Model Persistence

Model Persistence#

Saving models (production default)#

All 14 production models support safe bundles (corerec_safe_v1):

model.save("/artifacts/my_model")   # safe=True by default

This writes JSON metadata plus separate weight/array files — no pickle in the default path.

Loading models#

from corerec.engines import DCN

loaded = DCN.load("/artifacts/my_model")  # auto-detects safe vs legacy
recommendations = loaded.recommend(user_id=1, top_k=10)

Safe bundle specification#

See Safe Model Bundles (corerec_safe_v1) for file layout, ID map encoding, and migration from legacy .pkl / .pt artifacts.

Legacy formats#

model.save("legacy.pkl", safe=False)  # pickle/torch — untrusted files are unsafe

Migrate existing deployments:

legacy = SAR.load("old/sar.pkl")
legacy.save("prod/sar", safe=True)

Verify round-trip quality#

uid, iid = 42, 1001
assert abs(model.predict(uid, iid) - loaded.predict(uid, iid)) < 1e-2
recs = loaded.recommend(uid, top_k=10)

Model information#

info = model.get_model_info()
print(f"Model: {info['name']}")
print(f"Fitted: {info['is_fitted']}")

Checking model state#

if model.is_fitted:
    recs = model.recommend(user_id=1, top_k=10)

if model.knows_user(42):
    score = model.predict(user_id=42, item_id=100)
Contents