Save and load

gspice.save() writes a gspice.Model to a single .npz archive, and gspice.load() reconstructs it on disk.

Save a model

Fitting a model takes time, so save it once and load it where you predict. This model also has the block thresholds of Find and repair artifacts:

Python

import numpy as np

import gspice

rng = np.random.default_rng(7)
train = gspice.simulate(1500, 100, rng=rng)
clean = gspice.simulate(1000, 100, rng=rng)
model, _ = gspice.fit(train.flux, train.ivar)
model = gspice.fit_blocks(model, clean.flux, clean.ivar, width=50, stride=25)

gspice.save(model, "model.npz")

save replaces an existing file, and adds no suffix to the path.

Load it

Python

loaded = gspice.load("model.npz")
test = gspice.simulate(200, 100, rng=rng, artifacts=0.3)
before = gspice.repair(model, test.flux, test.ivar)
after = gspice.repair(loaded, test.flux, test.ivar)
print(np.array_equal(before.replaced, after.replaced))
print(np.allclose(before.flux, after.flux, rtol=1e-10, atol=0))

Text

True
True

The file holds every field of the model exactly, the blocks included. The loaded model replaces the same pixels, and predicts the same flux to rounding, because it computes its precision matrix again (What the file holds).

What the file holds

The file is an ordinary .npz archive, so numpy.load also reads it:

Python

print(np.load("model.npz").files)

Text

['schema', 'gspice_version', 'mean', 'n_train', 'edges', 'center', 'weighted', 'cov0', 'block_starts', 'block_width', 'block_nguard', 'block_thresholds']

The docstring of save describes each array. The file holds no precision matrices. A model computes the precision matrix of a bin at its first use, and keeps it (Cost). So the first prediction of a loaded model takes longer than the next ones.