Autoencoders in Keras | Abhishek Kumar | Skillshare

Playback Speed

  • 0.5x
  • 1x (Normal)
  • 1.25x
  • 1.5x
  • 2x

Watch this class and thousands more

Get unlimited access to every class
Taught by industry leaders & working professionals
Topics include illustration, design, photography, and more

Watch this class and thousands more

Get unlimited access to every class
Taught by industry leaders & working professionals
Topics include illustration, design, photography, and more

Lessons in This Class

7 Lessons (1h 26m)
    • 1. Introduction

    • 2. What are Autoencoders?

    • 3. Simple Autoencoder implementation in Keras

    • 4. Simple Autoencoder - Visualizing Encoded output

    • 5. Deep Autoencoder using Sequential API

    • 6. Deep Autoencoder using Functional API

    • 7. Convolutional Autoencoder in Keras

  • --
  • Beginner level
  • Intermediate level
  • Advanced level
  • All levels

Community Generated

The level is determined by a majority opinion of students who have reviewed this class. The teacher's recommendation is shown until at least 5 student responses are collected.





About This Class

Autoencoders are a very popular neural network architecture in Deep Learning. It consists of 2 parts - Encoder and Decoder. Encoder encodes the data into some smaller dimension, and Decoder tries to reconstruct the input from the encoded lower dimension. The lowest dimension is known as Bottleneck layer. So, it can be used for Data compression.

In this course we explore the different types of Autoencoders, starting from simple to complex models. We'll also look at how to implement different Autoencoder models using Keras, which one of the most popular Deep Learning frameworks.

Meet Your Teacher

Teacher Profile Image

Abhishek Kumar

Computer Scientist at Adobe


Computer Scientist @Adobe

See full profile

Class Ratings

Expectations Met?
  • 0%
  • Yes
  • 0%
  • Somewhat
  • 0%
  • Not really
  • 0%
Reviews Archive

In October 2018, we updated our review system to improve the way we collect feedback. Below are the reviews written before that update.

Why Join Skillshare?

Take award-winning Skillshare Original Classes

Each class has short lessons, hands-on projects

Your membership supports Skillshare teachers

Learn From Anywhere

Take classes on the go with the Skillshare app. Stream or download to watch on the plane, the subway, or wherever you learn best.


