This repository contains an experimental implementation of a Variational Autoencoder (VAE) using TensorFlow. The project was developed for learning purposes, with a focus on exploring generative models through unsupervised learning.
A Variational Autoencoder (VAE) is a type of generative model that learns a probabilistic mapping from input data to a latent space, allowing the generation of new data samples similar to the training distribution. VAEs are commonly used for tasks such as image generation and dimensionality reduction.
This implementation provides a customizable framework for building and training a VAE model on various datasets using TensorFlow.
- Encoder-Decoder Architecture: Implements the core VAE structure with an encoder to map inputs to latent variables and a decoder for reconstruction.
- Latent Space Regularization: Ensures the latent space follows a Gaussian distribution via the KL divergence loss component.
- Customizable: Modify the network architecture, including layers, units, and activation functions.
- Training and Evaluation: Scripts for training the model and visualizing results.
- Python 3.6+
- TensorFlow 2.x
- NumPy
- Matplotlib (for visualizations)
- Clone this repository:
git clone https://github.com/blackhat-coder/Tensorflow-VAE.git cd Tensorflow-VAE -
pip install -r requirements.txt
-
Prepare your dataset in a format compatible with the model (e.g., images in
.jpgor.pngformat). -
Train the model using the
train_vae.pyscript:python train_vae.py --dataset <path_to_dataset> --epochs 100
Visualize Results using the visualize.py script:
python visualize.py --model_path <path_to_saved_model> --latent_dim 2We welcome contributions to improve this playground 😉!
This project is licensed under the MIT License.
MIT License
This project is based on the concepts presented in the paper:
Kingma, Diederik P., and Max Welling. "Auto-Encoding Variational Bayes." arXiv preprint arXiv:1312.6114 (2013).