create_oilmm_from_data#

gpjax.models.create_oilmm_from_data(dataset, num_latent_gps, key, kernel=None, mean_function=None)[source]#

Create OILMM with data-informed initialization of mixing matrix.

Initializes U to the top M eigenvectors and S to the top M eigenvalues of the empirical covariance matrix of the outputs. Near-zero eigenvalues are clamped to 1e-6 for numerical stability. This can provide better initialization than random, especially when outputs have clear correlation structure.

Parameters:

  • dataset (Dataset) – Training data with y [N, P]

  • num_latent_gps (int) – Number of latent GPs (m)

  • key (Array) – JAX PRNG key

  • kernel (AbstractKernel) – Kernel for latent GPs (default: RBF)

  • mean_function (tp.Any) – Mean function (default: Zero)

Returns:

OILMMModel with U initialized to top M eigenvectors and S to top M eigenvalues

Return type:

OILMMModel

Example

>>> import gpjax as gpx
>>> import jax.numpy as jnp
>>> import jax.random as jr
>>> X = jnp.linspace(0, 1, 50).reshape(-1, 1)
>>> y = jnp.column_stack([jnp.sin(X), jnp.cos(X)])
>>> data = gpx.Dataset(X=X, y=y)
>>> model = gpx.models.create_oilmm_from_data(
...     dataset=data,
...     num_latent_gps=2,
...     key=jr.key(42)
... )

Expand for references to gpjax.models.create_oilmm_from_data

create_oilmm_from_data