Generative AI, from GANs to CLIP, with Python and Pytorch

Duration

Generative AI, from GANs to CLIP, with Python and Pytorch

2 days

What you'll learn

• How to code generative A.I architectures from scratch using Python and Pytorch
• How generative architectures work, in great depth, from GANs to multimodal A.I, understanding every little detail in the process
• In addition to the coding, every section begins with an in-depth review of the key concepts related to these architectures
• Examples: We will code a generative network that produces human faces, and also combine two advanced networks to transform text prompts into amazing images.
• Examples: We will learn to edit the clothes of a person in a picture by combining a segmentation architecture with the Stable Diffusion generative model

Prerequisites

• Basic knowledge of python. It's enough with the very basics, as we will code every little thing together, line by line
• Access to an internet connection, as we will use the free online Google Colab service to code together

Course Outline

The generative AI revolution

• The roadmap, from basic to advanced and beyond
• The generative revolution: coming home
• The present and future of AI is generative
• Applications of generative AI
• Latent spaces and representation learning
• Navigating latent spaces
• GANS: Generative Adversarial Networks
• Benefits and possibilities of Generative AI
• Coming home: generative AI and human nature

Coding a basic generative architecture

• Understanding the battle between generator and discriminator
• Understanding Cross Entropy in depth
• Understanding the equation to calculate the discriminator loss
• Understanding the equation to calculate the generator loss
• Coding: importing libraries and declaring a visualization function
• Coding: hyperparameters and the DataLoader
• Coding: the generator class
• Coding: the discriminator class
• Coding: the optimizer and testing the generator
• Coding: the loss values of generator and discriminator
• Coding: main training loop, discriminator part
• Coding: main training loop, generator and stats
• Coding: running the training
• Coding: results and conclusions

Coding an advanced generative architecture

• Challenges and issues of the basic GAN
• The Wasserstein Loss
• The Gradient Penalty
• Coding: setting up libraries and parameters
• Coding: Login and setup of the Wandb stats library
• Coding: Beginning the generator
• Coding: Understanding convolutions
• Coding: The generator class
• Coding: The critic class
• Coding: Alternative way to initialize parameters (optional)
• Coding: Loading the CelebA dataset
• Coding: Declaring dataset, dataloader and optimizers
• Coding: the gradient penalty
• Coding: saving and loading checkpoints
• Coding: training loop - critic training
• Coding: training loop - generator training
• Coding: stats and fixing issues
• Coding: reviewing the code before running the training
• Coding: running the training
• Coding: results after a few epochs
• Coding: results after a few more epochs
• Coding: results getting better and better
• Coding: morphing between points in latent space
• Coding: more morphing

Generating images from text by combining two advanced architectures

• Multimodal generation, an incredible adventure
• Coding: importing the libraries
• Coding: helper functions and hyperparameters
• Coding: Setting up the CLIP model
• Coding: Setting up the Generative transformer model
• Coding: Setting up the latent space parameters to be optimized
• Coding: encode the text prompts through CLIP
• Coding: creating crops from the generated image
• Coding: a function to display generated images and crops
• Coding: optimizing the latent space parameters
• Coding: the training loop
• Coding: running the training
• Coding: interpolating between points in the latent space
• Coding: creating a video of the interpolations and general review
• Coding: creating variations of the code
• Coding: Davinci Sfumato: Tweaking the code to create a new kind of texture
• Coding: Davinci Sfumato: reflecting about the process
• Final greetings from the spacecraft

Editing people's clothes by combining segmentation and generative AI models

• Intro: people's clothes replacement and editing using Generative AI
• Coding: Setting up libraries and the segmentation model
• Coding: Setting up the Stable Diffusion generative model
• Coding: Loading a picture and running the segmentation process to produce masks
• Coding: Visualizing the generated masks
• Coding: Inpainting, running and experimenting with the Stable Diffusion model
• Coding: Guide the segmentation process with text prompts
• Coding: run the generative model in this alternative setup

Programme scope, delivery arrangements and suitability are confirmed during enquiry.