[PyTorch] Learn GAN easily and quickly

It occupies a paradigm of deep learning.
Learn about GANs! 🙆🏻‍♂

I organized the lectures given at Inflearn and published them in a book titled 'Python Deep Learning PyTorch'.
Thank you for your interest : )
(Inflearn lectures have been updated as of 2020.10.06. We will continue to update the lectures.)

Learn GAN easily and quickly with Pytorch

Introduction to the course

Most of the artificial intelligence we talk about these days utilizes deep learning models. General machine learning or deep learning models ended with classification and regression. However, the emergence of GAN brought about a paradigm shift that is so great that it is no exaggeration to say that it has advanced the development of artificial intelligence by one step.

Going beyond classifying and predicting data to generating it was unimaginable at the time (4-5 years ago). The emergence of GANs, along with reinforcement learning (the basic principle of AlphaGo), has made them an indispensable field in artificial intelligence.

The image below is a fake image that does not actually exist in this world, created by the best performing GAN from a year ago. Now, more advanced models have emerged.
This lecture will explain the exact concept and learning principles of GAN step by step, and also talk about the shortcomings and future development directions of GAN.

🌈 Vanilla GAN

We cover the concept and learning principles of GAN.
Rather than simply explaining the concepts, we explain the proper learning principles and provide mathematical proofs.
We will explain the learning process and method, the shortcomings of GAN when it first came out, and the future direction of GAN.

🌈 DCGAN/LSGAN/CGAN (GAN that is gradually improving)

Since the advent of Vanilla GAN, GANs have been evolving at an incredibly rapid pace.
First, DCGAN, which applied the principles of GAN to CNN, was introduced, and then LSGAN, which only slightly modified the loss, appeared, and various GANs are being introduced.
The picture below shows a performance comparison between DCGAN and LSGAN.

🌈 CycleGAN (GAN, the basic model for style transfer)

[Beyond generating data] Using the generative principle of GAN, it has begun to develop into various fields. Among them, the representative model is CycleGAN, a style transfer model. It can change a picture into a photo or a photo into a painting, change day and night, and change seasons.
The GAN model that swaps the two domains of an image like this is CycleGAN. This CycleGAN has become the basic baseline model of GANs that utilize style transfer.

CycleGAN has been applied and developed as shown in the figure below. In addition to this model, it has been developed into various networks.

🌈 CAN (GAN model that generates art)

Generating data doesn't create anything new because it's ultimately generated from within the training data. That's why it's far from being art. Because if you generate from within the training data, it's just 'imitation'.
The CAN model slightly changes the learning principle of GAN to generate artwork. It is said that it obtained scores similar to those of real artworks by surveying humans.

🌈 Various fields where GAN is applied/developed

Besides, GANs are developing in a variety of fields.
Here is a brief introduction to various GANs, including Radial GAN for generating structured data for machine learning rather than images, DeliGAN, a model for how to generate diverse and high-quality images in situations where training data is limited, MGAN, which combines multiple GAN models, and SRGAN, which converts low-quality images into high-quality images.

Practice material link: https://github.com/LeeGyeongTak/torchgan

Go to related lectures

[PyTorch] Learn Deep Learning Quickly and Easily
Quickly learn the concepts and related knowledge of deep learning.