pictures of people.
And all we had to do was find new data sets and label them.
That was it.
And the code will work entirely fine on those.
You'll notice things are a little bit harder here,
where we still are validating things--
oh, nice.
David through your Facebook in there.
COLTON OGDEN: Shout out to David for throwing my Facebook in the chat there.
"Prof is here!"
Welcome, David sir.
NICK WONG: Yep.
Welcome back, David.
And so this validation accuracy is not necessarily
going straight to 100 anymore.
It is a little bit difficult to classify, really,
between cartoon images and real life ones.
But you'll notice we're doing way better than just guessing.
We are not at 50/50.
We're at some 80% and 80%, which means we're roughly--
oops-- not too terrible on this.
Ah, that's the same bug I had last time.
Let me get rid of that and let it run again.
rm img/predict/.DS.
There we are.
So annoying.
I told you at the very beginning of the Livestream that was going to come up,
and then I forgot about it.
COLTON OGDEN: It's like we've come full circle.
NICK WONG: Yeah.
We've come all the way back to what we had at the very beginning,
where there's a small bug, we fixed it, and now we
are doing what we promised at the beginning.
We are classifying between cartoon images of people
and real images of people.
And what that means is--
in concept, you could train this a lot better with way more comprehensive data
and doing all sorts of cool things.
But even with very limited data sets in roughly two minute training times
and in and hour and a half, two hours of coding,
we were able to get to a model that works pretty well
on what's going on here.
There's actually quite a few concepts that we demonstrate in this model.
There's a lot of things going on underneath the hood, which
is one of David's actual favorite phrases, I think.
What's the record?
40-something?
COLTON OGDEN: I think someone cut a video on our production staff
of how many times David said "under the hood," and it was a lot.
NICK WONG: It was pretty funny.
COLTON OGDEN: It was a good number.
A good number of it.
NICK WONG: So yeah, there's all sorts of things going on
in this even simple project.
And I think that there's a lot of places for it to grow and expand.
So you're welcome to take the code.
It's all yours.
There's no need to credit me with it.
And the images are not mine except for the pictures of me.
There is a picture of my face.
I use it as the profile picture.
And you'll notice, even here, we didn't do a great job of predicting.
We guessed that they were both cartoons.
One of them is certainly not a cartoon.
I'm a real person, I swear.
COLTON OGDEN: It's debatable.
NICK WONG: Debatable.
I'm certainly, not a cartoon version of myself, I think.
But yes, we basically were able to get from beginning
to end in about two hours.
And we have a machine learning model that allows you to predict on things.
There's plenty of room to modify this to make it a little bit more general,
to make it usable from the command line or as an app.
But this is certainly a--
I think if you were able to get through this even without fully
understanding everything, even if you were just sitting there and being like,
oh, this is kind of cool, that's fantastic.
COLTON OGDEN: Even I'll admit, I don't know every detail
of what we talked about, but it's been fascinating
and actually inspiring to go and dive a little bit deeper
into these technologies--
NICK WONG: That's awesome.
COLTON OGDEN: --into machine learning.
So I think if that's even what some people take away from this,
they're like things, different fields, different niches of CS or other fields
can be interesting or inspiring you without knowing all the details, that's
great.
That's how seeds are planted for future exploration.
NICK WONG: Yeah.
I thoroughly agree with that.
Couldn't be better said.
COLTON OGDEN: "Thank you very much," says DKBUXBOM.
"It has been instructive and very inspiring.
Special thanks for the encouragement."
Absolutely.
It's been a great session.
NICK WONG: Yeah, of course.
Thank you.
We appreciate having you guys sitting here talking with us.
It's really cool.
COLTON OGDEN: Yeah, the back and forth is what really makes this a ton of fun.
42FORCE-- "Many thanks for the effort Colton and Nick along with rest
of CS50 team!
This is something to be thankful for by being here in the 20th century.
Education is now a free resource."
Absolutely.
No, and it changes every day.
