Generative AI Bootcamp with Python: Gen AI & Agentic AI

Explore the basics of deep neural networks, including input, hidden, and output layers, neurons, edge weights, and activation functions; learn how a learning algorithm trains the network to recognize patterns.

Explore building a multilayer perceptron for CIFAR-10 image classification with TensorFlow, using dense layers with ReLU and softmax, one-hot encoding, and train/test splits.

Explore how convolutional neural networks learn spatial hierarchies of features using 2D convolutions, strides, and padding, then stack layers with batch normalization and dropout for robust image classification.

Learn how autoencoders enable generative modeling by encoding fashion MNIST images into a 2D latent space and decoding them back with a Keras TensorFlow model.

Explore variational autoencoders, mapping images to a multivariate normal latent space with a sampling layer and KL divergence loss in a keras build to generate new images.

Learn to design and train a deep convolutional GAN in Keras, generating 64x64 lego brick images from latent space inputs while mastering generator and discriminator training and common DCGAN challenges.

Explore the Wasserstein GAN with gradient penalty (WGAN-GP) to stabilize training, build from scratch in Keras, and generate faces while addressing discriminator strength, mode collapse, and training challenges.

Explore how a conditional GAN uses label conditioning to control generated images, such as blonde vs non-blonde hair. Implement it in Keras with one-hot labels and a critic input.

Explore autoregressive models, with LSTM at the core, for text generation and explicit data modeling in sequential data. Learn text preprocessing, tokenization strategies, and embedding-based RNN architectures in Keras.

Explore stacked LSTM extensions, GRU gates, and bidirectional layers to learn deeper text representations. Delve into PixelCNN with masked convolutions and residual blocks for autoregressive image generation on Fashion-MNIST.

Explore normalizing flow models, connecting their invertible mappings to variational autoencoders and autoregressive models, and learn how the change of variables with the jacobian determinant enables tractable sampling.

Explore real NVP normalizing flow models, using coupling layers to transform a 2d two moons dataset into a simple gaussian, with invertible mappings and easy jacobian determinants.

Explore energy-based models and their Boltzmann-based scoring, learn to sample with Longeverne dynamics, and train with contrastive divergence using a replay buffer on MNIST data.

Explore diffusion models for image generation, including forward noising and denoising with a Keras-based U-Net. Learn about cosine diffusion schedules and EMA training on the Oxford 102 flower dataset.

Explore transformers and attention for text generation, build a mini GPT in Keras with the wine reviews dataset, and learn pre-training, fine-tuning, and causal masking.

Explore progressive training with ProGAN to generate images, analyze StyleGAN, StyleGAN2, and self-attention GAN architectures, and learn key concepts like adaptive instance normalization, weight modulation, demodulation, and path length regularization.

Explore style-based image generation with StyleGAN and StyleGAN2, including mapping z to w, injecting style vectors at multiple layers, style mixing, stochastic noise, and artifact-reducing weight modulation and skip connections.