Explore autoencoders (AE), denoising autoencoders (DAE), variational autoencoders (VAE), and generative adversarial networks (GAN) in this comprehensive video lecture. Learn to implement these models using PyTorch, analyze their kernels, and understand their applications in generative modeling and image processing. Dive into practical code examples, compare different architectures, and grasp key concepts such as latent space interpolation, embedding distributions, and cost network design. Gain insights into state-of-the-art inpainting techniques and discover how these models function as energy-based models (EBM). Benefit from clear explanations, including an intuitive Italian vs. Swiss analogy for understanding GANs, and conclude with a hands-on PyTorch code reading session for GAN implementation.
AE, DAE, and VAE with PyTorch - Generative Adversarial Networks and Code
Overview
Syllabus
– 1st of April 2021
– Training an autoencoder AE PyTorch and Notebook
– Looking at an AE kernels
– Denoising autoencoder recap
– Training a denoising autoencoder DAE PyTorch and Notebook
– Looking at a DAE kernels
– Comparison with state of the art inpainting techniques
– AE as an EBM
– Training a variational autoencoder VAE PyTorch and Notebook
– A VAE as a generative model
– Interpolation in input and latent space
– A VAE as an EBM
– VAE embeddings distribution during training
– Generative adversarial networks GANs vs. DAE
– Generative adversarial networks GANs vs. VAE
– Training a GAN, the cost network
– Training a GAN, the generating network
– A possible cost network's architecture
– The Italian vs. Swiss analogy for GANs
– Training a GAN PyTorch code reading
– That was it :D
Taught by
Alfredo Canziani
