autoencoders is a PyTorch-first library for autoencoder-family models across deterministic, variational, and quantized latent spaces.

The distribution now also ships a sibling package, autoindexers, for indexing algorithms that operate directly on full embedding tables.

The project goal is simple: make autoencoders feel composable, serializable, and reusable in the same way transformers did for sequence models.

Why autoencoders

What It Covers

Current model families include:

• Deterministic models: AE, DAE, CAE, SAE, TopKSAE, KLSAE, WAE, AAE
• Variational models: VAE, DVAE, BetaVAE, BetaTCVAE, DIPVAE, InfoVAE, MMDVAE, FactorVAE, VampPriorVAE, HVAE
• Quantized models: VQVAE, GumbelVQ, FSQ, RFSQ, PQVAE, RQVAE, VQVAE2

Core interfaces include:

• Config + Model + Output + Export
• save_pretrained() / from_pretrained()
• encode() / decode() / reconstruct() / export()
• family-specific trainers for deterministic, variational, and quantized models

Bundled indexing interfaces include:

• BaseIndexer + Config + QueryResult
• build() / query() / save_pretrained() / from_pretrained()
• non-neural embedding-table indexers such as LSH, SimHash, PCA Hashing, and ITQ

At a Glance

Installation

Install the package:

pip install autoencoders

This single package install exposes both import namespaces:

• autoencoders
• autoindexers

Install with PyTorch dependencies:

pip install "autoencoders[torch]"

Install with encoder-backed text dataset support:

pip install "autoencoders[text]"

Install with CLIP-backed multimodal dataset support:

pip install "autoencoders[clip]"

Install everything commonly needed for experiments:

pip install "autoencoders[all]"

If you are working from source and plan to build or publish packages:

pip install "autoencoders[dev]"

Documentation

The repository now ships with an MkDocs site that documents:

• the dataset, backbone, and DataSpec surface
• the unified YAML training entrypoint
• tree-structured model parameter references for deterministic, variational, and quantized families
• sibling embedding-table indexers in autoindexers

Preview the docs locally:

mkdocs serve

Build the static site:

mkdocs build --strict

Quick Start

Build a basic AE + MLP model explicitly from a sample spec:

import torch

from autoencoders import AutoencoderConfig, AutoencoderModel
from autoencoders.data.base import TensorSpec

model = AutoencoderModel(
    config=AutoencoderConfig(latent_dim=16),
    sample_spec=TensorSpec(shape=(50,)),
    encoder="mlp",
    encoder_config={"hidden_dims": [64, 32], "activation": "relu", "use_bias": True},
    decoder="mlp",
    decoder_config={"hidden_dims": [64, 50], "activation": "relu", "use_bias": True},
)

inputs = torch.randn(32, 50)
outputs = model(inputs)

print(outputs.loss)
print(outputs.latents.shape)
print(outputs.reconstruction.shape)

Save and load checkpoints:

model.save_pretrained("artifacts/ae")
restored = AutoencoderModel.from_pretrained("artifacts/ae")

Inspect the model pipeline and layer-by-layer shape trace:

for step in model.get_pipeline_trace():
    print(step.name, "->", step.output_spec)

Train from YAML:

python examples/trainer.py --config examples/configs/glove/ae.yaml --epoch 5

Build an index directly from an embedding table:

import torch

from autoencoders.data.base import TensorSpec
from autoindexers import load_indexer

embeddings = torch.randn(1000, 128)
indexer = load_indexer(
    "simhash",
    sample_spec=TensorSpec(shape=(128,)),
    num_bits=64,
).build(embeddings)

results = indexer.query(embeddings[0], top_k=5)
print(results[0].item_ids)

Product Surface

Use the package at three different layers:

• Model layer: build or load latent models with typed configs
• Training layer: train deterministic, variational, or quantized families with dedicated trainers
• Experiment layer: run reusable YAML configs with one trainer entrypoint on real datasets
• Indexing layer: build classical embedding-table indexers with autoindexers

Model Loading

Load a model dynamically by name while still keeping backbone selection explicit:

from autoencoders import load_model
from autoencoders.data.base import TensorSpec

model = load_model(
    "vae",
    sample_spec=TensorSpec(shape=(50,)),
    latent_dim=16,
    kl_weight=0.1,
    free_bits=0.02,
    encoder="mlp",
    encoder_config={"hidden_dims": [64, 32], "activation": "relu", "use_bias": True},
    decoder="mlp",
    decoder_config={"hidden_dims": [64, 50], "activation": "relu", "use_bias": True},
)

Datasets

The library currently ships with embedding-first datasets plus one image dataset for CNN- and ViT-backed experiments:

• glove
• fasttext
• numberbatch
• snli
• multinli
• flickr30k
• cifar10

Load a dataset directly:

from autoencoders import load_dataset

dataset = load_dataset("glove", dim=50, max_vectors=50000)
loaders = dataset.get_dataloaders(batch_size=256)

Encoder-backed sentence datasets materialize embeddings during prepare() and cache the result just like static embedding tables:

dataset = load_dataset(
    "snli",
    encoder_name="sentence-transformers/all-MiniLM-L6-v2",
    max_vectors=50000,
)
loaders = dataset.get_dataloaders(batch_size=256)

CLIP-backed multimodal datasets follow the same cached artifact pattern:

dataset = load_dataset(
    "flickr30k",
    encoder_name="ViT-B-32",
    encoder_pretrained="laion2b_s34b_b79k",
    modality="both",
    max_vectors=50000,
)
loaders = dataset.get_dataloaders(batch_size=256)

Image data uses H x W x C specs end to end:

dataset = load_dataset("cifar10", max_examples=10000)
print(dataset.get_sample_spec())  # TensorSpec(shape=(32, 32, 3))

Backbone Semantics

Backbones are configured explicitly and built from the dataset-driven sample_spec.

