Variational Auto Encoders

This code implements a Variational Autoencoder (VAE) for generating and manipulating facial images using TensorFlow and Keras. Here's a brief description and definition of major concepts:

1. Variational Autoencoder (VAE):

   • A generative model that learns to encode and decode data.
   • Comprises an encoder that maps input data to a probability distribution in the latent space and a decoder that reconstructs the data from sampled points in this space.
   • Includes a sampling layer that introduces stochasticity, enabling generation of diverse samples.

2. Encoder and Decoder:

   • Encoder: Accepts input images and maps them to the latent space, producing mean and log-variance vectors.
   • Decoder: Reconstructs images from sampled points in the latent space.

3. Sampling Layer:

   • Introduced in the Sampling class, it samples from the learned distribution in the latent space during model training.

4. Training the VAE:

   • Trains the VAE using a custom loss function composed of a reconstruction loss and a KL divergence term.
   • The model is trained on a dataset of facial images to learn a representation in the latent space.

5. Data Preprocessing:

   • Images are loaded and preprocessed using TensorFlow's image_dataset_from_directory function.
   • Images are normalized to the range [0, 1] and converted to a float32 format.

6. Callbacks:

   • Utilizes various callbacks, such as model checkpointing and TensorBoard, for monitoring and saving the training progress.

7. Image Generation:

   • Generates images during training and saves them as checkpoints using the ImageGenerator callback.

8. TensorBoard Logging:

   • Logs training metrics and visualizations to TensorBoard for monitoring and analysis.

9. Testing and Visualization:

   • Samples images from the trained VAE and visualizes both real and reconstructed images.
   • Visualizes the latent space by sampling points and decoding them.

10. Attribute Manipulation:

• Defines functions to extract an attribute vector for a specific facial attribute (e.g., Blond Hair).
• Manipulates the latent space to emphasize or de-emphasize the chosen attribute.

11. Facial Morphing:

• Illustrates facial morphing by interpolating between latent space representations of two images.

12. Face Dataset and Labels:

• Uses the CelebA dataset, containing facial images of celebrities.
• Labels such as hair color are utilized to perform attribute manipulation.

The code demonstrates the training and use of a VAE for generating realistic facial images and manipulating facial attributes in the latent space.