And a special thanks to our production team for setting this up for us too.
This is only possible because of their hard work and for David--
NICK WONG: Snaps for the production team.
COLTON OGDEN: --being in support of the vision as well.
So shout outs to David and the production team.
Maybe want to read off TWITCH--
NICK WONG: Sure, yeah.
So TWITCHHELLOWORLD.
Nice.
You went for two smaller ones, and then--
COLTON OGDEN: Giving you.
NICK WONG: I appreciate it.
No, it's cool.
So TWITCHHELLOWORLD says, "Is the distinction between, say,
a PhD in AI for instance, those who develop the new techniques
and.or develop TensorFlow itself as opposed to the MS/BS degree holders
and self-trained as those that execute these techniques in various settings?
I hear sometimes PhDs are overqualified for some openings,
though it seems some companies prefer advanced degrees for their AI research.
Thank you so much for this."
So of course, you're welcome.
We appreciate it.
But also, that's a great question on the difference between PhD MS versus BS.
And actually, a lot of that can vary by country.
I know that in certain countries it's actually really common to get a BS,
then an MS, then a PhD.
In the United States, a lot of times you jump from BS to PhD.
There's maybe some years of research in between,
but you actually just go straight from one to the other.
And so that's a really good question for asking, maybe,
what are companies looking for?
And a lot of companies, especially the big five, you're looking at them
and you might say, OK, well I might want to go work
for like DeepMind, Google DeepMind.
They tend to look for a lot of PhDs.
I don't know of too many people who are undergrads
who are going to work on DeepMind.
That's very cool.
It's super technical work.
And they tend to do a lot more theory before they put it into practice.
But I think differentiating between people who are purely theoretical
and people who are purely practical is kind of a false dichotomy.
There are people who are very theoretical who can do
a lot of really cool practical stuff.
They can build apps out of nowhere.
And there are people who are super practical who
also happen to know a lot of theory.
And so I think that they go hand-in-hand.
The more theory you know, I think it makes it a little bit easier
to do things practically.
There are some classic counterexamples to this,
but I think in general, that's true.
And the more practical things you do, it motivates a lot of the theory.
Why am I curious about how this actually works underneath the hood
until I've done it?
And then I see where things break.
Now, maybe when I understand how things work,
I might say, oh, that makes a lot of sense.
I can actually fix this now.
And so it influences both directions.
COLTON OGDEN: We can answer questions like, what's a sigmoid?
NICK WONG: Right.
Right, exactly.
COLTON OGDEN: "What if we live in a cartoon?"
says MAGGUS.
NICK WONG: That would be really cool.
Actually, there is a friend of mine who posed a really cool thought
experiment on that, which was-- could you ever
believe that we would be living in a matrix?
I'm going to just absolutely bungle this.
But he said it really eloquently, which was
like, if you could believe that we could ever be a matrix,
then it's probably pretty possible that we're
in one, because if you think that AI could ever be representative of a human
being, then the fact that that could already exist
would make us exactly where we are now.
It would be on not identifiable.
And I think there's a very famous discussion or set
of quotations between Elon Musk and Stephen
Hawking and talking about the probability that we live inside
of a simulation.
And they say that it's really high.
The probability that we don't live in a simulation is very low.
And based on what we know, that's true, at least
based on that topic and my readings of the articles on that.
However, a counter to that example is, let's say
that I have a soccer team A and a soccer team B,
and I say, what's the probability that soccer team
A wins when they play together?
And you would pretty reasonably, hopefully, say it's about 50%.
And that's true.
That's the current probability given what you know.
However, if I tell you that soccer team A is Harvard's varsity boys soccer team
and soccer team B is my kindergarten soccer team,
you now know that the probability that soccer team A wins is roughly 100.
I would argue actually 100.
I hope it's 100.
So that changes--
COLTON OGDEN: Shout out to the Harvard soccer team.
NICK WONG: Yeah, go Harvard.
So that's a crazy shift in what our probability is.
