boltzmann9 0.1.4

Restricted Boltzmann Machine implementation in PyTorch

BoltzmaNN9

BoltzmaNN9 is a small toolkit for training Restricted Boltzmann Machines (RBMs) with a block-aware architecture, optional preprocessing, and a simple CLI for managing experiments.

Installation

• Requires Python 3.10+ and PyTorch (CPU or CUDA).
• From the repo root: python -m venv .venv && .venv/Scripts/Activate.ps1 (PowerShell) then pip install -e . or pip install -e ".[dev]" for linting/test extras.

CLI commands

• python boltzmann9.py new_project <path> – scaffold a project with config.py, data/, output/, and a synthetic data generator.
• python boltzmann9.py preprocess_raw --config <config.py> – run DataPreprocessor to turn raw CSV columns into binary features; relative paths in the config are resolved against the config file location.
• python boltzmann9.py train --project <project_dir> – train an RBM using the project’s config.py; creates a timestamped run under output/.
• python boltzmann9.py evaluate --run <run_dir or project_dir> – evaluate a saved run (defaults to the latest run if a project path is given).
• python boltzmann9.py list --project <project_dir> – list available runs.

Configuration basics

config.py is a plain Python dict (see src/templates/config.py for a template):

• device: "auto"/cpu/cuda:0/mps.
• data: csv_path (training data), optional drop_cols.
• model: visible/hidden block sizes, cross-block restrictions, initialization.
• preprocess: quantile bounds, category limits, missing-bit toggles.
• dataloader / train / eval / conditional: loader sizes, training hyperparameters, evaluation options. All relative paths in preprocessing are resolved relative to the config file; training also adjusts data.csv_path relative to the project directory.

Project template

python boltzmann9.py new_project demo_project creates:

• demo_project/config.py – prefilled with the data path data/data.csv.
• demo_project/data/synthetic_generator.py – generates a demo dataset and plot.
• demo_project/output/ – destination for runs (run_<timestamp>).

Development

• Run tests (if added): python -m pytest.
• Lint (if installed): ruff check . and black --check ..
• Key sources: boltzmann9.py (CLI), src/boltzmann/model.py (RBM), src/boltzmann/preprocessor.py (raw → binary), src/boltzmann/pipeline.py (training/eval flow).

See docs/QUICKSTART.md for a step-by-step guide and docs/TECHNICAL_DOCUMENTATION.md for module-level details.