Saving and Loading Models

Saving

After training, save the full model state to a directory:

model.save("my_model/", overwrite=True)

This saves model weights, the AnnData manager registry, and init_params_ so the model can be reconstructed exactly.

Loading

from spatialvi import SCVIVA

model = SCVIVA.load("my_model/", adata=adata)

The adata argument must have the same variables (genes) as the training data.

Transferring to a new dataset (scArches / ARCHES)

SCVIVA and ResolVI inherit scvi-tools’ ArchesMixin, enabling query-to-reference mapping:

query_model = SCVIVA.load_query_data(
    adata_query,
    reference_model="my_model/",
)
query_model.train(max_epochs=100)