sgdrf.sgdrf_config

Config for SGDRFs.

Module Contents

class sgdrf.sgdrf_config.SGDRFConfig

Streaming Gaussian Dirichlet Random Field configuration holder.

Parameters:

  • xu_ns (Union[int, Collection[int]]) – Number of inducing points, either one for each spatiotemporal dimension or a single number.

  • d_mins (Collection[float]) – A collection containing the minimum extent of each dimension

  • d_maxs (Collection[float]) – A collection containing the maximum extent of each dimension

  • V (int) – The number of observation types

  • K (int) – The number of latent Gaussian processes

  • max_obs (int) – The maximum number of possible simultaneous categorical observations

  • dir_p (float) – Initial uniform Dirichlet hyperparameter

  • kernel (pyro.contrib.gp.kernels.Kernel) – Latent Gaussian process kernel

  • optimizer (pyro.optim.PyroOptim) – Stochastic gradient descent optimization algorithm

  • subsampler (Subsampler) – Subsampling algorithm

  • device (torch.device, optional) – Pytorch device (e.g. torch.device(‘cuda’)), by default torch.device(“cpu”)

  • whiten (bool, optional) – Whether the Gaussian process covariance matrix is whitened, by default False

  • fail_on_nan_loss (bool, optional) – Whether to raise an exception if training loss is NaN, by default True

  • num_particles (int, optional) – Number of parallel posterior latent samples to draw, by default 1

  • jit (bool, optional) – Whether to JIT-compile the model and guide, by default False

K: int
V: int
d_maxs: Collection[float]
d_mins: Collection[float]
device: torch.device
dir_p: float
fail_on_nan_loss: bool = True
jit: bool = False
kernel: pyro.contrib.gp.kernels.Kernel
max_obs: int
num_particles: int = 1
optimizer: pyro.optim.PyroOptim
subsampler: sgdrf.subsample.Subsampler
whiten: bool = False
xu_ns: int | Collection[int]