Run multiple SmoothEM initializations in parallel — parallel

Runs initialize_smoothEM() for a set of initialization methods in parallel. Each method is fit for num_iter total SmoothEM iterations. Internally, initialize_smoothEM() warm-starts with exactly one EM iteration, and if num_iter > 1, continues via do_smoothEM() (where adaptive lambda updates may be enabled).

For compatibility with method="multi_scale", if K is not provided, then non-multi-scale methods default to K = 2^m_max + 1.

parallel_initial(
  X,
  methods = c("PCA", "tSNE", "random", "fiedler", "multi_scale", "pcurve"),
  num_iter = 1,
  num_cores = 2,
  m_max = 6,
  K = NULL,
  seed = NULL,
  adaptive = TRUE,
  lambda_min = 1e-08,
  lambda_max = 1e+08,
  quiet = TRUE,
  ...
)

Arguments

X: Numeric matrix (n x d).
methods: Character vector of methods to try. Defaults to c("PCA","tSNE","random","fiedler","multi_scale").
num_iter: Integer >= 1. Total number of SmoothEM outer iterations to run for each method.
num_cores: Integer >= 1. Number of cores for parallel execution.
m_max: Integer >= 1. Used by multi_scale and to set default K.
K: Optional integer grid size for non-multi-scale methods. If NULL, uses 2^m_max+1.
seed: Optional base seed for reproducibility. If provided, each method gets a deterministic derived seed.
adaptive: Logical; if TRUE, enables adaptive lambda behavior in initialize_smoothEM().
lambda_min, lambda_max: Positive bounds used to clip lambda when adaptive=TRUE.
quiet: Logical; suppress messages from workers.
...: Extra args passed to initialize_smoothEM() (e.g. rw_q, lambda, relative_lambda, modelName, etc.).

Value

A named list of smooth_em objects (or NULL for failed fits), with attributes:

summary: a data.frame summarizing last ELBO / last objective / last lambda for each method.