Keras Deep Learning & Generative Adversarial Networks (GAN)

Hi There!

Hello and welcome to my new course Deep Learning with Generative Adversarial Networks (GAN). This course is divided into two halves. In the first half, we will deal with Deep Learning and Neural Networks and in the second half on top of that, we will continue with Generative Adversarial Networks or GAN or we can call it 'gan'. So let's see what are the topics that are included in each module. At first, the Deep Learning one.

As you already know the artificial intelligence domain is divided broadly into deep learning and machine learning. In fact, deep learning is machine learning itself but Deep learning with its deep neural networks and algorithms tries to learn high-level features from data without human intervention. That makes deep learning the base of all future self-intelligent systems.

I am starting from the very basic things to learn like learning the programming language basics and other supporting libraries at first and proceed with the core topic.

Let's see what are the interesting topics included in this course. At first, we will have an introductory theory session about Artificial Intelligence, Machine learning, Artificial neuron-based Deep Learning, and Neural Networks. 

After that, we are ready to proceed with preparing our computer for Python coding by downloading and installing the Anaconda package and will check and see if everything is installed fine. We will be using the browser-based IDE called Jupyter Notebook for our further coding exercises.

I know some of you may not be coming from a Python-based programming background. The next few sessions and examples will help you acquire basic Python programming skills to proceed with the sessions included in this course. The topics include Python assignment, flow-control, functions List and Tuples, Dictionaries, Functions, etc. 

Then we will start with learning the basics of the Python Numpy library which is used to add support for large, multi-dimensional arrays and matrices, along with a large collection of classes and functions. Then we will learn the basics of matplotlib library which is a plotting library for Python for corresponding numerical expressions in NumPy. And finally, the panda's library is a software library written for the Python programming language for data manipulation and analysis. 

After the basics, we will then install the deep learning libraries theano, TensorFlow and the API for dealing with these called as Keras. We will be writing all our future codes in Keras.

Then before we jump into deep learning, we will have an elaborate theory session about the basic Basic Structure of an Artificial Neuron and how they are combined to form an artificial Neural Network. Then we will see what exactly is an activation function, the different types of most popular activation functions and the different scenarios we have to use each of them.

After that, we will see about the loss function, the different types of popular loss functions and the different scenarios we have to use each of them. 

Like the Activation and loss functions, we have optimizers that will optimize the neural network based on the training feedback. We will also see the details about the most popular optimizers and how to decide in which scenarios we have to use each of them.

Then finally we will discuss about the most popular deep learning neural network types and their basic structure and use cases.

Further, the course is divided into exactly two halves. The first half is about creating deep learning multi-layer neural network models for text-based datasets and the second half is about creating convolutional neural networks for image-based datasets. 

In the Text-based simple feed-forward multi-layer neural network model we will start with a regression model to predict house prices in King County USA. The first step will be to Fetch and Load Dataset from the Kaggle website into our program. 

Then as the second step, we will do an EDA or an Exploratory Data Analysis of the loaded data and we will then prepare the data for giving it into our deep learning model. Then we will define the Keras Deep Learning Model. 

Once we define the model, we will then compile the model and later we will fit our dataset into the compiled model and wait for the training to complete. After training, the training history and metrics like accuracy, loss etc can be evaluated and visualized using matplotlib. 

Finally, we have our already trained model. We will try doing a prediction of the King County real estate price using our deep learning model and evaluate the results.

That was a text-based regression model. Now we will proceed with a text-based binary classification model. We will be using a derived version of the Heart Disease Data Set from the UCI Machine Learning Repository. Our aim is to predict if a person will be having heart disease or not from the learning achieved from this dataset. The same steps repeat here also. 

The first step will be to Fetch and Load Dataset into our program. 

Then as the second step, we will do an EDA or an Exploratory Data Analysis of the loaded data and we will then prepare the data for giving it into our deep learning model. Then we will define the Keras Deep Learning Model. 

Once we define the model, we will then compile the model and later we will fit our dataset into the compiled model and wait for the training to complete. After training, the training history and metrics like accuracy, loss etc can be evaluated and visualized using matplotlib. 

Finally, we have our already trained model. We will try making a prediction for heart disease using our deep learning model and evaluate the results.