• MLPModule consumes tensor specs whose last dimension is the feature width.
• CNNModule consumes image-like TensorSpec(shape=(H, W, C)) values and handles HWC <-> NCHW conversion internally.
• VisionTransformerModule also consumes image-like TensorSpec(shape=(H, W, C)), patchifies them internally, and exposes sequence-shaped latent specs.

Auto-inferred decoders are intentionally strict:

• decoder: null is supported only when reversing the encoder produces a decoder whose runtime input spec matches the model's decoder input spec.
• Models whose decoder space differs from encoder output space, such as hierarchical or latent-shape-changing variants, must provide an explicit decoder config.

For explicit image decoders, use transpose: true when you want an upsampling transposed-convolution stack:

decoder:
  name: cnn
  config:
    channels: [64, 3]
    kernel_sizes: [4, 4]
    strides: [2, 2]
    paddings: [1, 1]
    activation: relu
    use_bias: true
    transpose: true

Downloaded datasets use a global cache:

• default: ~/.cache/autoencoders
• override with: AUTOENCODERS_CACHE=/your/cache/path

This makes the package useful both as:

• a standalone training library
• a latent-model subsystem inside larger PyTorch projects

Training API

Deterministic training:

from autoencoders import AETrainer, TrainingConfig

trainer = AETrainer(
    model=model,
    args=TrainingConfig(
        output_dir="artifacts/ae-run",
        epochs=5,
        batch_size=256,
    ),
)

trainer.fit(loaders, metadata={"dataset": "glove", "model": "ae"})

Variational training:

from autoencoders import VAETrainer, VariationalAutoencoderConfig, VariationalAutoencoderModel
from autoencoders.data.base import TensorSpec

trainer = VAETrainer(
    model=VariationalAutoencoderModel(
        config=VariationalAutoencoderConfig(
            latent_dim=16,
            kl_weight=0.1,
            free_bits=0.02,
            kl_warmup_epochs=20,
        ),
        sample_spec=TensorSpec(shape=(50,)),
        encoder="mlp",
        encoder_config={"hidden_dims": [64, 32], "activation": "relu", "use_bias": True},
        decoder="mlp",
        decoder_config={"hidden_dims": [64, 50], "activation": "relu", "use_bias": True},
    ),
    args=TrainingConfig(output_dir="artifacts/vae-run", epochs=10),
)

Quantized training:

from autoencoders import VQTrainer, TrainingConfig, load_model
from autoencoders.data.base import TensorSpec

trainer = VQTrainer(
    model=load_model(
        "rqvae",
        sample_spec=TensorSpec(shape=(None, 50)),
        latent_dim=16,
        codebook_size=256,
        num_quantizers=4,
        use_ema_codebook=True,
        dead_code_reset=True,
        encoder="mlp",
        encoder_config={"hidden_dims": [64, 32], "activation": "relu", "use_bias": True},
        decoder="mlp",
        decoder_config={"hidden_dims": [64, 50], "activation": "relu", "use_bias": True},
    ),
    args=TrainingConfig(output_dir="artifacts/rqvae-run", epochs=10),
)

Training Entry Point

Source checkouts now use one unified YAML-driven entrypoint:

• examples/trainer.py

The legacy examples/train_ae.py wrapper still forwards into the same code path for basic AE runs.

Useful examples:

python examples/trainer.py --config examples/configs/glove/ae.yaml --epoch 5
python examples/trainer.py --config examples/configs/glove/vae.yaml --epoch 5
python examples/trainer.py --config examples/configs/glove/vqvae.yaml --epoch 5
python examples/trainer.py --config examples/configs/cifar10/vqvae.yaml --epoch 5
python examples/trainer.py --config examples/configs/cifar10/vqvae_vit.yaml --epoch 5

Each config is organized into five sections:

• dataset
• model
• encoder
• decoder
• trainer

Each section uses name + config form except trainer, which is a flat config block. Runtime overrides such as --epoch 5, --lr 0.001, or --max_vectors 5000 resolve into ${...:default}$ placeholders inside the YAML files before training starts.

Launch-Ready Features

• 🗃️ Checkpoints: save_pretrained() and from_pretrained()
• 📤 Exports: standardized latent artifact export across model families
• 📚 Real datasets: static embedding tables, sentence corpora, and CLIP-backed image-text corpora
• 🎛️ Family-specific trainers: deterministic, variational, quantized, and adversarial flows
• 🧪 Packaging: buildable sdist and wheel, ready for PyPI publication

Design Direction

The library is organized around latent model families rather than a single monolithic interface:

• BaseAutoencoderModel
• BaseVariationalAutoencoderModel
• BaseVectorQuantizedAutoencoderModel

Matching outputs are also family-specific:

• BaseAutoencoderOutput
• VariationalAutoencoderOutput
• QuantizedAutoencoderOutput

This keeps the shared API stable without flattening away meaningful model differences such as posterior statistics or codebook indices.

Current Scope

autoencoders is intentionally embedding-first, with a growing image path for CNN-backed quantized models. The current core is aimed at:

• representation learning on embedding matrices
• latent compression
• variational latent modeling
• quantized latent tokenization

Future raw-modality frontends and multimodal adapters can be layered on top of this core.

Repository Status

This project is still early, but the current package already supports:

• trainable deterministic, variational, and quantized autoencoder families
• reusable checkpoints
• exportable latent artifacts
• real embedding datasets with download and cache support
• package metadata and distribution artifacts ready for publication workflows

Development

Build the package locally:

python -m build

Check the generated distribution:

twine check dist/*