Update to the deep learning and Python with tensorflow as well as now chaos tutorial

it's been a bit over two years since I did just a basic deep learning video in Python and

Since then a lot has changed. It's now much simpler to both like get into it

But then also just to work with deep learning models

So if you want to get into the more nitty gritty details in the lower-level

Tensorflow code you can still check out the older video

But if you're just trying to get started with deep learning that's not necessary anymore because we have these nice high-level

api's like chaos that sit on top of tensorflow and

Make it super super simple. So anybody can follow along if you don't know anything about deep learning that's totally fine

We're going to do a quick run-through of neural networks. Also, you're gonna want Python

3.6 at least as of the release of this video hopefully very very soon

Tensorflow will be supported on three seven and later versions of Python just happens to be the case right now

it isn't I think it's something to do with the

async

Changes, I'm not really sure anyways

Let's get into it starting with an overview of how neural networks just work

Alright to begin

we need to have some sort of balance between treating neural networks like a total black box that we just don't understand at all and

Understanding every single detail to them. So I'm gonna show you guys what I think is just the kind of bare essential to understanding

What's going on? So a neural network is going to consist of the following things. Like what's the goal of any machine learning model?

Well, you've got some input

So let's say X 1 X 2 X 3 and you're just trying to map those

inputs to some sort of output

Let's say that output is determining whether something is a dog or that something is a cat

So the output is going to be two neurons in this case. So it's just boom two neurons

Now what we want to do is is figure out how are we going to map to those things?

We could use a single hidden layer. Let's say we're going to do some neurons here and

That's our first

Hidden lair now

What's gonna happen is each of these X 1 X 2 and X 3 these are gonna map to that hidden lair

each of the

input data x' gets

Connected to each of the neurons in that first hidden layer. And each of those connections has its own

Unique weight now from here that first hidden layer could then map and connect to that output layer

the problem is if you did this the relationship between x1 and dog or cat and

All the other ones those relationships would only be linear relationships

so if we're looking to map and nonlinear relationships

Which is probably going to be the case in a complex question. You need to have two or more one

Hidden layer means you just have a neural network two or more hidden layers means you have a quote-unquote deep neural network

So we'll add one more layer and then we're gonna fully connect that one two

And then once that's fully connected again all unique weights, each of those blue lines has a unique weight associated with it

and then that gets mapped to

The output and again each blue line has a unique weight associated with it

so now what we're gonna do is talk about what's happening on an

individual

Neuron level. So again that neuron has certain inputs coming to it

It might be the input layer X values or it could be inputs coming from the other neurons

So we're gonna again we're gonna call the inputs x1 x2 and x3

But just keep in mind this could actually not be it might not be your input data

It might be data coming from another neuron

But regardless that data's gonna come in and we're just gonna get the sum of that data

So it's gonna come in and be summed all together

But remember we also have those weights each of the inputs has a unique weight that gets put you know multiplied

Against the input data and then we sum it together finally and this is kind of where the artificial neural network comes into play

we have an activation function and this activation function is kind of meant to

Simulate a neuron actually firing or not

So you can think of the activation function like on a graph, you know?

You got your X and your Y and then a really basic activation function would be like a step or function

So if X is than a certain value boom we step up and we have a value. So let's say here

This is zero here. The value is one

So let's say this is our x-axis 1 2 3

so if X, you know after being all the inputs are multiplied by their weights sum together if that value is

let's say

greater than 3

well, then this activation function returns a 1 but today we tend to use more of a

sigmoid activation function so it's not going to be a 0 or 1 it's going to be some sort of value between

0 and a 1 so instead we might actually return like a point seven nine or something like that

So coming back to this neural network here that we've been drawing

Let's say here on this final output layer. You've got dog and cat

well, this output layer is almost certain to have just a sigmoid activation function and

What's gonna say is maybe dog is a point seven nine and cat is a point two one

these two values are gonna add up to a perfect 1.0 but we're gonna go with whatever the

Largest value is so in this case

The neural network is you could say 79 percent confident that it's a dog 21 percent confidence a cat

We're gonna say we're gonna take the Arg max basically and we're gonna say hmm. We think it's a dog

All right. Now that we're all experts on the concepts of neural networks. Let's go ahead and build one. You're gonna need tensorflow

