Variational Autoencoders Explained | What are Autoencoders | Autoencoders Tutorial | Simplilearn

Key Takeaways

[Music] Hello everyone and welcome to today's tutorial on variational autoenccoders. In this tutorial, we are going to build variational autoenccoders from scratch using Python and PyTorch. Autoenccoders are one of the most powerful and innovative techniques in the world of generative AI and deep learning. Now, if you want to learn more about it, then watch this video till the end. Now, before we move on, just a quick info guys. Simply learn has got EOS graduate program in IC design in practice. This is developed by IIT Bombay's department of electrical engineering. You can earn IIT Bombay alumini status and the curriculum is also developed and delivered by IIT Bombay faculty. You are going to experience campus immersion at IIT Bombay and you can earn 36 outreach program credits and a diploma from IIT Bombay. So hurry up now and join the course. The course link is mentioned in the description box. Now let us discuss the agenda for our today's session. First we are going to understand what is variational autoenccoders. Then we are going to understand and build its model architecture and finally we are going to do a hands-on to understand the inference of this by example. Now before we move on there's a short quiz to test your knowledge. So what is the primary difference between a traditional autoenccoder and a variational autoenccoder? Your option A. A variational autoenccoder uses a fixed latent representation for each input while an autoenccoder uses a probabilistic distribution. Next option, a VAE introduces a probabilistic approach to the latent space mapping each input to a distribution rather than a fixed point or an autoenccoder uses a gausian distribution in the latent space whereas a VA uses uniform distribution. And finally, a VE only performs dimensionality reduction while autoenccoder is used for data generation. So guys, mention your answers in the comment section below. Now let's get started. Imagine guys, you walk into a giant library where shelf is filled with messy and handwritten notes and each note is like a image of a handwritten digit like 0, 1, 2 or 3. And also you want to store them neatly but you don't have enough space. And exactly guys, this is our real world problem. We have lots of data like images, audio, text, etc. And we need to compress it into something smaller without losing its meaning. A variational autoenccoder is a special type of neural network that learns how to compress data into a smaller form which is called as latent space and then it reconstructs it back. Unlike normal autoenccoders, variational autoenccoders don't just memorize, they learn distribution. meaning a range of possible values so that they can generate a new data similar to what they have learned. Now let's imagine something like this. Let's say you have a magical scanner and you feed a note say a handwritten 5 rupees note and when you put this into the scanner it doesn't copies it just it summarizes the important details into a shorter code and this shorter code we call it as latent representation. So, but here's a some twist. The scanner doesn't give you a single fixed code. Instead, it says um like this note looks like five rupees note or it could also look like a three rupees note. So, instead of giving you one code, it will give you the range of possible go. So, you can understand this in this analogy. And this range is like a cloud of points in a smaller hidden space. And we call this as latent space distribution. meaning means plus variance. Now here comes a decoder machine. Let's say you pick a point from that cloud or latent space and you put it into the machine and out comes a new node something at exactly the same as the original node but sometime it's a little bit different but always believable. So you can understand this in this analogy that when you feed a real digit into the variational autoenccoder it compresses it and reconstruct it and if you pick a random point from that given latent space it can generate a brand new digit that looks like something from the data set. So in nutshell guys this is actually what a variational autoenccoder does. Now let us dive into the hands-on part and let us try to build one variational autoenccoder. I'll be using PyTorch for the demonstration and I'm going to use VS code. Now as you can see on the screen first we're going to take an input image. Then we are going to put an encoder to it and then you can see there are two parameters all over here mu and log sigma square. Then the next process is reparameterization. Next is laten z. Then you have to put decoder and finally you get a reconstructed image. Now let me explain you this little bit in depth. What is this process all about? So if you look at the input image okay what is this? This is you can consider something like a image in 28 to 28 grayscale picture like a complete handwritten digit and you have to also think like how a computer sees it. Okay. So computer is going to see it in a grid of numbers like an array of 0 and one. Zero means black and one means white. Next you can see there is a encoder present and its goal is to compress the big image into a compact description. You can think of it like a smart summarizer that keep only the essence of the digit. What it outputs guys? So unlike a normal autoenccoder, it doesn't give you one fixed code. It will give you parameter of distribution. Here the variable mu is a center of the code meaning it's a mean and log sigma square is the variance. It is like the spread of the code means the log of the variance that we are calculating all over here. Now you'd be wondering why two outputs. So a variational autoenccoder wants a smooth continuous latent space where nearby points decode to similar images representing a distribution not a single point. Let us model uncertaintity and let us sample new plausible codes. Let's say you have a input image of B cross 74. So encoder has this hidden layers which produce features. So we calculated by mu is to b into latent dimension. You can take the latent dimension as 20. And next parameter is log of sigma square where we are calculating b cross latent dimension. Now why we are choosing log of variance instead of a variance that's a very good question to ask and the answer is see a network can output any real number. So applying a exponential of log of sigma square makes a variance is strictly positive that's why we are following up with the positive values. Now let us dive ahead and try the hands-on part. See all over here we have successfully installed. So guys for this tutorial we'll be using Google collab for our hands-on part. So generally collab already have a torch pytorch installed in it. But let's make sure everything is up and running. So type this command on Google collab and let's run it. Okay. Now you can see the requirements are already satisfied mean these things are installed. Now next step is importing all the given packages. So guys you can see all over here. First I have imported torch then torch.n and torch.n n dot functional as f torch.optim as optim then I've also used tors vision package from there I'm importing data sets and transforms and from torch.util UTIL.data I'm importing the data loader and finally I'm using mattplot lab as plt. Now in the next step we have the data utilities like the mnist the transforms batching from tors vision and also importing data sets transforms from torch utils data. Now uh next you can see I have used transform equals to transform.2 tensor. Okay. So