Deep-ensemble fine-tune of a Pillar 4 foundation encoder — foundation_finetune

Trains K_ens independent heads on the same encoder with different random seeds and collects them into a single object. Each member is a full foundation_fit_classifier() or foundation_fit_regressor() fit.

Usage

foundation_finetune_ensemble(
  encoder,
  x,
  y,
  task = c("classification", "regression"),
  K_ens = 5L,
  base_seed = 301L,
  ...
)

Arguments

encoder: A MoCo/SimCLR encoder as returned by foundation_weights_load() (or an equivalent nn_module).
x, y: Training data; same shape requirements as the base fine-tune functions (4-D array (N, C, H, W) + vector y).
task: "classification" or "regression".
K_ens: Integer; number of ensemble heads. Defaults to 5L.
base_seed: Integer; each member uses base_seed + k - 1L.
...: Additional arguments forwarded verbatim to foundation_fit_classifier() or foundation_fit_regressor() (e.g. epochs, batch_size, lr, dropout, hidden, freeze_backbone, device, verbose).

Value

A list with class edaphos_foundation_ensemble containing members (list of K fits), task, K_ens, encoder (a reference to the fit-time encoder), and final_losses / loss_histories.

References

Lakshminarayanan, B., Pritzel, A. and Blundell, C. (2017). Simple and scalable predictive uncertainty estimation using deep ensembles. NeurIPS 30.