2. What are Autoencoders?: elephants in discourses. Study word or two in quarters, and we'll see how to implement them in a very popular deep learning tremble called Farrah's . We will start with the basic quarter in quarters and OpenGL amend it, and then gradually we will move on to more complex or doing quarter architectures, and then we will see some practical or two in quarter mortal examples. So let's suppose to understand what is or Queen Quarter. So are doing Quarter is a data compression technique, so this architecture consists mainly off two parts. The first part is then quarter, and the second part is Dick Order and this middle ear. This compares the compressed representation. And how old does this continue? Compressed strippers? Interesting. Is there If you have entertained that, say, toes in diamonds, then given have different layers in neural network. So the first layer then accept these one present inputs so we can have 1000 more here just learning to input and then in the next little it Sylvia 800 north and these are connected, and then we have lets it 400 nodes and off in supervised learning. The seed input to the network and also don't level or so the network, with help off these many or notes in between multiple layers, tries to figure out on the complex function on DE so that it tends represent these values. So, in this case, in case of for two orphan quarters, there are some layers. Then this is the middle ear, and then this dick, or to the very sole here, what it has been the same is door. So it's a supervised learning setting. You can see we are doing the input. We're not explicitly giving or put the importance. A minute or so the job is off. This network is to recreate. Or but the constraint is that we have kept Let's it 1000 North's here by 800 norms here and 400 notes here and then again, 800 then potion So that thing could over Damon's in matches. So the networks job is to minimize the loss function on our loss function will be comparing the input and the generated or so this difference would be minimum, and the jaw off any network is to minimize their loss. So but in the middle, we are just full damage. 400 dimensional value. So in essence, various people worried many examples than the network will learn how to recreated under our input and you sleep the are rolled off this live, then it will ah kind of represent this important later because initially, the the network of will exam features off the input and discard some features of doing in court. Ah, but ultimately, by minimizing these lost functions were being put to, say, menarche, it will pick the most important things. Presenting this input, that is, it will capture the main content of them and ignore some very personal contents like noise or some less important information. So if the after it sufficiently trained and the loss function as Sophie's empirical world, then we can ignore this dick order part and even take this in quarter part. So what this in quarter will do the never be given great off 100 dem insulin It will return and guitar 400 eminence and visually sincerely consist off almost time in formation. Some information will definitely be most, but it really sincerely campaign most of the information and let's say we want to transmit the later then we can transmit this just 400 is later and somebody else receives it over. A network can use this train recorder, and this dick order will take 400 dimensional input and it won't return. It goes in demons'll or so that person can use this 400 demonstrate, feed it to this recorder and get record 1000 dimensional data. So you see that we have We are representing her 1000 diamonds. No detained with 400 still a little. So that's where it can be used as a data compression technique. So what can we do? Simplest or going quarter. So we will have one fully connected live in neural network. And then let's say it does those in dimension and then in the next layer, we have alerts 200 dimension. So this is the second left and in the third earlier were again 1000 Germans. So this is the simplest ordering quarter there we can think off and not necessarily simples in terms of dimension, but the number off layers. So when and we also call this layer bottleneck near because you see an important if you have a border like this, this nick is the slimmest part of the bottom. So in this network, this layer is the slimmest art. So that's where it's called Mortal Nicolle. So this is a simple, uh, ordering quarter. Then we gonna have oh, many more layers, like at a certain a little sample 1009 100. You tendered gradually up to, let's say, 100 and then deconstruct Barre quarter. So this is slightly more complex, but still using fully connected feed forward Network began for their this in convolution network. So we have in poor payments that's in this case, the input Hillary and image. Then we will have the initial convolution earlier. Ah, and we represented like this. I'd return the number of channels so we will have a few conversational layers on some, then some pulling layers, which will sink it to a smaller conversely earlier and keep drinking. So this one reporter Nick and this will this part will form the in quarter part and then leaving reversed. The architecture's and required are protein. It's from the same protein is so that will be or two in quarter in Khan Volitional neural network setting. No, let's see some off the mortal examples off between quarter one that is simple or two in quarter based on fully connect earlier, been sparse or two in quarter, then be fully