basically to convert the image and uh by using the tensor I'm normalizing it into between 0 and one. In the next step what I'm doing is I'm downloading and training the data set. So I'm using the data sets from mist model and you can say the root has been added to dot / data train equals to true download equals to true and transform put equals to transform. So this is how our uh training part is going on and this is a test data sets. So similarly data sets domin root data train false download true transform true and next I'm using data loaders with batch size of 128 you can see train loader data loader train data sets batch size 128 and I have put shuffle equals to true similarly I have done it for test loader now let us move ahead to the next part now here I'm building my ve model so basically The encoder is putting uh you know transforming uh the given thing with the help of mu and log sigma square and then it is going to do the reparameterization after that and then we can pass it to decoder and you can see we have starting with definite so this is our class v and this is nnn module so you can see I have used and then all over here as torch okay and uh then you can see all over here we have uh initialized our constructor so I can say self and I have put the input dimension as 784 hidden dimensions as 400. So 400 hidden layers and latent dimension as 20. Now I have used the super keyword. Okay. So this is VA self and I have initialized it with the init. Now here is the encoder part. After this we are putting our encoder. So you can see this is the first linear image. Okay. So this is taking the dimension as input dimension. Then the hidden 400 layers. Then it is calculating the mean of the given thing. So it has taken hidden dimension plus the latent dimension available. And finally it is calculating the variance all over here. Now these parameters are very much important because same parameter will be used for reparameterization. Then your decoder comes in. So here is self fc.2. So this is taking the linear image and this is the basically hidden image. So decoder is doing the same thing. So guys if you are uh like not much aware about encoder decoder and all those thing then watch the same video on our uh YouTube channel. Okay you can see the deep learning uh full course and in that there are all these things are covered. Now finally I calculated a function called define code. I have passed self as a parameter and one next is x and you can see I'm calculating h all over here. Then I'm calculating mu then I'm calculating log and finally I'm returning all over here mu and log var. So this is a mathematical thing what basically we do to calculate it. Similarly I'm reparameterizing this with the same thing mu and log var. So I have stored in variable called std and eps. And finally I'm returning mu plus eps std. Okay. So these are bit mathematical part. Okay. Then similarly in the decode uh function I'm passing zed as a parameter and I'm calculating h all over here. Then I'm using the sigmoid function which is a kind of activation function which is helping me to do this. Then I have put forward as one function. Then I've used x as one parameter and I could say x= to x dot view minus 1784. I'm flattening that git image. Now there you can see there are two things mu and log bar and I'm finally putting up into the encoding. So I've used this function and I have used the same parameter as x. So this all process what we are doing all over here we are basically trying to flatten the given image. Okay after the reparameterization now guys in the next part I have calculated the loss function which means the reconstruction plus the divergence which I'm creating. For that I am defining the loss function which is going to take recon x mu and log var. So basically this is bit technical and um um when you study about autoenccoders. So there are various mathematical functions involved to understand it's like how it is optimizing the given thing and creating uh a new flattened image uh with the help of you know the given process that we discovered. Uh so by optimizing these two parameters. So similarly so here uh our binary cross entropy we have calculated and here I have passed recon x x dot view and I've taken the range from -1 to 70 784 and reduction equals to sum now finally this given function the loss function is returning the values of these two which is going to tell us how much the data has been lost then you can see all over here we have a training loop so I'm defining our uh model all over here so in the parameter that is model train loader then we have optimizer which we have uh installed as optimal all over here from toss optim package and u next uh for a batch of idx for a batch of data basically I'm trying to enumerate this uh you know train loader so uh next is you know I'm trying to do data equals to data to device then optimize with the zero gradient then basically I'm using these two parameter recon batch mu log var and I'm modeling that into the given data. Finally, we are calculating the loss function. Then here loss dot backward I have called. Then finally I have done is train loss uh appended the value of the train loss to train loss plus loss item and finally I am calling the optimizer to optimize the given model. So this seems pretty much but ultimately the uh thing that we are doing is I'm training the entire process trying to optimize it okay by using the optimizer and uh this is our main functions we can use the device to do our device so if uh this is available this part of the CPU space is available then you can say print using the given device I'm transforming it to the sensor I'm loading this uh MLS it data okay I'm initializing the model and the optimizer also right now in the main function and I'm just training it for five epochs. Okay. And uh finally calling our train function in which I'm putting the model train loader optimizer and device and finally calling the main function. So in this uh five steps uh you know we have uh actually built our autoenccoder. So the initial step was the same that we discussed earlier that how this process is going to go like from uh you know taking the input of the data size then you know putting up the encoder then uh you know optimizing it using these two parameter which is law of sigma square and mu and then reparameterizing this and this passing it into the decoder so that our image uh gets flattened and optimized with the latent space. Now let us try to run this code and see what is our output. So you can see guys it has started calculating. So using this device as CPU has printed. This is on epoch one. Let's wait for a couple of time. So you can see it has loss function has calculated the average loss from the given data which is coming around 165.5998. Now let us see in epoch 2 how much it calculates. So you can see in the second iteration it has calculated 122.188 as a loss function. In the next epoch it has calculated 114.3751 and finally you can see at the final epoch it has calculated this value 109.5371. So basically we used our uh you know loss function all over here and it is generating how much you know the average loss is occurring on based on these if you are trying to train the given model. Okay. So this was a short idea regarding variational autoenccoder. How do we build it? What's the exact process? And how encoders and decoders are involved in this. Thank you guys for watching this video. If you like these kind of videos, then do not forget to hit the subscribe button and click the bell icon below so that you don't miss out any update.