Deep generative models for image...
summary
Walks through building a convolutional Variational Autoencoder + KL loss in Keras/TensorFlow on the Fashion dataset (60,000 train / 10,000 test), trains (demo uses ~50 epochs on Colab GPU), shows a smooth 2D latent space but generated images remain blurry and suggests combining VAE decoder with a GAN discriminator. Also provides RNN/LSTM text models with sequence length 10, vocab 14,833 and temperature sampling producing Shakespeare‑like fragments, notes long‑range dependency failures and points to Transformer/GPT‑2 as the practical fix.
AI-written summary · about
chapters (21)
- 0:00Generative models intro
- 2:05Generative model basics
- 3:56Variational autoencoders
- 9:44VAEs solve autoencoder issues
- 16:16Combining variational autoencoders with GANs
- 19:10Setup image generation
- 21:20Exploring notebook
- 22:40Fashion dataset loading
- 28:29VAE model training
- 33:35Convolutional layers
- 36:03Model training & results
- 41:03Latent space visualization
- 47:41Transition to text RNNs
- 48:10Introduction to text generation with recurrent neural networks
- 57:46LSTM networks
- 1:01:24Transformers for text
- 1:02:05Setup text generation
- 1:02:48Shakespeare preprocessing
- 1:06:57RNN model training
- 1:10:53Text generation sampling
- 1:14:35Text generation limits
related talks
24mA Journey through Generative AI's Frontier Python Libraries
30mHow to train a large diffusion model
41mBuilding Generative Image & Video models at Scale
22mEfficient Pretraining of Text-to-Image Models
28mPosterior Collapse in Deep Generative models
30mVisual data abundant relevant labelled cheap Pick two
37mHarnessing the power of Generative Adversarial Networks (GANS) for supervised learning
44mGenerative Adversarial Networks (GANs) and Stable Diffusion