After the text-based binary classification model. Now we will proceed with a text-based multi-class classification model. We will be using the Red Wine Quality Data Set from the Kaggle website. Our aim is to predict the multiple categories in which a red wine sample can be placed from the learning achieved from this dataset. The same steps repeat here also. 

The first step will be to Fetch and Load Dataset into our program. 

Then as the second step, we will do an EDA or an Exploratory Data Analysis of the loaded data and we will then prepare the data for giving it into our deep learning model. Then we will define the Keras Deep Learning Model. 

Once we define the model, we will then compile the model and later we will fit our dataset into the compiled model and wait for the training to complete. After training, the training history and metrics like accuracy, loss etc can be evaluated and visualized using matplotlib.

Finally, we have our already trained model. We will try doing a prediction for wine quality with a new set of data and then evaluate the categorical results.

We may be spending much time, resources and effort to train a deep learning model. We will learn about the techniques to save an already-trained model. This process is called serialization. We will at first serialize a model. Then later load it in another program and do the prediction without having to repeat the training.

That was about text-based data. We will now proceed with image-based data. In the preliminary session, we will have an introduction to Digital Image Basics in which we learn about the composition and structure of a digital image.

Then we will learn about Basic Image Processing using Keras Functions. There are many classes and functions that help with pre-processing an image in the Keras library API. We will learn about the most popular and useful functions one by one.

Another important and useful image processing function in keras is Image Augmentation in which slightly different versions of images are automatically created during training. We will learn about single image augmentation, augmentation of images within a directory structure and also data frame image augmentation. 

Then another theory session about the basics of a Convolutional neural network or CNN. We will learn how the basic CNN layers like the convolution layer, the pooling layer and the fully connected layer work. 

There are concepts like Stride Padding and Flattening in convolution for image processing. We will learn them also one by one.

Now we are all set to start with our CNN Model. We will be designing a model that can classify 5 different types of flowers if provided with an image of a flower in any of these categories. We will first download the dataset from the Kaggle website. Then the first step will be to Fetch and Load this Dataset from our computer into our program. 

Then as the second step, we have to split this dataset manually for training and then later testing the model. We will arrange them into training and testing folders with each class labelled in separate folders.

Then we will define the Keras Deep Learning Model. Once we define the model, we will then compile the model and later we will fit our dataset into the compiled model and wait for the training to complete. After training, the training history and metrics like accuracy, loss etc can be evaluated and visualized using matplotlib. 

Finally, we have our already trained model. We will try doing a prediction for five different types of flowers with a new set of image data and then evaluate the categorical results.

There are many techniques that we can use to improve the quality of a model. Especially an image-based model. The most popular technique is doing dropout regularization of the model. 

The next technique is doing the optimization and adjustment of the padding and also the filters in the convolution layers. 

And finally, optimization using image augmentation. We will tweak different augmentation options in this session.

Doing these optimization techniques manually one by one and comparing results is a very tedious task. So we will be using a technique called Hyperparameter tuning in which the Keras library itself will switch different optimization techniques that we specify and will report and compare the results without having to interfere with it.

Even though these techniques and the creation of a model from scratch are fun. It's very time-consuming and may take ages if you are planning to design a large model. In this situation, a technique called transfer learning can help us. 

We will take the world-renounced, state-of-the-art, most popular pre-trained deep learning models designed by experts and we will transfer the learning into our model so that we can make use of the architecture of that model into our custom model that we are building.

The popular state-of-the-art model architectures that we are going to use are the VGG16, VGG19 designed by deep learning experts from the University of Oxford and also ResNet50 created in  ImageNet challenge to address the vanishing gradient problem.

We will at first download these models using Keras and will try simple predictions using these pre-trained models. Later we will try the network training for our flower dataset itself using the VGG16. We will make a few changes in the model to incorporate our dataset into it. Since the network architecture is not that simple, on our computer it will take a lot of time to complete the training. 

So instead of a CPU, we have to use a GPU to enhance parallel processing. We will be using a cloud-based Free GPU service provided by Google called Google Colab. At first, we will try training with VGG16 in Google Colab. We will prepare, zip and upload the dataset into Google Colab. Then we will extract it using Linux commands and then do the training. The training is almost ten times faster compared to the local computer. Once we have the trained model we will serialize the model and will do the prediction. 

