Skip to contents

K-seed deep ensemble of stacked ConvLSTMs

Source: R/temporal_posterior.R
temporal_convlstm_ensemble_fit.Rd

Trains K_ens independent temporal_convlstm_fit() models with different random seeds and collects them in a single object. This formalises the hand-rolled ensemble loop from the v1.5.0 Pillar 3 Cerrado runner (data-raw/temporal_cerrado_run.R) and produces the natural forecast-ensemble input for temporal_kalman_update() or for as_edaphos_posterior().

Usage

temporal_convlstm_ensemble_fit(
  sequence,
  target,
  hidden_dims = c(8L, 4L),
  kernel_size = 3L,
  return_sequence = FALSE,
  epochs = 80L,
  lr = 0.02,
  K_ens = 10L,
  base_seed = 101L,
  physics_lambda = 0,
  physics_k_in = 0.03,
  physics_k_out = 0.015,
  physics_driver_channel = 2L,
  verbose = FALSE
)

Arguments

sequence, target, hidden_dims, kernel_size, return_sequence, epochs, lr, physics_lambda, physics_k_in, physics_k_out, physics_driver_channel, verbose

Passed through unchanged to temporal_convlstm_fit().

K_ens

Integer; number of ensemble members. Defaults to 10L.

base_seed

Integer; each member uses base_seed + k - 1L.

Value

A list with class edaphos_temporal_convlstm_ensemble containing

members

List of K edaphos_temporal_convlstm fits.

K_ens

Integer, the ensemble size.

final_losses

Numeric vector of per-member final training losses.

loss_histories

List of K numeric vectors (one per member).

See also

temporal_convlstm_rollout() to roll each member forward, temporal_kalman_update() to assimilate observations into the forecast ensemble, as_edaphos_posterior() for the unified uncertainty wrapper.