It went from something that is literally probabilistic to something
that is quite hopefully deterministic.
And that, I think, points to a lot of these questions about--
are we a cartoon?
Are we in a matrix?
Are we in these sort of things?
There are people who are very smart and very
relevant to this sort of discussion, and they have a lot of cool opinions on it,
but it's good to also temper those opinions
with the reality of the probability that we know very much
about our world and our universe even.
Pretty minimal.
We don't even really know what's in the oceans.
So to say that we might understand something about the universe
is pretty limited to what we can actually
say given that the amount of information we know, the total breadth of knowledge
of humanity-- pretty small if you imagine that the breadth of knowledge
that actually exists is infinitely larger, or almost infinitely larger.
COLTON OGDEN: We're into metaphysical conversations here on--
NICK WONG: Yeah, very metaphysical, very philosophical.
COLTON OGDEN: --CS50.
Twitch.TV/CS50TV.
Let's see, make sure we didn't miss anything.
NICK WONG: Right.
COLTON OGDEN: "Colton, Nick, you're amazing.
Thank you for the great work you're doing."
NICK WONG: Thank you very much.
COLTON OGDEN: Thanks so much, BELLA.
Appreciate you coming in time and time again.
I think I've seen you on every stream so far, so thanks so much.
NICK WONG: Wow, that's awesome.
COLTON OGDEN: "Thanks for the live stream," says BHAVIK.
"I really enjoy being here."
Glad to have you, BHAVIK.
Also a regular.
David's heading out.
Thanks, David--
NICK WONG: Thank you, David.
COLTON OGDEN: --for joing us on the chat.
We got some thank yous for David there in the chat.
"Which IDE are you using?" says MARLINDO71.
NICK WONG: Visual Studio Code.
And I think BHAVIK_KNIGHT actually answered that right below.
COLTON OGDEN: Oh yes.
NICK WONG: But yeah, I'm using VSC.
It's one of my favorite IDEs.
I also sometimes switch to Adam if I'm feeling kind of fun.
But yeah, whatever works for you.
COLTON OGDEN: "How to choose the best partition
after preparing a Gini or Shannon tree?
How to identify the prediction exceptions?"
I can't say I know.
NICK WONG: OK.
So I don't know super well what exactly Gini and Shannon
trees are, though I would imagine that it's related to forest decision
problems in machine learning.
And I can explain a little bit of that.
It's basically you pass a set of data into a set of-- they're
called trees that make decisions.
And at first, they're random, and then they
start to really successfully choose based on decisions.
And this goes iteratively all the way through.
That's a very rough, leaving out a lot [INAUDIBLE]..
COLTON OGDEN: It's like a decision tree pipeline almost?
NICK WONG: Yeah.
I would think of it like that, and there's some nuance
that I'm pretty much omitting.
And I believe if you're talking about something along those lines,
then it's very similar in the decision to choose a rectified linear unit
versus a sigmoid versus different activations
then what we saw here in our linear models.
And that's roughly what you end up seeing.
But because I don't actually know what a Gini or Shannon tree is or are,
I can't necessarily speak to how you would actually identify the prediction
exceptions on those.
COLTON OGDEN: Cool.
Cool.
"Thank you," says TWITCHHELLOWORLD.
Astley says, "If we lived in a cartoon, we could run in the air
until we looked down."
NICK WONG: Possibly, yeah.
COLTON OGDEN: That's one of my biggest desires right there.
"VSCode.
If you are looking for an IDE to do Python stuff, PyCharm is very good.
I use PyCharm for Python."
I haven't used it myself--
NICK WONG: [INAUDIBLE].
COLTON OGDEN: --but I've heard the same thing, yeah.
People tend to really enjoy it.
Patrick Schmidt tends to really like that a lot.
NICK WONG: Oh, that's awesome.
COLTON OGDEN: "Love Professor Margot Seltzer's phrase, if you do good,
you will do well."
Yeah.
Do good, do well.
That's one phrase that actually definitely says.
It's very true, very true.
