scitomo/deepdewedge-tutorial

DeepDeWedge tutorial checkpoint — fresh Scitomo FORMAT 2 package

This is a fresh FORMAT 2 export from the authoritative original Lightning checkpoint, not a migration of any earlier Hugging Face package. Normal runtime uses Scitomo's generic FORMAT 2 loader and Safetensors only; it does not require PyTorch Lightning or the upstream DeepDeWedge source checkout.

Package identity

• package id: deepdewedge_tutorial; package revision: 3
• learned-checkpoint format: 2; manifest schema: 4
• Scitomo conversion checkout: 2832957f69daff0d7baec5df17a7c54954623eed
• minimum Scitomo version: 0.7.3
• previous Hugging Face commit: 87db06570dd874a99af1289e62b79ea99f87f006 — HISTORICAL ONLY; NOT CONVERSION INPUT

Authoritative provenance

• upstream repository: https://github.com/MLI-lab/DeepDeWedge
• upstream revision: 072075692a44a8f17394214369e6e762abe52bc3
• Figshare DOI: https://doi.org/10.6084/m9.figshare.25043435.v1; file id: 45582309
• original archive SHA-256: 7c871342e51f5a66a773fe427d72944b5d2cc8ff41c5b7415ab38dbfc9ac6d58
• original checkpoint member: tutorial_data/fitted_model.ckpt
• original checkpoint size: 327952642 bytes
• original checkpoint SHA-256: 5262f6c11e85fd662b02e59efe936fa7b69913758e841235be2683f7bd03ec76

DeepDeWedge Tutorial Data is attributed to Simon Wiedemann and is distributed under CC BY 4.0. The pinned DeepDeWedge implementation is BSD-2-Clause; its license text is included below LICENSES/. See ATTRIBUTION.md.

Scientific inference semantics

The pure persisted Network owns only the lowered U-Net architecture and its 54 canonical tensors. The fitted affine values remain outside Network state in the typed deepdewedge_inference profile:

• network_affine_loc: -0.14898751676082611
• network_affine_scale: 1.3237642049789429
• input layout: (..., Z, Y, X); Network layout: (..., C, Z, Y, X)
• paired halves are refined independently then averaged; full-width missing wedge: 50 degrees
• 96³ patches, 32³ overlap, trailing-reflection coverage, linear-ramp reassembly
• preconditioning recomputes patch statistics; output uses the checkpoint-fitted affine

Fresh conversion and validation

refresh_format2.py is the exact one-off implementation and records the verified source, explicit 54-tensor mapping, strict Network lowering, and generic export. It was run with Python 3.12.13, Torch 2.12.1+cpu, Lightning 2.6.5, Safetensors 0.8.0, and Scitomo 0.7.3.dev0 on Windows-11-10.0.22631-SP0.

The generic exporter freshly serializes weights.safetensors; no previous Hugging Face Safetensors, manifest, construction, or inference record is read. The conversion record lists every source checkpoint tensor to canonical target mapping. The validation record binds package state closure, generic loader reload, external-affine semantics, and deterministic forward parity.

For a deterministic directional, non-symmetric CPU float32 input of 4,096 elements, authoritative upstream output versus FORMAT 2 pure-Network-plus-profile output passed rtol=1e-5, atol=1e-6: maximum absolute error 0, relative L2 error 0.

Files and closure

manifest.json is the authoritative, closed inventory of every package file, with each fresh size and SHA-256. It declares only FORMAT 2 construction, inference, Safetensors, conversion, validation, and documentation/license resources; there is no format-1 or migration artifact. Validate and load with:

import scitomo as st
loaded = st.api.load_learned_network("/path/to/package")

This operation uses the generic Scitomo FORMAT 2 loader and does not import Lightning or DeepDeWedge. It is a checkpoint package, not a claim of scientific approval for a new dataset or acquisition protocol.