How julienexr/autoencoders-lab works
This repository is a PyTorch autoencoder implementation suite with CLI training and an interactive Streamlit demo. The training workflow selects a model and dataset, creates dataloaders, trains the chosen model, and saves checkpoints; the demo loads models for reconstruction or generation. The model families are distinct branches, with VQ-VAE extended by a Transformer prior. README and excerpts establish the main capabilities; unsampled training and app code limit some call-level relationships, so those links are documented or omitted rather than inferred.
Main entry points: app.py Interactive demo · main.py Training CLI · visualization.py Visual outputs · training.py VQ-VAE training
Components 17
User entry points
- Training CLI
main.py - Interactive demo
app.py
Data and outputs
- Image datasets
src/data.py - Model checkpoints
models - Visual outputs
src/visualization.py
Autoencoder models
- Autoencoder
modules/autoencoder/ae.py - Variational AE
modules/vae/vae.py - Conditional VAE
modules/cvae/cvae.py - Vector-quantized AE
modules/vq_vae/vq_vae.py - Transformer prior
modules/vq_vae/tranformer_prior.py
Training and visualization
- AE training
modules/autoencoder/training.py - VAE training
modules/vae/training.py - CVAE training
modules/cvae/training.py - VQ-VAE training
modules/vq_vae/training.py - VQ prior inference
modules/vq_vae/inference.py - Perceptual loss
src/loss.py
Other components
- User
Connections 23
- User → Training CLI: chooses training
README.md - User → Interactive demo: opens demo
README.md - Training CLI → Image datasets: loads data
main.py - Training CLI → AE training: trains AE
main.py - Training CLI → VAE training: trains VAE
main.py - Training CLI → Model checkpoints: saves weights
main.py - AE training → Visual outputs: renders outputs
modules/autoencoder/training.py - VAE training → Visual outputs: renders outputs
modules/vae/training.py - VAE training → Perceptual loss: computes perceptual loss
modules/vae/training.py - CVAE training → Visual outputs: renders outputs
modules/cvae/training.py - VQ-VAE training → Visual outputs: renders outputs
modules/vq_vae/training.py - VQ-VAE training → Vector-quantized AE: encodes batches
modules/vq_vae/training.py - VQ-VAE training → Transformer prior: trains prior
modules/vq_vae/training.py - VQ prior inference → Transformer prior: samples indices
modules/vq_vae/inference.py - VQ prior inference → Vector-quantized AE: decodes samples
modules/vq_vae/inference.py - Interactive demo → Image datasets: loads examples
app.py - Interactive demo → Autoencoder: runs model
app.py - Interactive demo → Variational AE: runs model
app.py - Interactive demo → Conditional VAE: runs model
app.py - Interactive demo → Vector-quantized AE: runs model
app.py - Interactive demo → Transformer prior: uses prior
app.py - Interactive demo → Model checkpoints: loads weights
- Image datasets → User: downloads datasets
src/data.py