Skip to content

autoencoders

autoencoders is a PyTorch-first library for deterministic, variational, and quantized autoencoder-family models with one consistent config, module, trainer, and checkpoint surface.

The same distribution also includes autoindexers, a sibling package for classical indexing and hashing algorithms that consume full embedding tables.

What this documentation covers

  • how to build models from sample_spec
  • how dataset, encoder, core, and decoder specs fit together
  • how to train from one YAML-first entrypoint
  • how different model families extend the shared base configs
  • how sibling embedding-table indexers fit next to learned latent models

Core ideas

  • Datasets expose canonical DataSpec objects.
  • Backbones are explicit and are built from those specs.
  • Models distinguish:
  • sample_spec
  • encoder_output_spec
  • core_spec
  • decoder: null is intentionally strict and only works when reversing the encoder produces a decoder whose runtime input spec matches the model's declared decoder input spec.