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
DataSpecobjects. - Backbones are explicit and are built from those specs.
- Models distinguish:
sample_specencoder_output_speccore_specdecoder: nullis intentionally strict and only works when reversing the encoder produces a decoder whose runtime input spec matches the model's declared decoder input spec.