xnn.common.models.hub#

Model hub: one-line from_pretrained() loading of pre-trained models.

One API for every pre-trained model, whatever its origin:

from xnn.common.models import from_pretrained, list_models

list_models()                                       # registry + cache
model = from_pretrained("mace-off23-small")          # MACE foundation model
model = from_pretrained("xnn-mace-argon")            # xnn-trained model
model = from_pretrained("doi:10.5281/zenodo.18957344",
                        filename="mace_csfapbbri_al_5_1_stagetwo.model")
model = from_pretrained("runs/exp/best.pt")          # local checkpoint

Downloads are MD5-verified and cached under cache_dir (by default default_model_cache_dir()), one directory per model. Every cached model is stored as a portable directory (card.json, config.yaml, model.pt) with no pickled objects and no absolute paths, so a cache, or one model of it, can be copied to another system and loaded there; foreign formats (mace-torch) are converted once on download and need no upstream package afterwards. Write your own models in the same layout with save_pretrained(), and add them to the registry with register_pretrained() (or an entry in the packaged models.json).

Modules

`cache`	Where the model hub keeps downloaded and converted models.
`card`	Model cards: the metadata that describes one pre-trained model.
`checkpoint`	Reading, writing and building models from checkpoints.
`cli`	`xnn models`: list, inspect, download and package pre-trained models.
`formats`	Model file formats the hub can read, beyond xnn's own.
`loader`	`from_pretrained`: one call to load any pre-trained model.
`registry`	The registry of pre-trained models and `list_models()`.
`zenodo`	Resolve Zenodo DOIs and record links to downloadable files.