This project implements a Conditional Variational Autoencoder (CVAE) for generating and exploring handwritten digits from the MNIST dataset. The CVAE conditions image generation on digit labels (0-9) while capturing style variations (e.g., thickness, slant) in a low-dimensional latent space. The model is trained to maximize the conditional Evidence Lower Bound (ELBO), balancing reconstruction fidelity and latent regularization.
-
Data Preparation: Loaded the MNIST dataset using PyTorch's
torchvision. Visualized examples with one digit per class (0-9) and a random 5x5 batch to confirm data integrity. Images are grayscale (1x28x28) normalized to [0,1]. -
Model Architecture:
- Encoder: Convolutional layers process concatenated image and one-hot label, followed by FC layers to output mean (μ) and log-variance (logσ²) for the latent distribution (z_dim=16).
- Decoder: FC layer expands concatenated z and label, followed by transposed convolutions to reconstruct the image.
- CVAE Wrapper: Handles reparameterization trick for sampling z ~ N(μ, σ) during training.
-
Loss Function: Negative ELBO minimized via Binary Cross-Entropy (reconstruction) + KL Divergence (latent regularization to N(0,I)).
-
Training: Trained for 30 epochs on a DataLoader (batch_size=128) using Adam optimizer (lr=1e-3). Monitored per-example total loss, reconstruction, and KL terms.
-
Inference and Visualization:
- Generated samples for a chosen digit (e.g., 2) in a 3x5 grid.
- Created a horizontal strip of generated digits matching a 9-digit student ID (e.g., 404405406).
- Explored latent space with 9x9 traversal grids for digits 0,1,3,5,7,8, varying two random orthogonal directions (±3σ).
- Generated a style-controlled student ID, progressively varying one latent dimension (e.g., dim=2 for thickness) from thin to thick across the 9 digits.
The implementation uses PyTorch for all components, with device-agnostic code (CPU/GPU). Training converges to stable generations with diverse styles.
- Horizontal strip of generated digits for ID 404405406:
- ID 404405406 with progressive style variation (thin to thick along latent dim 2):
