PyAutoNerves

PyAutoNerves is the Nerves of the PyAutoScientist — the configuration, serialization, and I/O foundation (package autonerves) connecting the workspace's conventions to every library. It provides a layered configuration system with workspace overrides, dict / JSON / CSV serialization of arbitrary objects, and FITS I/O.

PyAutoFit, PyAutoArray, PyAutoGalaxy, and PyAutoLens all depend on autonerves: it supplies their packaged default config, the object-serialization used to persist models and results, and shared utilities (test_mode, jax_wrapper). Centralising these here keeps a single, consistent config and I/O layer beneath every library. As a released library it ships with every nightly train — see the PyAutoHands release board for the current version.

Install

pip install autonerves

Examples

Layered config — read a directory of YAML into a queryable Config:

from autonerves.conf import Config

config = Config("path/to/config")          # directory of YAML files
value = config["general"]["model"]["section"]["value"]

JSON serialization — round-trip arbitrary Python objects:

from autonerves.dictable import output_to_json, from_json

data = {"sersic_index": 4.0, "centre": [0.0, 0.0]}
output_to_json(data, "model.json")
restored = from_json("model.json")         # == data

FITS I/O — write and read a NumPy array:

import numpy as np
from autonerves.fitsable import output_to_fits, ndarray_via_fits_from

arr = np.arange(12.0).reshape(3, 4)
output_to_fits(values=arr, file_path="demo.fits", overwrite=True)
loaded = ndarray_via_fits_from(file_path="demo.fits", hdu=0)   # np.allclose(arr, loaded)

Nerves board

https://pyautolabs.github.io/PyAutoNerves/ — a read-only browser of every config file and option across the organism: each library's <package>/config/ (PyAutoFit, PyAutoArray, PyAutoGalaxy, PyAutoLens, PyAutoCTI) and each workspace's config/ that overrides them.

• Per repo (repos/<Repo>.html): every YAML file with its top-level keys, its source line-numbered with the comments kept (the comments are the option docs) and a GitHub link; prior files render as a table (Class · param · type · mean/σ or bounds · width modifier · limits).
• Search: the index page's box searches every key, file and comment across all repos and jumps to the line.
• Override map: autonerves resolves a key workspace → last-imported library → … → PyAutoFit (Config.push(keep_first=True), keys lowercased), so each workspace file is compared, by relative path, against the same file across its library stack in that order (lens → galaxy → array → fit; cti → array → fit). Keys whose value differs, keys no library defines (orphans) and keys that fall through to the libraries are listed per file. build/*.yaml is workspace tooling and is grouped apart.
• Possibly unused config keys: each library's Python (and autonerves' own) is scanned for conf.instance[...] lookups, and every key of a library settings file is classed used (a lookup reads it, or reads a key under it — PyAutoLens reading a PyAutoFit key counts, since autonerves merges every layer), section-read (a lookup reads an ancestor section whole, or a non-literal subscript such as ["plots"][section][name] ends the chain above it) or unused (nothing references it). Unused keys get a chip on the repo page, per-file counts and an index section grouped by library with GitHub links. Prior files are exempt (looked up by class name). Limits: the scan is static — it follows literal subscript chains (including multi-line ones and .get("k")), sections bound to a local and indexed later in the same function or a closure, and the should_output helper; a key read any other way (built key strings, getattr, a section passed to another function and indexed there) shows as section-read at best and may show as unused. Treat the list as candidates, not a verdict.
• Environment variables: the PYAUTO_* switches autonerves reads.
• Cockpit feed (state.json): green when every file parses and no workspace key is orphaned; yellow with one item per unparseable file or orphan-carrying workspace file; grey when nothing was collected. Never red. Possibly unused keys add one info item per library — never yellow, until the scan is trusted.

Rendered daily by .github/workflows/nerves_board.yml from sparse checkouts of the config folders (plus each library's package .py files, for the lookup scan); the renderer is scripts/board.py (not part of the autonerves package). Nothing on the board edits config.

Links

• Source & tests: autonerves/, test_autonerves/
• Agent/contributor instructions: AGENTS.md
• The organism this repo is the Nerves of: PyAutoBrain/ORGANISM.md, documented in full at https://pyautoscientist.readthedocs.io
• Ecosystem: PyAutoLabs on GitHub

JAX compatibility

The package retains JAX/JAXlib >=0.7,<0.12, excluding 0.10.* and 0.11.0. Those excluded releases contain CPU batched LAPACK scheduling that can deadlock while materializing a vectorized likelihood (PyAutoHeart#274). The failure was captured on 0.10.2; tagged source identifies the same path in the other excluded releases. Open-source builds disable it from 0.11.1. This restriction does not replace the separate historical FFT/Eigen workaround.

The compatibility diagnostics exercise 0.9.2 as an older endpoint and 0.11.2 as the recommended validation baseline. They do not certify every version permitted by the retained range or promise identical performance across versions. A new minimum of 0.11.2 is deliberately avoided: it would unnecessarily force newer NumPy/SciPy requirements on users with working older environments.

Updated family packages require autonerves>2026.10.4.1 so their resolver cannot fall back to a Nerves release without the exclusions. This is a release ordering requirement: publish a Nerves wheel containing this policy before the corresponding family wheels. No release version is chosen by this bound, and existing published wheels are not retroactively repaired. An unpinned family package can still be backtracked to an entirely older family release; select the repaired family release explicitly when applying the fix to a constrained environment.

Keep JAX, JAXlib and any CUDA plugin on a mutually compatible version. Use pip's resolver rather than --no-deps, and verify the resulting environment with python -m pip check. Existing lockfiles need regeneration to adopt the policy.