tbhugging/gummybear-tomography

GummyBear Tomography Dataset

Synthetic optical tomography corpora and reproducibility artefacts for the gummybear-tomography project.

This Hub package sits beside the Git repository layout. Generated corpora are stored as zip archives so upload/download stays tractable (many small files). The Dataset Viewer does not need those zips extracted: preview JPEG/PNG bytes are embedded in the parquet Image columns.

Hub layout:

• data/generated/m8_1.zip — full single-particle optical corpus (Milestone 8/9)
• data/generated/m10_illumination.zip — multi-illumination corpus (Milestone 10)
• data/huggingface_metadata/ — parquet indexes + this dataset card (viewer)
• checkpoints/ — trained weights (optional download)

Unzip the archives locally if you need the full multi-view float TIFF stacks. After unzip, paths match the git-repo layout (data/generated/m8_1/…, data/generated/m10_illumination/…).

Source code, notebooks, and generation configs live in the GitHub repository. Float32 .raw.tif sidecars under each sequence remain the authoritative ML image representation; JPEG/PNG previews exist for display and for the Hub Dataset Viewer.

Configurations

Config	Sequences	Workbook	Hub archive	After unzip
m8_1	900	configs/m8/localization_single_particle.xlsx	data/generated/m8_1.zip	data/generated/m8_1/single_particle/
m10_illumination	1500	configs/m10/localization_m10_illumination.xlsx	data/generated/m10_illumination.zip	data/generated/m10_illumination/
m10_illumination	3000	configs/m10/m10_variable_optics_generation.xlsx	data/generated/m10_illumination.zip	data/generated/m10_illumination_variable_optics/

m8_1 (default)

Single embedded particle, three optical regimes (low / medium / high), 36 camera views per sequence. The parquet index shows one preview row per sequence (camera angle 0°) with labels particle_x/y/z and regime metadata.

Split counts: train=630, validation=135, test=135.

m10_illumination

Same localisation task with six illumination setups (opt_m10_illum_{000,060,120,180,240,300}). Use the Hub config dropdown to switch corpora in the Dataset Viewer.

Split counts: train=1050, validation=228, test=222.

m10_illumination_variable_optics

Same localisation task as m10 illumiation, with six illumination setups, and random scattering/absorption coeffiecients of the gummybear (centered on "medium" properties from m8_1) (opt_m10_illum_{000,060,120,180,240,300}). Use the Hub config dropdown to switch corpora in the Dataset Viewer.

Split counts: train=2100, validation=450, test=450.

Columns (viewer)

Column	Meaning
observed / clean / particle / anomaly	Preview images (JPEG/PNG bytes embedded; path is the filename only, not a Hub tree path)
particle_x / particle_y / particle_z	Ground-truth particle centre (mm)
optical_regime	M8 regime label (low / medium / high) when applicable
illumination_angle_deg	M10 light angle decoded from optical_setup_id
observed_raw_path / anomaly_raw_path	Repo-relative float TIFF paths (valid after unzip)
sequence_dir	Repo-relative sequence folder (valid after unzip)

Load in Python

from datasets import load_dataset

ds = load_dataset("tbhugging/gummybear-tomography", "m8_1")
row = ds["train"][0]
print(row["sequence_id"], row["particle_x"], row["particle_y"], row["particle_z"])
row["observed"]  # PIL image from embedded bytes (no unzip required)

ds_m10 = load_dataset("tbhugging/gummybear-tomography", "m10_illumination")

load_dataset reads the parquet indexes only. Unzip data/generated/m8_1.zip / m10_illumination.zip locally if you need the float .raw.tif stacks referenced by observed_raw_path.

Regenerate metadata locally

From the repository root (after optical corpora exist on disk):

pip install ".[hf]" -c requirements.txt
python -m tomography_ml_validation.huggingface_metadata

Upload data/huggingface_metadata/ together with the zip archives data/generated/m8_1.zip and data/generated/m10_illumination.zip. Do not upload extracted trees: preview Image columns already contain JPEG/PNG bytes for the Dataset Viewer. Unzip locally only when you need the full stacks.

License

Apache-2.0 (same as the source repository).