nervecode 0.1.0

Intrinsic surprise scoring for PyTorch via statistical coding.

Nervecode Docs

This directory will contain documentation for the Nervecode project.

Developer Setup

• Install dev tools (in your virtualenv): pip install -e .[dev]
• Install git hooks: pre-commit install
• Run on all files once: pre-commit run --all-files

Included hooks:

• Formatting: ruff-format
• Linting and import sorting: ruff (with --fix)
• Type-checking: mypy

Tests

• Unit tests: pytest tests/unit (fast correctness checks)
• Integration tests: pytest tests/integration (heavier, cross-module)
• Smoke tests: pytest tests/smoke (repo wiring and API surface)

Deterministic runs

• Tests auto-seed RNGs per test using NERVECODE_SEED (default: 1234).
• You can control reproducibility in your own scripts via:

  from nervecode.utils.seed import seed_everything, seed_from_env
  seed_everything(seed_from_env())
  

Performance Notes

• Overhead estimates and guidance for pooled Conv2d coding: see docs/overhead.md.
• Scaling and tradeoffs for layer selection and coding dimension: see docs/scaling.md.