CLI Reference¶
The rddac command line interface downloads the dataset from DaRUS and prints summary information. Because RDDAC is the experimental counterpart to DDACS, the full download also fetches the matching DDACS simulations (skip with --no-sim).
rddac info¶
Display dataset information and the list of available versions.
rddac download¶
Download dataset files from DaRUS.
Arguments¶
| Argument | Default | Description |
|---|---|---|
[VERSION] |
2.0 |
Dataset version (default: 2.0) |
Options¶
| Option | Description |
|---|---|
--files FILE... |
Specific filenames to download |
--small |
Download the small sample bundle |
--no-sim |
Download only the real measurements (skip the DDACS simulations) |
--out PATH |
Output directory (default: ./data) |
-y, --yes |
Skip confirmation prompt |
-q, --quiet |
No output or progress bars; implies --yes |
--extract |
Extract downloaded zips into their directory |
--remove-zip |
Delete zips after extraction (with --extract) |
--small fetches ~174 MB: sample.zip (one experiment per category), process_parameters.csv, metadata.json.
Default behaviour¶
Zip files are kept on disk by default and are not extracted. This keeps the dataset readable in place by mlcroissant, which references zip members through the Croissant manifest. Pass --extract to additionally write the HDF5 files to disk, and --remove-zip to delete the zip afterwards.
The --out directory defaults to ./data. The same value is used by rddac.load(data_dir=...) and RDDACDataset(data_dir=...), so files written by rddac download are picked up by the Python API without additional configuration.
DDACS simulations¶
A full download (no --small, no --files) additionally fetches rddac.zip (~9 GB) from the DDACS dataset into ./data/simulation/. It contains the FEM simulations matching the RDDAC parameter grid, for simulation-to-reality comparison. Pass --no-sim to skip this second download; the simulations are also listed and confirmed separately, so answering n at the prompt has the same effect.
Examples¶
# Download the small test set (~174 MB)
rddac download --small -y
# Download the full dataset: measurements + matching DDACS simulations
rddac download
# Real measurements only
rddac download --no-sim
# Download to a custom directory
rddac download --out /path/to/data
# Download specific files only
rddac download --files process_parameters.csv metadata.json sample.zip
# Extract the zip files in place, keep the zip alongside
rddac download --extract
# Extract and remove the zip after a successful extraction
rddac download --extract --remove-zip
After --extract --remove-zip, the HDF5 files are no longer wrapped in zips and mlcroissant cannot resolve the FileSet. See the Loose HDF5 recipe for the appropriate iteration pattern in that case.
rddac preprocess¶
Process raw measurements into the ML-ready processed layout; see Preprocessing for what each modality does. Raw files are never modified; processed files are written separately.
Arguments¶
| Argument | Default | Description |
|---|---|---|
[MODALITY] |
Subset of: force, sheet, oil, pointcloud (default: all) |
pointcloud covers both scan grids, z and luminescence. Without an explicit modality list, pointcloud is skipped with a notice when the DDACS simulations (or the rddac[preprocessing] extra) are missing; named explicitly, it fails with an actionable error instead.
Options¶
| Option | Description |
|---|---|
--data-dir PATH |
Directory holding the raw dataset (default: ./data) |
--out PATH |
Output directory for processed files (default: |
--ids IDS |
Experiment id selection, e.g. '0-999' or '42,1035' |
--split {train,val,test} |
Restrict to one of the predefined splits |
--workers N |
Parallel workers (default: 1) |
--overwrite |
Recompute modalities that already exist in the output files |
--config TOML |
TOML file overriding processing parameters (see --dump-config) |
--dump-config |
Print the default processing parameters as TOML and exit |
--rebuild-models |
Force retraining of the cached fin-classifier models (pointcloud stage) |
-y, --yes |
Skip confirmation prompt |
-q, --quiet |
No output or progress bars; implies --yes |
Default behaviour¶
Output lands in <data-dir>/processed as loose <id>.h5 files. Re-runs append: modalities that already exist in an output file are skipped unless --overwrite is passed. The command refuses an output directory that looks like a raw dataset directory (see output directory rules).
Examples¶
# Everything, reference recipe
rddac preprocess
# Only the traverse measurements, 8 workers
rddac preprocess sheet oil --workers 8
# A parameter variant, published next to your code
rddac preprocess --dump-config > my.toml
rddac preprocess oil --config my.toml
# One split, custom locations
rddac preprocess --split train --data-dir /mnt/rddac --out /mnt/rddac-processed
Global Options¶
| Option | Description |
|---|---|
-V, --version |
show program's version number and exit |
--token TOKEN |
DaRUS API token (for draft access) |