scitomo/deepdewedge-tutorial

DeepDeWedge tutorial checkpoint for scitomo

This repository contains the official DeepDeWedge tutorial checkpoint converted to a scitomo-native, non-executable Safetensors package. It reinstantiates the scitomo UNet3D used by the deepdewedge restoration method. It is not a PyTorch Lightning trainer-resume checkpoint and does not contain optimizer, scheduler, callback, random-number-generator, or dataloader state.

Package identity:

• scitomo catalog name: deepdewedge_tutorial
• native package id: deepdewedge_tutorial
• package revision: 1
• minimum scitomo version: 0.5.4
• learned-checkpoint format: 1
• method: restoration / deepdewedge
• construction fingerprint: sha256:c9da7decd489f0f4f893dce565f37db25abc63d0d3efcacad9835305ffd97cb2
• inference fingerprint: sha256:c2c5eba72579510265258dcfce3a5be2f851f9276c7e53d92330d78eb37720dd

Source and attribution

The original checkpoint is part of DeepDeWedge Tutorial Data, authored by Simon Wiedemann and published on Figshare under CC BY 4.0:

• DOI: https://doi.org/10.6084/m9.figshare.25043435.v1
• Figshare file id: 45582309
• archive: tutorial_data.zip
• archive member: tutorial_data/fitted_model.ckpt
• archive SHA-256: 7c871342e51f5a66a773fe427d72944b5d2cc8ff41c5b7415ab38dbfc9ac6d58
• original checkpoint SHA-256: 5262f6c11e85fd662b02e59efe936fa7b69913758e841235be2683f7bd03ec76

The method and upstream implementation are described by:

Simon Wiedemann and Reinhard Heckel. “A deep learning method for simultaneous denoising and missing wedge reconstruction in cryogenic electron tomography.” Nature Communications 15, 8255 (2024). https://doi.org/10.1038/s41467-024-51438-y

Upstream code: https://github.com/MLI-lab/DeepDeWedge/tree/072075692a44a8f17394214369e6e762abe52bc3

The conversion changed the serialization, state names, and executable construction mechanism. It did not intentionally alter tensor values. Two learned normalization values that upstream stored as non-trainable parameters are native scitomo buffers with identical values. See ATTRIBUTION.md, conversion/conversion-record.json, and LICENSES/.

Native construction and inference

construction.json specifies a residual single-input/single-output 3D U-Net with 64 initial channels, three downsampling blocks, strided-convolution downsampling, transposed-convolution upsampling, leaky-ReLU activations, and the learned input-normalization location and scale from the tutorial checkpoint.

Volume tensors use scitomo's canonical trailing-axis layout (..., Z, Y, X). The network layout is (..., C, Z, Y, X), with one scalar volume channel. The canonical rotation axis is Z; at theta zero the beam axis is Y, and detector (V, U) corresponds to (Z, X).

The frozen DeepDeWedge inference profile requires:

• paired half-tomograms, refined separately and averaged;
• a 50-degree full-width missing-wedge Fourier mask on each half;
• 96 x 96 x 96 patches with overlap 32 x 32 x 32;
• trailing reflection padding for full coverage;
• patch-statistic normalization and network denormalization;
• linear-ramp weighted patch reassembly; and
• no full-tomogram standardization.

The package was converted from the tutorial's fitted network. It assumes the same scientific meaning, preprocessing, normalization, missing-wedge convention, and paired-half workflow. It is not a general-purpose cryo-ET foundation model.

Files and identities

File	Bytes	SHA-256
construction.json	2,175	d4c7eced057042438827de168d47b7900311fba742ce52445675be7f40343432
inference.json	968	217636919d17d9332aa49474e893064ed1dfb29f016a468cf22d439cf16887a0
weights.safetensors	109,294,940	2a34aeb61dba5da5c7b78f4cc26de9b8a9134ac12dc08876a8e93bf02776c795
conversion/conversion-record.json	13,953	e55c35f9e233e0ba04ea1332306b179b96db12158b4eb0cd8d5dcd3609ac8e54
validation/validation-record.json	1,946	eb2f9153d49e7124a8aeb1951230b6124f5fdc18ea1f53b2c9e28db55273cc3f

manifest.json binds these files plus this model card, attribution, and license resources by exact size and SHA-256. The immutable Hugging Face commit and scitomo learned-weight catalog bind the complete distribution, including the manifest and documentation resources, without a self-referential checksum inside this README.

Validation

All 56 source tensors were mapped one-to-one and exactly matched after native assignment.

Predetermined CPU float32 checks:

Case	Tolerance	Result
Synthetic forward parity	atol=1e-6, rtol=1e-5	bit-exact; max absolute error 0; relative L2 0
Real tutorial-volume crop	relative L2 <=1e-4	max absolute error 7.152557373046875e-7; relative L2 1.2612566990810592e-7

The real-data case used a centered 32 x 32 x 32 crop from tutorial_data/tomo_even_frames.rec. Vendor and native outputs were finite, had identical shapes, and had the same absolute-peak spatial landmark at (Z, Y, X) = (25, 13, 0).

The evidence proves native network-state and reviewed forward parity for the frozen inputs. It does not establish accuracy on every microscope, specimen, acquisition protocol, missing-wedge angle, or preprocessing pipeline, nor does it replace validation of a complete user workflow.

Safe loading

The repository contains declared data files only. Loading does not execute remote code, import the vendor project, or use Python pickle. scitomo requires the exact catalog commit and verifies every declared size and SHA-256 before opening weights.safetensors. Hugging Face trust_remote_code is never used.

Install the learned and catalog extras before resolving the catalog package:

pip install "scitomo[learned,catalog]"

Normal runtime loading is owned by scitomo's central learned-checkpoint API. Do not load the original Lightning checkpoint in an ordinary runtime.

Licenses

• Converted weights and their source tutorial dataset: CC BY 4.0. See LICENSES/DeepDeWedge-Tutorial-Data-CC-BY-4.0.txt.
• Upstream DeepDeWedge code and behavior used for construction/conversion: BSD 2-Clause. See LICENSES/DeepDeWedge-Code-BSD-2-Clause.txt.

The CC BY 4.0 attribution and modification notice are provided in ATTRIBUTION.md. No endorsement by the original authors or rights holders is implied.