"Also relevant."
We have a Reddit post here.
I'm going to pull it up on my computer.
NICK WONG: Yeah, I was going to say, that's a dangerous--
COLTON OGDEN: DRUNKBEAR-- "What is happening here?"
OK, the number one data--
maybe if you want to pull it up, it looks OK.
NICK WONG: Oh, all right.
COLTON OGDEN: Pull it up on your web browser there.
NICK WONG: Let's see.
Twitch.tv.
COLTON OGDEN: Oh, actually-- oh, yeah, yeah.
You have to go to the chat, I guess.
NICK WONG: CS50TV.
COLTON OGDEN: Yeah.
And then I'll paste it in there again when you're in the chat.
NICK WONG: Beautiful.
Gotta love the Wi-Fi.
Anytime you are actively observing the Wi-Fi, it's millions of times slower.
But then if I look away, it's gone.
Look at that.
Cool.
Oh, this is going to be pretty meta.
You're going to be watching the stream.
Oh, well you get an ad.
I was going to say, you're going to be watching the stream--
COLTON OGDEN: Watching the stream within a stream kind of.
NICK WONG: --while it's streaming.
Yeah, kind of.
COLTON OGDEN: Kind of.
I think the chat is--
I think you have to expand that a little bit.
NICK WONG: I'm going to expand this a little bit.
There we go.
COLTON OGDEN: And there.
You should be there.
So let's paste that bad boy right in there.
NICK WONG: Beautiful.
All right.
So we get this beautiful gem.
It almost looks like an XKCD comic.
All right.
So "The number one data scientist excuse for legitimately slacking off."
COLTON OGDEN: Yeah.
That's like the old "my code's compiling" joke from back when--
NICK WONG: That's my favorite.
COLTON OGDEN: --back when programs took 30 minutes to an hour.
NICK WONG: My build.
COLTON OGDEN: Some probably still do take that long.
NICK WONG: I believe so.
COLTON OGDEN: Probably much longer than that.
Things like-- I don't know about Chrome, but maybe
Microsoft products like Word and Excel.
NICK WONG: Yeah.
They're known for taking forever.
COLTON OGDEN: Those probably take ages, right?
"What is happening here?"
DRUNKBEAR.
We are just recapping the conclusion to our machine learning stream here
with Nick Wong, but it's going to be on YouTube.
The VOD should be up on Twitch.
We built a simple binary classifier.
And then I guess this is a perfect way to close it off too with having the--
NICK WONG: Yeah, I figure we can loop all the way back around.
COLTON OGDEN: --the visual of Nick's custom little shell screensaver here,
which is really cool.
I don't know if maybe you want to link to how you got that if it's simple.
NICK WONG: Yeah.
Actually, I can link to what that is.
COLTON OGDEN: Because that's really cool.
NICK WONG: I can post that here.
Well, let me log into Twitch.
COLTON OGDEN: "Yeah, you're genius guys.
Gini and Shannons are algorithms to identify the best fitting model
to adopt so you can build a classifier."
Yeah it sounds like what you were saying.
NICK WONG: OK, I can't log in.
I can tell you what it is.
It's CMatrix is the program that actually runs.
I piped CMatrix into lolcat.
So I can actually show you what that looks like.
COLTON OGDEN: "Atom is a bit slow to start.
I like Vim's key binding very much, so I use it inside the IDE," says BHAVIK.
Yea, Atom is a bit slow.
Try VSCode out, see if that's a little bit faster.
I know it uses less memory than Atom does.
But the two are, I believe, both built on Electron,
so they both do use a bit more memory than they need to.
Well, not than they need to, but than is--
compared to other--
NICK WONG: Than standard-- yeah.
COLTON OGDEN: --IDEs or other text editors.
PUPPYLIONHD-- "When is the next stream?"
The next stream is tomorrow, so we'll have CS50's Bryan Yu, our head teaching
fellow, who'll be up here.
NICK WONG: Yeah, Brian Yu.
