Skip to content

manylatents

Dimensionality reduction and neural network analysis. Built on PyTorch Lightning and Hydra.

Quick Start

git clone https://github.com/latent-reasoning-works/manylatents.git
cd manylatents && uv sync

uv run python -m manylatents.main data=swissroll algorithms/latent=pca

Documentation

  • Algorithms — LatentModule (fit/transform) and LightningModule (trainable) algorithms, networks, and losses
  • Data — Synthetic manifolds, precomputed data, and sampling strategies
  • Metrics — Three-level evaluation system: embedding, dataset, and module metrics
  • Evaluation — Algorithm dispatch, sampling strategies, and shared caching
  • Cache Protocol — Shared kNN cache, config sleuther, metric expansion
  • Callbacks — Embedding callbacks (save, plot, wandb) and trainer callbacks (probing)
  • Extensions — Install, use, and develop domain extensions
  • API — Programmatic API for agent-driven multi-step workflows
  • Probing — Representation probing for auditing algorithm internals during training
  • Testing — CI pipeline, namespace integration testing, and mock package patterns