Variational Autoencoder (VAE)

This project implements a (Linear) Variational Autoencoder (VAE) following the paper "Auto-Encoding Variational Bayes" by Diederik P. Kingma et al. The model is trained on the MNIST dataset and can be used to generate new samples.

Overview

The Variational Autoencoder is a generative model that learns a probabilistic mapping between input data and a latent space. It is trained to encode input data into a distribution and decode samples from that distribution back into the input space. This project provides a basic implementation of a VAE using PyTorch and demonstrates how to train the model on the MNIST dataset.

Requirements

• PyTorch
• torchvision
• Matplotlib (for visualization)

Usage

Training

To train the linear VAE model, run the train.py script. You can customize training parameters such as learning rate, epochs, and batch size using command-line arguments. Example:

python train.py --learning_rate 0.001 --epochs 20 

Digit-Specific Generation:

Users can specify a digit to generate, and the model will generate samples based on the learned characteristics of that digit.

generate.py  --checkpoint_path  <path/to/checkpoints.pth> --digit_to_generate 0

Interpolation:

The project supports interpolating between two specified digits in the latent space. Users can visualize the transition between two digits by interpolating their latent representations.

generate.py  --checkpoint_path  <path/to/checkpoints.pth> --from_digit 0 --to_digit 1