Train CNN contrastive retrieval on PubMed

This is the integrated version of the original best-recipe experiment. The released mixed-source autoencoder initializes the brain encoder by default; fine-tuning is never selected implicitly.

from neurovlm.training import ContrastiveTrainConfig, train_contrastive

config = ContrastiveTrainConfig(
    domain="pubmed",             # pubmed | nilearn | neurovault
    variant="mixed_baseline",  # explicit alternative: finetuned
    output_root="runs",
    epochs=100,
)
result = train_contrastive(config)

print("run:", result.run_dir)
print("best checkpoint:", result.best_checkpoint)
print("epoch metrics:", result.run_dir / "metrics/history.csv")
print("recall curves:", result.run_dir / "metrics/curves.csv")

Explicit local initialization and resume

Leave from_run unset for released Hugging Face initialization. Set it only when chaining from a locally trained autoencoder. Resume retains the original run_id and uses the existing run directory.

# config = ContrastiveTrainConfig(domain="pubmed", from_run="runs/<ae-run-id>")
# config = ContrastiveTrainConfig(domain="pubmed", run_id="my-run", resume="runs/my-run")