Save and load¶

gamfit persists fitted models (.gam) and posterior samples (.npz).

Models — `.gam`¶

model = gamfit.fit(train, "y ~ s(x)")
model.save("model.gam")

loaded = gamfit.load("model.gam")
preds  = loaded.predict(test)

In-memory transport:

blob   = model.dumps()              # bytes
loaded = gamfit.loads(blob)

The .gam file is the JSON serialisation of the Rust Model struct (produced by serde_json::to_writer). It contains:

coefficients, smoothing parameters, basis specifications;
formula and family / link metadata;
the data schema used by Model.check(...) and Model.predict(...).

It does not contain the training data.

After loading, every method works as on the original model:

loaded.predict(test, interval=0.95)
loaded.summary()
loaded.diagnose(test)
loaded.sample(test, seed=42)

Posterior samples — `.npz`¶

posterior = model.sample(train, seed=42)
posterior.save("posterior.npz")

restored = gamfit.load_posterior("posterior.npz")
# equivalent:
restored = gamfit.PosteriorSamples.load("posterior.npz")

The archive (written via numpy.savez) contains:

Array	Shape / dtype	Meaning
`samples`	`(n_draws, n_coeffs)` float64	Raw draws.
`mean`	`(n_coeffs,)` float64	Per-coefficient posterior mean.
`std`	`(n_coeffs,)` float64	Per-coefficient posterior std.
`rhat`	scalar float64	Convergence diagnostic.
`ess`	scalar float64	Effective sample size.
`converged`	scalar bool	Convergence flag.
`model_bytes`	1-D uint8	The fitted model's serialised bytes.
`metadata`	0-D object (JSON)	`coefficient_names`, `method`, `model_class`, `family_kind`, `config`.

PosteriorSamples.load requires allow_pickle=True to round-trip the metadata object array; only load archives produced by save.

Because the model bytes are bundled, predict() and predict_draws() work after a round-trip:

restored = gamfit.load_posterior("posterior.npz")
bands    = restored.predict(test, level=0.95)

Version compatibility¶

.gam payloads are version-gated by the Rust loader (validate_for_persistence). Models written by an older release may fail to load in a newer release if a required field has been added. For long-term archival, pin the gamfit version or refit after upgrades.

Patterns¶

Save model and posterior together¶

model.save("model.gam")
model.sample(train, seed=42).save("posterior.npz")

The posterior archive carries the model bytes, so loading just the posterior is enough to call predict().

Inspect a model without the training data¶

m = gamfit.load("model.gam")
print(m.summary())
print(m.model_class, m.formula)

Round-trip in tests¶

blob = model.dumps()
assert (gamfit.loads(blob).predict(test)["mean"]
        == model.predict(test)["mean"]).all()