gfa.config#

Configuration classes for GFA model fitting and simulation.

Classes

GFAFitConfig

Configuration for GFA model fitting.

GFASimConfig

Experimental design parameters for GFA simulation.
class GFAFitConfig(
*,
x_dim_init: int = 1,
fit_tol: float = 1e-08,
max_iter: int = 1000000,
prune_x: bool = True,
prune_tol: float = 1e-07,
save_x: bool = False,
save_c_cov: bool = False,
save_fit_progress: bool = True,
random_seed: int | Sequence[int] | None = None,
min_var_frac: float = 0.001,
)[source]#

Configuration for GFA model fitting.

All parameters have sensible defaults. Create with keyword arguments only. Instances are immutable (frozen).

Parameters:
x_dim_initint

Initial number of latent dimensions (before pruning). Must be >= 1.

fit_tolfloat

Convergence tolerance for ELBO relative change. Must be > 0.

max_iterint

Maximum EM iterations. Must be >= 1.

prune_xbool

If True, remove latent dimensions that become inactive during fitting. Improves speed and memory for high initial x_dim_init.

prune_tolfloat

Variance threshold for pruning. Latents with mean squared value below this are removed. Must be > 0.

save_xbool

If True, save posterior latent estimates. Can be memory-intensive for large N.

save_c_covbool

If True, save loading covariances. Can be memory-intensive for large y_dim and x_dim.

save_fit_progressbool

If True, track ELBO and runtime per iteration.

random_seedint, sequence of int, or None

RNG seed for reproducibility. Accepts a single non-negative integer, a non-empty sequence of non-negative integers (for structured seeding in parallel experiments), or None for random initialization.

min_var_fracfloat

Private variance floor as fraction of data variance. Must be in (0, 1).

Examples

>>> config = GFAFitConfig(x_dim_init=10)
>>> config.x_dim_init
10

>>> # Configs are immutable
>>> config.x_dim_init = 20  # Raises FrozenInstanceError
Traceback (most recent call last):
    ...
dataclasses.FrozenInstanceError: cannot assign to field 'x_dim_init'
class GFASimConfig(
*,
n_samples: int,
y_dims: ndarray,
x_dim: int,
snr: float | ndarray = 1.0,
random_seed: int | Sequence[int] | None = None,
)[source]#

Experimental design parameters for GFA simulation.

Defines the structure and reproducibility settings for generating synthetic GFA data. The probabilistic model specification (hyperprior) is provided separately to simulate().

All parameters except random_seed are required. Instances are immutable (frozen).

Parameters:
n_samplesint

Number of data points to generate. Must be >= 1.

y_dimsnp.ndarray

Dimensionalities of each observed group, shape (n_groups,). Must be 1D array of positive integers.

x_dimint

Number of latent dimensions. Must be >= 1.

snrfloat or ndarray

Signal-to-noise ratio. Either a scalar (broadcast to all groups) or per-group array of shape (n_groups,). Must be > 0.

random_seedint, sequence of int, or None

RNG seed for reproducibility. Accepts a single non-negative integer, a non-empty sequence of non-negative integers (for structured seeding in parallel experiments), or None for random initialization. Required for reproducible recipe saves.

Examples

>>> config = GFASimConfig(
...     n_samples=100,
...     y_dims=np.array([10, 10, 10]),
...     x_dim=5,
...     snr=1.0,
...     random_seed=42,
... )
>>> config.n_groups
3
>>> config.y_dim
30
On this page