connected. So just deep, deeper version of this 1st 1 So here this is the simplest neural network with or three layers. Then this is the be perversion on their deep means and terrorists deft in this network because of multiple layers and then become volitional order in quarter. So this next example that I give a kept representing the image in a small in a smaller room is, um and then tried to recreate that team is. And then you see that business network lost the sense off the letter. The men come content on the letter and diagnosed in significant part or very essence return less important, recreating the so quite naturally it will ignore the noise. And then we will see sequence to sequence or two, including for the reduced. And then we will see various, milord, when quarters. So in the next video, we will see hard to implement a simple order in quarter just three layers using tear us. So see you in the next video. Thank you. 3. Simple Autoencoder implementation in Keras: a difference in this real see, a simplest influences, um, off a simple or two in quarter. So if he remember or this slate from the last will you We talked the world Not the simplest or two in quarter can be that you turn just one in place which will take minuto on this foolery equal to their dimension of the input data and then immortal nuclear off smaller sage. And then in the quarter, which will gender door put off the same Sager's in police in pretty Stojan Demonstrate holding Cross one and effective tajin cross from And this mortal nuclear is 200 Cross win andan this or put will 3000 cross one director So this mortal nuclear will represent the representation off this important later in the lower dimension that is in compressed form and in this redoing particular with Hillary Ah, using this mist hand written notice it. So this is those one of the simplest it a certain it is available in almost all the deep learning frameworks on the demons and off 18 years in this feminist tender, tender atrocities 28 crossed frontiers But we will not use the con volitional layers in this case because we're currently trying to build a simplest possible order in quarter. So we will use simple neural network. So if this is images 28 cross frontier been door Perea 20 it across 20 It excels in the Siemens on that is seven it before. So we have 74 pixels in the image and each pixel will take a value from 0 to 2 55 So we since Illyria victor off 74 Elements 74 cross one So we can build a network in which we will have 7 84 neural network norms in the first layer. And then we will love and Inter immediately off 32 notes. Let's And finally we will have 7 84 works in the and we can re separate to create our pro team. It's so we will take them put emit recent and converted to 74 across one Richmond in turn, weaken water. Tow this Lord immense and then back to 74 rich people can work back to 28 Crossed appeared . So this is one of the books off this sort 1/4. This too is recreated. And this is the compressed representation. So let's start rating the court for this. So I have opened a new Ukraine notebook. So first we will importantly since deliveries. So we will import for real use in pro players. And we will use dense or fully connected lier in which every note in the next 11 next layer is connected to all the notes in the previously ER alluded network may learn and discard some of those connections or give a very lower it to them so effectively they will not be connected. Onda. We will import duties. It feminist 100 injured is it? Then we would also need numbly and then Matt broke Libs be like Clark literally and let's run it so it will take some time and imported So it was important deliveries. So no, we will define the according Demons'll 64. Let's see. Then we will read the important moments define import and this is the input layer of the network input and its cities two indeed cross 20 years. So we will recon warding it too. 74 growth one and then the in quartered leaders. It will be a dense layer and we will use the bed with prison and it will play it on the previously that is in for a clue. And then the record dick ordered burials were intensely and fully connected earlier, and it's or produce image important notice 74 And then we will build up or two in quarter more. And in this morning, input reader in cochlear will wither in court e Miss Lou and or put with video in the quarter live. And then we will also create an in quarter modern and your simple decision in 40 minutes on or produce and quartered and we will create a placeholder for and in coordinating. So we will call it in, quartered in court and we will extract the last live from oh Norton quarters. And then we will keep it in the quarter Mortal. And we will know compile the model until we come from ordering quarter a little too, including Morton Norc complain Andre will use optimizer. So we have compiled the modern and no world is in the NATO from the Mr Do does it and returned it Is it so it will give for search Extreme right in That is the training set and the cast warning levels. Then just sit And Chris morning levels will work here in this case, in case off ordering quarters, the important Seamus Dark. So we will not need the why values Lord extreme and then just a placeholder for way value dungan X just and wait. Just so we have loaded these in extreme and waiting and know, what we can do is we can divide the values way to 55 so that it's known rice toe giron one between Giron one and the same thing for extra est. And then we will re separate to 74. So this will have Ah, I think 10,000 to Mrs. So the first the limp off this is 10,000 and then each off Those 10,000 later have done it enough. 