Conditional Variational Autoencoder (CVAE) on MNIST

This project implements a Conditional Variational Autoencoder (CVAE) trained on the MNIST dataset. It allows the generation of handwritten digits conditioned on their class labels (0–9), and demonstrates digit-to-digit transformation by manipulating labels during decoding.

Project Summary

• Base Model: Variational Autoencoder (VAE)
• Innovation: Conditional VAE — label is input to both encoder and decoder
• Dataset: MNIST (handwritten digits 0–9)
• Framework: PyTorch
• Creative Twist: Change the label of a digit to reconstruct it as a different digit (e.g., 7 ➝ 2)

Features

• Conditional image generation using digit labels (0–9)
• Modular code (train, test, model, dataloader)
• CVAE with reparameterization trick
• PyTorch-based implementation

How to Run

1. Clone the repository

git clone https://github.com/aunraza19/Conditional-variational-autoencoder.git

2. Change directory into the project

cd CVAE-MNIST

3. Install dependencies

pip install -r requirements.txt

4. Train the model

python main.py

Outputs:

outputs/sample_epoch_5.png — Generated digits at epoch 5

outputs/sample_epoch_10.png — Generated digits at epoch 10

outputs/loss_plot.png — Training loss over epochs

outputs/digit_transform_7_to_2.png — Creative twist result: 7 transformed into 2

Evaluation

Visual Samples: Synthetic digits generated from latent space and labels

Reconstruction Loss: Binary cross-entropy + KL divergence

Creative Twist

Use of the decoder with altered labels to transform one digit into another

License

This project is released for academic purposes. Feel free to fork and build upon it.