doodleVAE

Yet another disentangled VAE ... but for quick drawing doodles.

The Model

This project implements a Variational Autoencoder (VAE) trained to generate hand-drawn doodles. It learns a compressed latent representation of doodles from the Quick, Draw! dataset and uses it to generate new, human-like sketches.

The model consists of:

• Convolutional encoder that compresses input images into a latent vector
• Reparameterization layer to sample from the latent space
• Convolutional decoder that reconstructs images from latent vectors
• Latent space exploration, saved as an animation
• Loss plotting

The training pipeline supports configurable hyperparameters (e.g. latent dimension, beta, batch size, epochs) through a configuration file or command-line arguments.

(Back Top)

Getting Started

Follow these steps to set up and run the project locally.

Prerequisites

Ensure you have Python installed (>= 3.8 recommended).
You can install it from python.org.

Installation

1. Clone the repository

   git clone https://github.com/yassa9/doodleVAE.git
   cd doodleVAE

2. Install dependencies

   You can install everything with pip:

   pip install torch torchvision matplotlib numpy

3. Prepare your dataset

   • Provide a file <file>.npy path using --data-path.
   • You can get data from Quick, Draw!.

Training

To train the model, run:

python train.py --data-path path/to/<file>.npy

You can customize training with command-line arguments:

python train.py --data-path cat.npy --epochs 50 --latent-dim 20 --beta 4

Argument	Description
--epochs	Number of training epochs
--batch-size	Training batch size
--latent-dim	Dimensionality of latent space
--beta	Beta value for KL divergence term
--lr	Learning rate
--save-dir	Directory to save model and plots
--no-explore	Skip final latent interpolation animation

(Back Top)