20 across 20. So this Lynn Extreme, that is 10,000. So it's 10,000 cross 28 cross tinted. So we can re separate 10,000 cross 10,000. You chose 74 days. So this post is the length which will be preserved. And then we will multi play no other to them in since. And the same thing for just and then Let's also print the values off their sips and let's run it. So the train, sir, turns 60,000 images and the tests a test, 10,000 images and this 2nd 28 Christian treat we have converted to 74. And no, we will fit this NATO in over or two in quarter model. So what we will do will do or two in quarter the name of our modern that is sort 1/4 Lord fit and we will pass the extreme and X test and in this case or in quarter, then produce a mazar. So we'll waas extreme and again extreme because our way train is also extreme and we will run it for 50 approx and then Matt Sides. Let's keep it to 56 and we will so full editor. And for validation, we will use the X test. And again, why justice him as X test in the case of ordering quarters? So alerts running so it will run for 50 of books and you will see the lost funds and decreasing swords after 1st 2 pockets going 34 been after second book. It's 2 56 than to 30 when you know, you see its peak raising. I said, Well drawn for 50 bucks and it will almost converts. So let it be learning. We will write the forecourt. So once it has room, we will do the prediction on that morning, who we will have been quartered images. So first we will in court Dame IDs. And then we will use over recorder so recorded. So this is recorded in Mrs will contain the region. Richard images so alerted complete first men will run it. So I will create another cell just for or plotting purpose. I will not run Bissell until we have this model fit. So let's right the court for plotting. So for how many digits we will display then So it has completed 35 books. It's is still running. - So while plotting, we will reset no or puts to 28 cross 20 it. Okay, so, no, it has completed 50 of folks. And the law says I was induced significantly. So we can run this in quarter and then Dick order on. No, let let me complete the remaining part off floating first. - So we're not so nexus on display the reconstruction and alerts, Run it and let's see what are the values. So you see, the first true is the exact in the important data from the detested. And then it was in corded to lower dimension and finally regard lacto indeed, cross 20 it or port image? You see, there are police Ah, quite similar to input. One that means a bow mortal nuclear. What is sufficiently good represent is uncompressed representation of this impersonator because the decoder was able to recreate nor put from that. But there is some loss, as you can see it slightly blurry long ages. But that's expected. So this organ quarter can be used for compressing the return. Also getting reduced redundancy in the data, for example, noise. So this was the simplest influences in off a simple or two in quarter. In the further reduce, we will see more complex mortal souls disorder in quarters to see in the next video 4. Simple Autoencoder - Visualizing Encoded output: in our previous listen we had seen that are our simplest or two in quarter, which consisted of just three layers like this Ah were referred. Oh, the amnesty 100 gender does it which was 20 odd crossed frontiers. And we made it into one dimension and fed it to our input layer on the intermediate layer was 64 neutrons and then or put away again reconstructed and we be saw that it was It will do generate other data very well. And we had thwarted the input and the corresponding generated values. And we had also used these defined these two extra models in quarter and Dick order. Although we could have used this just order in quarter because we were not plotting the intermediate values. But But we had but first calculated the included value out of the maze. And then using the recorder Morton, we recorded it in this way we will just visualize that what is the value? Intermediate value. So it will be 64 Ah neurons so we can reset it to eight. Cross it and see what is our put. Obviously, this input will be some handwritten or digit lay claim on or quit will be also very close to what in court is. And we have to visualize what is the compressed? Will you? So ah, let's begin. So for our previous work, we could simply comment over these Or, you know, uh, we can just ignore these middle It even commenting Full began coming pen or we can comment or these tools. This is just too so that if we want to just see the airport, we we can safely ignore these and we can work with or doing quarter immortal Let's run it so it should be extinct. This too being put image, we will directly, feet or turn quarter and we'll get there recorded value and no, let's around this ill and you see we'll get to similar. We had just broken it into this in orderto understanding better so that frustrating in court demons and then decoded on. We display their recorded or it's also ah, big Lord, the included value. So no, this will be three rules Onda. This will be two times in so in more Matt shortly we will. So there were plotting three and in eastern three cross 10 out of that. This is the first image. You start from zero. So I plus one is one. So then I will be too. And three and Thorne and similarly here it will be 11 because we're heading. Tend to it, Harry, Like 22. So this But we averted from continuing to 30 