Tutorials¶
These tutorials walk through the typical workflows on top of the ddacs Python surface, from a first download to a PyTorch training loop. Each page is short and code centric: read top to bottom or jump to the section that matches the task at hand.
- Getting Started: Install, download the small bundle, load the manifest, open one simulation, render the first plot.
- Build your own view: Use
ddacs.add_viewto compose a custom Croissant RecordSet from the field map. - PyTorch training: Stream a view through
DDACSDataset, batching withDataLoader, multi worker sharding, DDP, and a reproducibility note. Includes the per simulation read benchmark. - Visualization: Mesh, point cloud, vector field, and projection plots on top of
ddacs.open_h5. - Loose HDF5 recipe: Iterate over loose
.h5files afterddacs download --extract --remove-zip. - Streaming and numpy export: Iterate any view with
ddacs.streaming.iter_view(no PyTorch, no mlcroissant FileSet walk), and materialise it as flat.npyshards withstreaming.export_to_numpyfor ~1000x faster training reads.
Prerequisites¶
Notebooks¶
The companion Jupyter notebooks live in the notebooks/ directory of the repository. Each tutorial page has a matching notebook that reproduces the code blocks end to end.