The same procedure will be repeated for VGG19 and also for ResNet.

After having enough knowledge about Deep Learning, we will then move on with our Generative Adversarial Network or GAN.

At first, we will have an introduction to GAN. We will discuss about the the basic things inside the GAN, the two different types of networks called the Generator and the Discriminator. And then we will attempt to do a simple Transpose Convolution using a grayscale image. 

Transpose convolution is the opposite of the convolution operation that we had in Deep Learning and in the next session, we will once again have a thorough discussion about the Generator and the Discriminator mechanism inside a GAN. After that, we will try to implement a Fully Connected, simple GAN using the MNIST dataset. We will have a step-by-step approach to creating this fully connected GAN or fully connected GAN. 

At first, we will be loading the MNIST dataset. Then we will proceed with defining the Generator function. Once we have the Generator function, we will then define the Discriminator function. Then we will combine these Generator and Discriminator models for our fully connected GAN so that we will have a composite model. And then this model will be compiled so that it can be used for training. After the compilation, we will proceed with training. We will train the discriminator at first and then we will proceed with the generator training. We will also define functions so that during the training of this fully connected gan, we will save the log at regular intervals and we will plot the graph. Along with the graph, the image is generated after each batch of training. And once the training is complete, we will

see how it performed. We will see the generated images by this fully connected GAN and also we will save this model

so that we can use it for future image generation without having to train it again. We will see how we can do that. 

Once we complete the fully connected GAN, we will then proceed with a more advanced Deep Convoluted GAN or DCGAN. For DCGAN also, we will discuss what is a DCGAN. What's the difference between a Da CGAN with a Fully Connected GAN? Then we will try to implement the Deep Convolution GAN or DCGAN At first we will define the Generator function and then we will define the Discriminator function. After that, we will combine this generator and discriminator models and we will have a composite model. We will then compile this model. Once the compilation is complete, we will proceed with training the model. Since deep convoluted GAN or DCGAN is more complex than a fully connected GAN, it will take much time to train it.

So we will move the training from our local computer to our Google Colab using GPU. We will train the model and we will use that to generate the MNIST handwritten digits.

We will use this same deep convolution GAN with another dataset also like the MNIST Fashion dataset. We will train the model using GPU and we will generate images. Then we will have the MNIST handwritten digits and also the MNIST Fashion dataset are all simple grayscale images. 

We will try colour images also. We will be having the CIFAR-10 dataset. We will define the generator first. Because it is a colour dataset, we need to adjust the model. So we will define the generator, and then we will define the discriminator once again. Then we will proceed with training using the CIFAR-10 dataset. The training will be done using Google Colab GPU and then we will try to generate images using that trained model.

Then we will have a brief discussion about Conditional Generative Adversarial Networks or Conditional GAN. We will compare the normal Vanilla GAN with Conditional GAN. What's the difference between a normal GAN and a Conditional GAN? And then we will proceed with implementing the Conditional GAN, which is a bit different from our normal GAN. We will first define the basic generator function and then we will have a Label Embedding for this generator. Then we will define the basic discriminator function and also we will have the label embedding for this discriminator function. Then we will combine this generator and discriminator. We will combine and we will compile. After that, we will proceed with

training the GAN model using our local computer. And we will try to display the generated images. Then we will use the same code, upload it to Google Colab and do the training with Google Colab. And that was for the MNIST handwritten digits dataset. We will then proceed with training the Conditional GAN using the MNIST Fashion dataset. The same dataset that we used for our fully connected and also deep convoluted GAN. Once we complete the training, we will try to generate images using the conditional GAN for our Fashion MNIST dataset. In the next session, we will discuss about the other popular types of GAN and also I will provide you with a git repository that was shared by a deep learning and machine learning expert. I will share that link so that you can try those exercises by yourself and also I will teach you how you can fork that repository into your pit repository so that you can try the code, change the code

and see how it performs.

And that's all about the topics which are currently included in this quick course. The code, images, models and weights used in this course have been uploaded and shared in a folder. I will include the link to download them in the last session or the resource section of this course. You are free to use the code in your projects with no questions asked.

Also after completing this course, you will be provided with a course completion certificate which will add value to your portfolio.

So that's all for now, see you soon in the classroom. Happy learning and have a great time.