COLTON OGDEN: Shout out to Brian.
He'll be teaching us how to React, the basics of React.
And his web course is--
I think I pasted it earlier, but just to do it one more time.
He taught that course, a web course, so if you
want to maybe glance over some of the course materials there.
I don't recall offhand if he taught React in that class.
I know that Jordan's class did teach React or React
Native, which part of that was React.
"Recently installed VSCode."
Oh sorry, did you put the thing in the--
NICK WONG: I did.
Sorry.
I put the command up here.
It's this command is technically how to run.
Or I guess this is the actual command.
COLTON OGDEN: OK.
So they need to install a CMatrix and a lolcat, and then once they have that,
they can alias that in their Bash profile,
and then they're all set to go.
NICK WONG: Yeah.
And I like to call it rest because it's what I do when I'm resting
and I'm not actually teaching anything actively.
COLTON OGDEN: Nice.
Love it.
I love it.
It's nice.
NICK WONG: Cool.
COLTON OGDEN: We'll keep that on there for just a second.
NICK WONG: All right.
Sounds good.
COLTON OGDEN: But yeah, this was Nick Wong
building a binary classifier, a humble introduction to machine learning.
I didn't absorb probably more than 50% or 60% of it, but it's very inspiring.
NICK WONG: We went through it very fast.
Thank you, I appreciate it.
COLTON OGDEN: It looks like a deep rabbit hole.
Looks like a--
NICK WONG: Yeah, it just goes.
COLTON OGDEN: --lot of very interesting stuff, a lot of fancy terminology
that I would love to understand that the theoretical meaning of.
But machine learning powers so much of the modern world,
so it's nice to get a glimpse into how it works.
NICK WONG: Thank you for having me, Colton.
I really appreciate it.
It's been awesome.
COLTON OGDEN: Yeah.
No, it was a terrific--
NICK WONG: Looking forward to all of the future talks we get to do.
COLTON OGDEN: Terrific stream.
And join us again next week.
On next Friday, we'll have Nick talking to us about some of this fancy terminal
stuff.
He'll be showing us--
NICK WONG: Terminal tricks.
COLTON OGDEN: --basic Linux commands.
So a bit more of an entry level stream, some how do we use the command line?
What are some common commands?
Things like ls, cd, and maybe some other things like piping,
or however much time we have.
NICK WONG: Yeah.
I think we'll go for a little while.
COLTON OGDEN: Which is a great foundation
upon which we can build other streams and get people more fluent
in using the command line.
NICK WONG: Yeah.
COLTON OGDEN: "What password managers do you guys use?
I keep forgetting to ask that.
David Sir talked about it in cyber security lecture,
but I'd like some recommendations to try."
We use 1Password for CS50, so take a look at 1Password.
I don't know if you use--
NICK WONG: I use LastPass, but all very--
COLTON OGDEN: They're all probably more or less--
NICK WONG: They're pretty much the same, yeah.
COLTON OGDEN: --identical or feature compatible.
But yeah, using a password manager is a very handy tool.
NICK WONG: I just got a student discount for LastPass.
COLTON OGDEN: Oh, nice.
OK, a little bit of a interest there.
All right.
Well, thanks so much everybody.
This has been CS50 on Twitch, our seventh episode, Binary Classifier
with TensorFlow, Keras, and Python.
Thanks for everybody who's tuning in.
As always, if you have suggestions, definitely let us know.
We'll be joined again by Nick next week.
And this week, we'll be joined by Brian tomorrow for some React.
And then on Friday, we'll take a look at some Unity.
We'll make 3D Pong.
NICK WONG: Nice.
That's awesome.
COLTON OGDEN: So yeah, that'll be it.
Any last things you'd like bring up or mention?
NICK WONG: No.
Thank you, guys.
Really appreciate it.
I'll see you guys next Friday.
COLTON OGDEN: Cool.
See you, everybody.
Thanks so much again.
Looking forward to the next stream.
NICK WONG: Awesome.