Reproducibility#

Probabilistic models involve random sampling and random initialization. Reproducing results requires controlling these sources of randomness. Latents uses NumPy’s modern Generator API throughout, and every config class accepts a random_seed parameter.

Setting a random seed#

Pass random_seed to any config to make results reproducible. The default None draws fresh entropy each run.

from latents.gfa import GFAModel, GFAFitConfig, GFASimConfig

# Reproducible simulation
sim_config = GFASimConfig(
    n_samples=100,
    y_dims=np.array([10, 10]),
    x_dim=5,
    random_seed=42,
)

# Reproducible fitting
fit_config = GFAFitConfig(x_dim_init=10, random_seed=42)
model = GFAModel(config=fit_config)
model.fit(Y)

# Running again with the same seed produces the same result
model2 = GFAModel(config=fit_config)
model2.fit(Y)

What the seed controls#

The seed controls different things depending on the context:

  • Simulation (simulate()): sampling loading matrices from the prior, sampling latent variables, and generating observation noise. The same seed produces identical synthetic data.

  • Fitting (model.fit()): initializing the posterior estimates. The variational EM updates that follow are deterministic given the starting point. The seed determines which local optimum the optimizer finds.

Because fitting uses the seed only at initialization, two runs with the same seed and same data will converge to the same solution.

Structured seeding for parallel experiments#

When running a sweep of independent fits, each run needs its own seed. Manually choosing unrelated integers works but is fragile. A better approach is to pass a sequence of integers:

base_seed = 42

for run_idx in range(10):
    config = GFAFitConfig(
        x_dim_init=10,
        random_seed=[run_idx, base_seed],
    )
    model = GFAModel(config=config)
    model.fit(Y)

This works because NumPy’s SeedSequence hashes integer sequences into statistically independent bit generators. Sequences that differ in any element — such as [0, 42] and [1, 42] — produce unrelated random streams. You can add as many index dimensions as your experiment needs (e.g., [sweep_idx, run_idx, base_seed]).

Further reading#