So do a pip install - - upgrade tensorflow you should be on tensorflow version 1.1 or greater. So

one thing you can do is import tensorflow and then

Actually touch flow as TF and then TF dot version will give you your current version

so mine is

1.10

Now let's go ahead and get started. So the first thing we're going to do is import a data set. We're going to use em

nacelle

of

Data sets with machine learning. It is a dataset that consists of 28 by 28 sized images

So it's like the resolution

images of handwritten

Digits 0 through 9. So it'll be like a 0 1 2 3 and so on and it's a handwritten kind of unique image

so it's actually a

Picture we can graph it

soon enough so you can see it's actually an image and the idea is to feed through the pixel values to the neural network and

Then have the neural network output

Which number it actually thinks that image is

So that's our data set, and now what we want to do is

Unpack that data set to training and testing variables

So this is a far more complex

Operation when it's actually a data set that you're kind of bringing in or that you built or whatever

For the sake of this tutorial we want to use something real basic like M inist

so we're gonna unpack it to X train Y train and

then we're going to do X test Y test and

that's gonna equal m n--

Astana score data, so that's gonna unpack it into there now

Just to show you guys what this is

We're gonna use Matt plot Lib you can pip install or just look at it with me, but we're gonna import matplotlib

Pipe lot as a PLT. And what we're gonna do is peel TM show and we're gonna do X

train

And we'll do the zero width

So one thing we could do just just for the record

Let me just print so we can you can see what we're talking about here. So this is just going to be an array

It'll be a multi-dimensional array which is all a tensor is by the way

So this is this is here's your tensor

right

Okay, so that's the the actual data that we're gonna attempt to pass through our neural network and just to show you if we were

To actually graph it and then do a peel t touch show. It's gonna be the number and you can just excuse the color

It's definitely black and white. It's a single color. It's a binary

So one thing we could say is the color map is equal to P LTCM for color map

Binary Reap lot it and there you go. It's a it's not a color image

So anyways back to our actual code up here

Once we have the data one thing we want to do is is normalize that data

so again, if I print it out, you can see it's data that seems to vary from 0 to

Looks like we have as high as 253. It's 0 to 255 4 pixel data

So what we want to do is scale this data or normalize it but really what we're doing in this normalization is scaling it

So we're going to just redefine X train and X testing but it's gonna be TF caras dot utils dot

Normalize and we're gonna pass X

Train and it'll be on axis 1 and then we're gonna copy

paste and we're gonna do the exact same thing for X test and

All this does let's just run that and then we'll run this again and you can see how the 5 has changed a little bit

looks like I got a little lighter and

Then we come down here and we can see the values here are now

Scaled between 0 and 1 and that just makes it easier for a network to learn we don't have to do this

But at the end of this only probably won't have time

But if you want on, you know, comment those lines out and see how it effects the network. It's it's pretty significant

Ok. So the next thing we're gonna do now is actually build the model

So the model itself is gonna start as TF karosta model's dot and then it's going to be the sequential type of model

There's two types of models

The sequential is your your most common one. It's a feed-forward like the image we drew

So we're gonna use this sequential model and then from here we can use this like model dot add syntax

so the first layer is gonna be the input layer and now right now our images are 28 by 28 in this like

Multi-dimensional array we don't want that

We want them to be just like flat if we were doing like a confident or something like that

We might not want it to be flat

but in this case

we definitely want to flatten it so we could use that we could use like numpy and reshape or

We can actually use one of the layers that's built into chaos, which is flattened. So

So we're gonna do ad and what we're gonna add is TF. Chaos layers dot flatten

so one of the reasons why you you want this to actually be a layer type is like when you have a

Convolutional neural network a lot of times at the end of the convolutional neural network. There'll be just like a densely connected

Layer, and so you need to flatten it before that layer. So it's it's it's used for more than then the input layer

We're just use it for the input layer

Just to make our lives easier. So once we've got that

That's our input layer. Now. We want to do our hidden layers again

We're going to go with I think just two hidden layers. This isn't a complex problem to solve

So again, we're going to use the model set up model dot add syntax and we're gonna add and in fact

I think what I'm gonna do is copy paste and then rather than a flattened layer it's a dense layer in the dense layer

We're gonna pass a couple parameters here. So the first one is gonna be how many units in the layer. So we're gonna use

128 units or 128 neurons in the layer, then we're gonna pass the activation function

This is like the function. Like I said like a stepper function or a sigmoid function

What is gonna make that neuron fire or sort of fire whatever so we're gonna use TF tenon