Train CNN contrastive and text-to-brain branches

Contrastive retrieval and text-to-brain projection are independent tasks. Choose one of the three domains in one place. Both default to the released mixed-source autoencoder.

from neurovlm.training import (
    ContrastiveTrainConfig, TextToBrainTrainConfig,
    train_contrastive, train_text_to_brain,
)

DOMAIN = "pubmed"            # pubmed | nilearn | neurovault
VARIANT = "mixed_baseline"  # explicit alternative: finetuned
AE_FROM_RUN = None            # e.g. "runs/<ae-run-id>"

Contrastive retrieval

contrastive = train_contrastive(ContrastiveTrainConfig(
    domain=DOMAIN, variant=VARIANT, output_root="runs",
    from_run=AE_FROM_RUN,
))
print(contrastive.run_dir)
print(contrastive.run_dir / "metrics/history.csv")
print(contrastive.run_dir / "metrics/curves.csv")

Text-to-brain projection

text_to_brain = train_text_to_brain(TextToBrainTrainConfig(
    domain=DOMAIN, variant=VARIANT, output_root="runs",
    autoencoder_from_run=AE_FROM_RUN,
))
print(text_to_brain.run_dir)
print(text_to_brain.run_dir / "metrics/history.csv")
print(text_to_brain.run_dir / "metrics/summary.csv")

Resume

To resume either task, keep its original run_id and pass its run directory through resume. This restores the optimizer, epoch, best-metric state, and accumulated metric history.