teammates. And ah, here we've been Brenda and quartered image. And these are just 68. Cross 64 or mitts, India 74. So we had received 28 Crossed wanted. But we cannot do that for this case. So if you tried to run it, Terry sort of even error. So cannot receive 64 into 28. Cross 28. So we have to recipient. Tow it, cross it. So I think this is a small limits where every image as God compressed. But these are off different dimension. This is it. Cross it. This is 28 cross tribute so we can lord it outside also so that these original images remain as cities and let's try to run it. So it's again. It's treating it to same limited. Let's leave it off. In fact, you can see this is the compressed image. Even you can come the pixel. This is 12345678 Similarly, eight in the sec so you can see all the 64 pictures. So this ah order in quarter. Ah compresses. The seam is 28. Cross 20 it into this represent ism which yet close it on from this It regenerates this image and you can see that the results, Jeremy. Amazing. And from this image, it's able to generate bagged original image. So you know, you see Ah, sincerely. It's representing the same. Oh, you made in this a small form. 5. Deep Autoencoder using Sequential API: reference will see very simple ordering quarter where we will huge a sequence, ills or model off Cara's on Ben. We've learned a few layers, dense layers only fully connected. And then we will build an orphan quarter, and our aim would be to recreate, Although original, input off and missed 100 gender tests, a Z images, ornaments and 20 across 20 years so purple we have 74 pixels. So first we will one roll these 28 cross 20 it and feed it to input clear. So we have 74 excessive, which are fed to input player, also having won 28 neurons. Then do you are further connected? Toe Next, live with 64 oh neurons. On day, we will start recording it back to 1 20 It andan seven it before. So this partisan recorder on this party's in quarter and we will try to compare the input and output so we can re separate back to 28 cross 20 it and then we will plot both low input as a waiter or put and see that it's this network is ableto a generating plant back or not, And this layer, as you remember we will call it more of a nuclear as we had seen in the previous reviews. And we will compare the input in our proxy. Let's begin an hour just to notebook. So first we will import the required level res. So we will need to import, see conceal model. We will just be using C conceal model of cross And now we will import Onley one type of layer called and slave that is fully connected. Lee, we will not use any other layers. Later, In this real usage, we will see Ah, we can use conversely small ears on other pulling layers, another to make more complex modern. But in this case, we will stick to just simple sequence immortal having fully connected layers. Then we need number and we need the innocently to Sirte. It's and finally legal need. Matt Lord lives even plot library for porting and there No, we will define over morning So we will call it Morgan and we never see conceal mortem and then they will air the issue layers to it. So one continued sort of a 1,000,000 could oh has 7 84 pixels. So the first layer it will from then 74 converted to 1 28 Neurons on activation will use Jialu and we need to pass in poor governance and for the first kiss. So this 74 comes from Ah, this 28 cross 28 emits. So we unroll it in tow. A single Richter, a number dance live No, This we have in quartered. No, We will start recording who on compute and then 74 and finally really using more tourism. So now we have defined in quarterly years. So these Denard in quarter on this denotes decoding and no, we will compile over morning. So before trying to figure data, we need toe combined the modern mind recourse in tropical lows and no well lord And it is it So we're not concerned with the way levels because in the case off in quarter, why eat same as X in pretty similar. So we are ignoring those. We just load the training just X values not wavering. We'll ignore those a Mr Lord Nitta Andan. We will normalize these x and y values two values between zero and one. So we need to divide way 2 55 and the same for X test andan. We will re separate toe or single Levinson. So first parameter is the number off data sets. So in this case, Strangelove. Oh, I think 60,000 deserts and just we looked impose into deserts, so we will just pick it from Lynn Len Function. So this is the first perimeter and then, markedly, all little meaning Levinson's Yeah, just excluding the festoon and still land. So whatever is the diamonds and multiply that and receive it. Same for the just and no, we have re subject. So we'll start fitting our model toe. These data fool. So we'll see the data using more than looks extreme and again extreme. We go to avoid cinematics and let's run it for who do your books and let's keep them. That's a as 2 56 earn. Let's run it. So it had started fitting the data, and it will run for 50 books by the time it completes 50 books. Let's write gold for or proved images so we'll store all oh, our courts in old images and we will come very three x test, so art images will be off once the model is trained we will use the predict function to predict for new values. So we're this morning is fitting on the training set. We will do the predictions for the pistes it. So we will pass X just so this will do the prediction so we can run it only after it has completed training. So looks right for therefore, our lord anything. So we report for 10 values on defined