v1.0.0rc4

Full Changelog: v1.0.0rc3...v1.0.0rc4

v1.0.0rc3

torchfits 1.0.0rc3 — collaborator soak

Please install, try the examples, report bugs, and contribute.

Highlights

• Examples: Learn Astropy parity (M13 stack, MEF header, Spitzer cutout), real Lupton RGB (SDSS g/r/i), MaNGA LOGCUBE, CFHT MegaCam MEF cutouts, custom transform + make_loader vs DataLoader
• Docs: API button on landing, cache how/when, Dataset/loader tables, custom transform how-to, persona pass; fixed escaped pipes in API tables
• Jules: AGENTS.md + JULES.md — one weekly researched bug/perf pass (delete old cosmetic schedules in the Jules UI and paste the prompt from JULES.md)
• Fixes: hdu="EXTNAME" on read_tensor/read_subset, 3D + zero-size cutouts, table where= + TNULL, remote prefetch race, Lupton empty bands, clear .bz2 error

Benchmarks

• Host scorecard: MPS exhaustive_mps_20260718_180230; Linux CPU/CUDA still Jul-17 soak (CANFAR Jul-18 sessions still Pending)
• ML loader: ml_20260718_191908
• MegaCam: 20260718_124403

Docs edge: https://astroai.github.io/torchfits/edge/

v1.0.0rc2

Install

pip install torchfits==1.0.0rc2

Requires Python 3.10+ and PyTorch 2.10. Docs: https://astroai.github.io/torchfits/

Highlights

• Concurrent FITS reads open a private CFITSIO handle per call (safe under multi-worker DataLoaders and multi-HDU access).
• Dataset peers: FitsTensorDataset, FitsImageDataset, FitsCubeDataset, FitsSpectrumDataset; configurable TORCHFITS_CACHE_DIR / cache_dir=.
• CLI short options, probe SSRF hardening (all DNS records + redirect hops), convert/table filters.
• Contiguous table/image writes, header integer overflow checks, empty-primary MEF compress fix.

Breaking / notes

• Cross-thread shared fitsfile* handle LRUs are gone (correctness). Metadata/pread caching remains.
• SemVer 1.0.0 is not this tag — soak on the rc line first.

Links

• Changelog: https://astroai.github.io/torchfits/changelog/
• Compare: v1.0.0rc1...v1.0.0rc2

torchfits 1.0.0rc1

Install

pip install torchfits==1.0.0rc1

Requires Python 3.10+ and PyTorch 2.10. See the compatibility matrix.

Highlights

• Release candidate for the 1.0 API (formal stamp in docs/reviews/release-api-freeze-1.0.0rc1.md)
• Empty table filters preserve schema; stretch a > 0 validation; verify treats missing checksums as OK (exit 0)
• CLI/docs polish: convert --to inference, CFITSIO image sections on cutout, --format text|json|jsonl
• Root read_table / stream_table / read_table_rows emit DeprecationWarning — prefer torchfits.table.*

Breaking / behavior change

• torchfits verify on files without checksum keywords now exits 0 (was 4). Use status == "fail" / exit 4 for corruption.

Links

• Changelog: https://github.com/astroai/torchfits/blob/v1.0.0rc1/docs/changelog.md
• Compare: v1.0b1...v1.0.0rc1

Wheels publish via trusted publishing when build_wheels.yml completes.

What's Changed

• release: v1.0b1 by @sfabbro in #217

Full Changelog: v1.0b1...v1.0.0rc1

v1.0b1

What's Changed

• Docs gallery: KaTeX, transform viz, real data, CLI recipes by @sfabbro in #215

Full Changelog: v0.9.3...v1.0b1

torchfits 0.9.3

Install

pip install torchfits==0.9.3

Requires Python 3.10+ and PyTorch 2.10. Pre-built wheels for Linux x86_64 and macOS arm64.

Docs: https://astroai.github.io/torchfits/
Changelog: https://astroai.github.io/torchfits/changelog/

Highlights

Convert more catalog formats

torchfits convert catalog.fits out.parquet --to parquet --hdu 1
torchfits convert catalog.fits out.csv --to csv --hdu 1
torchfits convert catalog.fits out.tsv --to tsv --hdu 1
torchfits convert catalog.fits out.arrow --to arrow --hdu 1
torchfits convert r.fits g.fits b.fits rgb.png --to png

CSV/TSV export flat columns; nested columns use parquet or arrow. PNG is a Lupton RGB preview (stdlib, no Pillow). --to ppm is removed.

Header fitsort + probe

torchfits header *.fits --fitsort --keyword OBJECT --keyword DATE-OBS
torchfits probe science.fits --json

Optional vos: / vos:// probe when the vos package is installed.

Transforms package

Same public API, split by domain:

from torchfits.transforms import ArcsinhStretch, Compose, ZScaleNormalize

Notebooks

Lean _repr_html_ previews on TensorHDU, TableHDU, and TableHDURef.

Performance

Scorecard refresh: CUDA 0 strict deficits; Linux CPU 1; Mac MPS 16. Methodology: Benchmarks.

Security

Filenames using CFITSIO’s sh:// protocol are rejected (same class of checks as | pipes).

Breaking changes

Before	After
torchfits convert … --to ppm	--to png

Links

• Pull request
• Full changelog
• Compare: v0.9.2...v0.9.3

What's Changed

• 🛡️ Sentinel: [CRITICAL] Fix command injection vulnerability via CFITSIO sh:// protocol by @sfabbro in #207
• 0.9.3: fitsort header tables and vos probe by @sfabbro in #212

Full Changelog: v0.9.2...v0.9.3

torchfits 0.9.2

Install

pip install torchfits==0.9.2

Requires Python 3.10+ and PyTorch 2.10. Pre-built wheels for Linux x86_64 and macOS arm64.

Docs: https://astroai.github.io/torchfits/
Changelog: https://astroai.github.io/torchfits/changelog/

Highlights

Command-line tools

pip install provides a torchfits CLI for multi-extension FITS (MEF) workflows:

torchfits info science.fits
torchfits header *.fits --keyword OBJECT --json
torchfits verify science.fits
torchfits stats science.fits --hdu 0 --jsonl
torchfits table catalog.fits --hdu 1 --preview 5
torchfits cutout science.fits cutout.fits --box 100,100,256,256

Full command list and exit codes: CLI guide.

Clearer Python imports

• Root imports are FITS I/O and HDU helpers only.
• Transforms import from torchfits.transforms (for example from torchfits.transforms import ZScaleNormalize).
• read_fast and read_image are removed — use read / read_tensor.

Performance

Lab scorecard refresh: Linux CPU and CUDA report 0 strict-gate deficits; Mac MPS reports 4. Methodology and tables: Benchmarks.

Breaking changes

Before	After
from torchfits import ZScaleNormalize	from torchfits.transforms import ZScaleNormalize
torchfits.read_fast(...) / read_image(...)	torchfits.read(...) / read_tensor(...)

Links

• Pull request
• Full changelog
• Compare: v0.9.1...v0.9.2

What's Changed

• Zero-deficit CANFAR scorecard + engine fixes by @sfabbro in #210
• Release 0.9.2: lean API, CLI, scorecard by @sfabbro in #211

Full Changelog: v0.9.1...v0.9.2

v0.9.1

What's Changed

• Release 0.9.1 with truthful PyTorch ABI metadata by @sfabbro in #204

Full Changelog: v0.9.0...v0.9.1

v0.9.0

Full Changelog: v0.7.0...v0.9.0

v0.7.0

Full Changelog: v0.6.0...v0.7.0