Master GAN with TensorFlow: Create Art with Neural Networks

Explore vanilla GANs, which use implicit density models where a generator converts noise into images and a discriminator differentiates real from fake, trained alternately.

Learn to implement a vanilla GAN with dense layers in TensorFlow, building the generator and discriminator, training loops, and visualizing results with GIFs on MNIST.

Train a deep convolutional gan (dcgan) with convolutional layers, leaky relu, and batch normalization, while comparing sigmoid and tanh activations and noting a switch to layer normalization.

Explore how Wasserstein GANs use Earth Mover's distance and a critic network to solve instability, mode collapse, and convergence issues in GANs, with Lipschitz constraints and gradient penalties.

Learn how gradient penalty stabilizes WGAN training by enforcing a one Lipschitz constant and replacing weight clipping, using per-example interpolated images and a penalty term.

Implement the Wasserstein loss in TensorFlow for WGAN-GP, compute real vs fake critic scores, train the critic multiple times per generator, and apply gradient penalty with epsilon and lambda.

Build a generator and a critic in a wgan-gp setup, train the critic multiple times per cycle with gradient penalty and wasserstein loss, using tensorflow on fashion-mnist 28x28 grayscale images.

Explore how conditional GANs enable image generation based on user labels, using embedding layers, one-hot encoding alternatives, and label integration in the generator and discriminator.

Implement conditional GAN in TensorFlow using Fashion-MNIST, embedding labels, and a high batch size. Train generator and discriminator with embedding-based conditioning, and assess real and fake outputs.

Explore pix2pix, a pixel-to-pixel image translation using conditional adversarial networks to convert sketches or edge maps into real images via a U-Net generator and patch GAN discriminator.

Train the pix2pix architecture on a Kaggle dataset mapping cityscapes to shoes for cats, using segmented masks as input and real images as targets with a TensorFlow pipeline.

Explore how CycleGAN enables unpaired image-to-image translation using cycle consistency and identity losses across two translators, enabling zebra to horse or day to night conversions without paired data.

Implement CycleGAN in TensorFlow using U-Net generator, prepare Kaggle data by splitting real and fake halves, normalize, and train with gradient tape using mean squared error, cycle, and identity losses.

Explore neural style transfer by merging a content image with the painting style from a style image using a pre-trained VGG19 network, Gram matrix, and iterative stylized image optimization.

Learn to build your first neural style transfer model in TensorFlow, using VGG features, gram matrices, and gradient updates to merge content and style images.

Explore feedforward style transfer with a trainable convolutional network that stylizes images in a single pass, using a two-part architecture, VGG features, gram matrices, and perceptual loss for real-time effect.

Develop a feedforward style transfer network using VGG-based content and style activations with a U-Net, trained on a raw JPEG Kaggle dataset with perceptual loss.

Master arbitrary style transfer in a single pass with adaptive instance normalization, matching style statistics via mean and variance of VGG activations, using a VGG19 encoder and decoder.

Set up and train an arbitrary style transfer pipeline using a VGG-based encoder, adaptive instance normalization, and a decoder to blend content and style.