the sage off the fear. - So first, we will put a little just a lowly Mrs from Just did a sit in first rule and in the second room, uh, or put or put images. - So this Liston, as converted this will move it, explain to this one or pretty easier. And this through as 10 or puts than the corresponding or put will be below that for second . Same thing. Liston. So that's a plus in here and here. It should be this old images. - So it's running for a second, a poke, and it's almost complete. So no, it's in the last two book and it has completed training, nor alerts predict the values. Oh, - 321 or so. I have wrongly ordered this one. And now it has predicted the values and started out images. No, we will compare the test and or damage it. So this is the best images on Deezer the generator demanded by the order in quarter so you can see it's able to are generating mrs. Although we are compressed it toe more dental times from 74 to 64 in the border nuclear. But still it has. The network has learned. I don't mean oh core components off the network. That's wasteful. It was able to generate it back. So this was a very simple implementation off ordering quarter using is he conceal model and just Bensel ears? 6. Deep Autoencoder using Functional API: In our previous video, we saw a very Celo Oh, are doing quarter where we are just one in pro players. This was 74 norms in our case because we were working with Amnesty 100 Inditex, it reaches 28 cross 2030 minutes, and if you multiply 28 cross 20 it you will get 74. So we have 74 excels in this and this is a great skill single channel limits. So we unruly to 74 cross one. So this form number input layer which we strung to 64. So this was ever a bottle in a clear And then we decoded it back to the original size and that it's 7 84 So in this case, we will go for the So this this was 28 cross country. If you don't understand what we're talking or and this was dozens and Richard 28 cross 20 it on this middle ear, we kept it a 64 cross one on this. In the previous example, it was 7 84 close one, and we're just single area you can see. So it's a very celo network and draft the simplest or doing quarter where we're just single little the least Wharton unclear. And we dick ordered to backdoors no sides. So in this case, we will aired a few more years. So our first layer being quickly relive against its 74 pixels. And then we will sink it toe No. 56 and these were reconnected among themselves. And then we will sink it further to want to indeed and then finally 64. So obviously these are not drawn to skill. And then this will for mother in quarter part and then leaving be called it further toe again. 1 28 been to 56 and then finally seven it before Lord no. And they will re separate back toe 28 crossed frontiers. So essentially, this is the border nuclear, and the network will try to capture the most important information within this 64 on neurons Should information that was earlier containing 74 neurons is now imbedded in this small bottling Clear and then this dick order, because it back toe original sage and other lost function is such that we compared the input and are and retargeted to be as close as possible so that the network learns to embed the most important information Very funerals for considerate accomplishing techniques. Let's begin. So this word over earlier nor book so this will remain the same Onley thing is that we will add a few layers here. So really, we're just one in quartered full. Let's make good 2 56 then one continued and foreign released 64 And the import of this will be the previous in quarterly. So whatever they are out of this and then again in quartered. So no were recorded it to 64 Ah neurons starting from 7 84 No, we need to be called. So we've learned a few layers for recording it and this is over mortal nuclear. No, Here it will be one dont it seamers this one and then to 56 then finally 74 and then put off this layer is the corporate of previous that is in quartered and it will a play this on previous or their distant recorded and again dick ordered. And these will be Luke terrorism's. We will place ignored Onley in there and no, let's run it. So there is some mirrors for it says are imported. Zero in comfortable. It took scripts. Will 2 56 Gordon 1 64 on the Arisen This lane dick order. So earlier in our last video of your the cellar network, where we're just one layer of recording. So with rigid minus one, we re stare the border nuclear, which was 64 neurons? No, we have a few more off does. So it will do minus one. We reached to statistics, so it expects to up to six. So we need to do my industry and then feed it to this minus two, then minus one. So we need to modify ever cool ribbit. So we will have a few of these and let's make it my ministry than minus two. And here we can put this input and recorded andan here. No, we should not require the soon. No, let's turn it so it seems fine extra in it for the load the data sect andan fridge it so it will run for 50 approx and it will take more time than in our previous case because our Lear beard a cellar network. So I will deposit and resumed then the Sinuses for three books. So now it's in the last tip book and the errors seem toe become words you. No, you test completed 50 books and this is over. Or put it over, deporting corner and again. Ah, there seems to be some blurriness around ages, and that is expected because we're compressing the indictable around 74 to 64. So that's roughly more than 12 times. But that is the sense of 14 quarter hope You enjoyed it. In the further reduce, we