Hugging FaceVAE
Overview
From-scratch Variational Autoencoder trained on CelebA face images at 128×128 resolution. Demonstrates reconstruction, novel face generation by sampling from the Gaussian prior, and latent space arithmetic (e.g. adding/subtracting attribute directions). Based on Auto-Encoding Variational Bayes (Kingma & Welling, 2014).
Architecture
Encoder: 4× Conv2d (3→128→256→256→256, stride=2) → linear → μ and log σ² (32D latent)
Decoder: Linear → 4× ConvTranspose2d → 128×128 RGB image
- Reparameterisation trick for differentiable sampling
- Loss: MSE reconstruction + KL divergence
- Activation: LeakyReLU (slope=0.01)
- WandB tracking
Training
| Hyperparameter | Value |
|---|---|
| Dataset | CelebA (202,599 images, 80/20 split) |
| Epochs | 200 (checkpoint at epoch 240) |
| Optimizer | Adam, lr=5e-4 |
| Batch size | 32 |
Paper
Auto-Encoding Variational Bayes — Kingma & Welling, 2014