Normalizing Flows Explained | Flow Matching Part-1 | Generative AI

Описание: In this tutorial video, we dive deep into Normalizing Flows - both explanation and implementation.

We’ll begin with why normalizing flows are important when we already have VAEs and GANs in generative modeling. Once we have understood the motivation, we will get into what normalizing flows are, starting with the foundation behind flow-based models which is - Change of Variables Theorem for probability densities. As part of understanding change of variables theorem for multi dimensional cases, we’ll explore the role of the Jacobian in normalizing flows. At this point we would have the understanding that normalizing flows are just modelling single transformations and now from modelling a single function, we move to using normalizing flows to model compositions of invertible functions, enabling us to convert simple distributions to complex ones with decent success.

As an example of a deep generative model using the normalizing flow technique, we will cover Real NVP paper but focusing mainly on affine coupling layers rather than entire paper.

Finally, we’ll walk through a step-by-step PyTorch implementation of a RealNVP-like model on the MNIST dataset and look at the results we get.

⏱️ Timestamps
00:00 Introduction
00:25 Why Normalizing Flows
02:22 Change of Variables Theorem Explained
11:46 Role of Jacobian in Normalizing Flows
15:30 Composition of invertible functions
19:32 Design Constraints of a Normalizing Flows model
21:09 Real NVP and Affine Coupling Layers
29:18 PyTorch implementation of a Normalizing Flows model(RealNVP like)
39:58 Results

📖 Resources
Github Implementation Link(to be uploaded soon) - https://github.com/explainingai-code/...

🔔 Subscribe:
https://tinyurl.com/exai-channel-link
Email - [email protected]