will see more variance, self or train quarters, so seeing the next really 7. Convolutional Autoencoder in Keras: reference to review will see were gone volitional or pain quarters? So decision. Very kroosee. Listen, because the concept off or two in quarters is very widely used in quite popular in with images were we tried toe or extract the eastern sort of the media and also get rid off the redundant noise. So it's very popular with conversational your in network in order in quarters. So just toe recap. We're working with numbness tender, tender, toxic and we will be defining an in quarter and 1/4 of each other the two components in an order in quarter. And then we will generate a new image. And the main thing is that this middle layer has the least number off North's in the so in a sense, we're representing this whole input data into a very few nodes or in a compressed form so earlier were defined just simple neural networks of it. Some lords. So in this case, we will apply to conversational setting. So the first little bill be input clear or size. Don't be gross to indeed cross one, because over in miss tender, tender desert sorrows, 28 crossed indeed. Ah, great skill images so dizzy over input, clich? And then we will learn a convergence Manlier Cornwallis None. When we will urge 16 Tamil street and we will keep the parenting same and we can use some three cross tree senator. So yes, we are doing the same parting. Then we can have 28 cross 20 it, but the number of channels will be 16. So this is our next live, so we can say corner one Onda we cannot play a pulling on top of it because first, we will be shrinking in the in quarter part and then we will record it. So right now we're constructing the oh in quarter parts of disease 20 it grossed 20 across 16 Then we have played pulling to cross to straight to in our step place to do it will shrink to 14 So no, we have 14 gross 14 a number of channels His same 16 and we will be applying Max fooling And now we have 14 grows 14 cross 16 So you came from 28 Cross frontier to 14 cross for being. Then we will play for the convolution on proper fit 100 It filters so we can shrink it toe for being gross for being gross it. And then we will play pulling against him to cross to Max fooling and our reduced to seven cross seven. So this is not drawn to scale. It's a very much smaller. So we have this seven cross on grocery it volume, and then we're gonna play for the max pulling and certain get further 24 growth for gross it and no, begin. They're the corn layer on top of the same parting. So we really again left with four cross for gross it. And no, we will treat descends over in quarter. Well, this is over compressed representation, or we also call it important nuclear because this is the smallest demonstrating over network. So no, we're done with the according for Then we can up skillet and tried to decode whatever is captured in this this email Ivorian prettier and we sink it so this layer and are probably the same its image for reducing the loss function. Andi, With enough training, the network will learn to represent the seam it in this many damn instance. So no even I heard the recording part so by up, I mean apps m bling. So we have apps, Templin, Le'Ron, carers. And we will shortly see it in the implementation. And with apps ambling, we will scale it toe. Oh, this is a present isn't so while ah, playing Max feeling it will be reduced to seven plus one by if it would have been eight cross it crosses then then also told that come to four cross for growth there. But for no, it's it growth it growth. You're so we haven't done abs employing you. Then we will look convolution and it will come back to it. Cross it, grows it And then we can a plea for their up simply and make it 16. Grow 16 Grocery. Now you're seeing that when we're doubling, it will ultimately come to 32. Cross 32. We're care were started with 28. Close turn to you. So the trick is that we were doing same convolution air So we were reading some parenting. So there. So you remain same after cannibalism. But here we will play the lead conversion. There will not bear Jews. So no, we have 16 gross 16 dress a eight and then we can apply EU convolution here three gross three and we will not paired anything that is valid Cannibalism. Then it will sink to foot in gross 14 Gross it because you feel him 16. Ah, all right. To 16 or 16 and redo colonialism. Then we will overlap this tree. Cross three on profit So this will generate one excel. Then we will move it here This filter and religion right next to him. So busy that when we have covered three cells we have generated one in court and very steerforth one. Then we're who picks suit. And when we reached to 16 we will have generated for in often these excels values and we every still and so we will not slate this window further similar Linda Vertical case. So this 16 cross 16 cross it will reduce the 14 cross border and cross it is We did not play any padding through Well, right, Bill? Not right. Same here. Same means repaired. So there in Park City, its image or concept. So now we have come through this 14 cross for being gross It which is similar to this. And now we can do further up simply and reach 28 cross 20 across it and then we cannot play more convincing clear with one filter so that we reach back to this 28 Cross 28 cross of one. So it will be more clear. I think there is too much in this laid, so it will be more clear when we write Implement this layers ourselves. So the first part it same over previous reviews. We will be importing