neurovlm.models.load_model

Contents

neurovlm.models.load_model#

neurovlm.models.load_model(name=None, *, family=ModelFamily.MLP, task=None, domain=None, variant=None)[source]#

Load a packaged model by legacy name or structured fields.

Parameters:

  • name (str, optional) – Existing packaged-model name. Every name supported before the structured API remains an alias to the same checkpoint.

  • family ({"mlp", "cnn"}, default: "mlp") – Architecture family. MLP remains the global default.

  • task (str, optional) – Structured task identifier. Required when name is omitted.

  • domain ({"pubmed", "nilearn", "neurovault"}, optional) – Required for CNN contrastive/text-to-brain models and for fine-tuned CNN autoencoders.

  • variant (str, optional) – CNN models default to mixed_baseline. Domain-specialized CNN checkpoints require the explicit value finetuned.

Return type:

model

Examples

Existing calls continue to work:

>>> model = load_model("autoencoder")

Structured CNN selection defaults to the mixed baseline:

>>> model = load_model(family="cnn", task="contrastive", domain="nilearn")
Contents