some libraries and in particular you see one additional and to Europe sampling because we will be using abs ambling layer off. Oh, get us and then, ah, in some of the video you may have seen Sequent, Ill. AP is being used to there. If you have all the layers of the network are stacked one after them and you can you see conceal a pH of carers. But it has limited, and although it's very simple, you just need to keep on adding layers here the model and and what a lot earlier. But when you have complexes, structures like when two inputs are coming to your Lord or when one Nord is giving to our computer network is diverting from the do you need more finer control and then you need the functional appears off heroes. So I will be using functional Libya. In this case, all do not require for our case. So that's right. This input Larry's here, and I have not important see conceal abuse. So no, Once we have imported the leverage, let's load the data. Is it? And there again, working with dimness tender, tender desert and the main part off this Ah luxuries this defining the convoy listening or two in court earlier. So just really finding Brooklier, which was 28 close to indeed close one. And we can also, additionally give names to our lives. So if we don't give name, Kira's will automatically give some minimum like in one input. To what? Let's explicitly you. Some named lower layers. So, dissident, put clear, then we will. Oh, you can see here were added on convalescent live with 16 all filters solar. It's exactly do the same thing in this court here, so we will use this convolution booty. It's and we will have 16 features here. Three cross three filters may have been using and we keep parting equal to same. So it repaired appropriate number of Jerusalem. There are probably Samos input and activation is real. And we will again given name to overlook. Let's keep it going one and then return. It has input. The include clear the previously er and then we will do Max swelling. So here, after conversation, we did Max willing to sink it to 14 gross for dangerous 16. - So know we learned a few more off these conversion and pulling. So in the next level Revilla, it filters cardinal. So you will be seen. Everything would be seem activation real ooh, and new gone to on it We applied on X layer and then similarly for pulling. And then Ah, a few more off these for this time again. Eight on three. So, no, we have to find in quarter part and no living a place their dick according part forward in quarters defined No. And no one believes the assembling. So in pulling we were sinking the sizing up something river no one in this age So in the last case, we did not play Ah seem ho convolution because it would notable toe 32 across 32. In this case who we we will skip seem in the last one. And then finally we need and or pro clear. - Let's give a name to it. So no, we have defined, never learn layer and on dribble define the mortem. And in our case, for this convoy listening Lord, 1/4 model in courts Is this in pro players and or produce this or prepare on, we can give a name to We can call it C for conversational or two in quarter. And then we're ready to compile the modern, where we will need to suicide up to my ears and the loss foursome. You can drive it different optimizers and lost from since. Or you can even define your own loss function. Look, soldier prender, some Rio model so we can see what slayers are having work differences and no, let's run it. Oh, in the corporate clearly held in some stick. So now our model is compiled successfully, and with this mortal lot summary, you can print the structure of the model. So this is the input layer. Don't eat cross to get cross when and this and understands formats age When viewed, it's not known and heard the moment with envy or freaked. Actually, eternity, they re satisfied. The matter is so this demons'll and will represent the battery. None. So this is 20 cross 28 Cross one day and 28. Question did cross 16. And these names are coming from who are doing the names of the order given and then sitting to 14 74 And this is over immortal nuclear or the compressed representation. And then we started recording it. So level it eight than 16. Then we play valued coalition toe, make it 14 then double it example 28 Christian did cross it and finally do indeed cross 20 across one. We just say marry our input. Demons'll So, Noah, we will, uh, their data. So we have already loaded the administrators it and in the case, off ordering quarters. Why eats image X for we us extreme place and hear whispers filed. Nomad sage from bad sides denotes after home any ah buses through the different Ditto. It will have no in the weeds. So even if we don't want, the matters were before list 32. So you can skip that and we will run it for 50 folks full of seven it so it will take some time. Oh, I will get back when they complete 50 books. So no, it's in the last epoch and it has no completed training. So we will do prediction on some new data and we will store it in old images who were done with the prediction. And no, we'll just floored their data to see ah ho correctly it has generated. So you can see your this is the generate 30 minutes and you see it's should be close to the original one. So the organ quarter was able to learn a compressed representation of data in its mortal nuclear. So this what very concealed Listen. And you understand it completely because it will be helpful in or starting further the various non ordering quarters in the high end value in sof orphan quarters. Thank you.