Fine-tune or linearly probe a Pillar 4 encoder for regression

Regression counterpart of foundation_fit_classifier(). Attaches a scalar-output head (linear or MLP) on top of a self-supervised encoder and trains it against a numeric target y. Supports the same two regimes — linear probing with a frozen backbone or full fine-tuning with a two-group learning rate — and the same device ∈ {"cpu", "mps", "cuda"} dispatch.

Usage

foundation_fit_regressor(
  encoder,
  x,
  y,
  freeze_backbone = TRUE,
  head = c("linear", "mlp"),
  hidden = c(64L, 32L),
  dropout = 0,
  epochs = 30L,
  batch_size = 32L,
  lr = 0.001,
  weight_decay = 0,
  backbone_lr_mult = 0.1,
  loss = c("mse", "huber"),
  val_split = 0.2,
  device = c("cpu", "mps", "cuda"),
  seed = NULL,
  verbose = FALSE
)

Arguments

encoder

See foundation_fit_classifier().

x

A 4-D array (N, C, H, W) of patches.

y

Numeric vector of length N with the target values.

freeze_backbone, head, hidden, dropout, epochs, batch_size, lr, weight_decay, backbone_lr_mult, val_split, device, seed, verbose

See foundation_fit_classifier().

loss

One of "mse" (default) or "huber". Huber is more robust to outlier pedons with extreme SOC / clay contents.

Value

An edaphos_foundation_regressor list with the same slots as the classifier counterpart plus y_mean, y_sd (target normalisation constants) and val_rmse_history.

Details

Target normalisation is handled internally: y is centred and scaled before training and un-scaled at predict() time so the user never has to think about the numerical range of the head.

See also

predict.edaphos_foundation_regressor().

Examples

if (FALSE) { # \dontrun{
  moco <- foundation_weights_load("edaphos-cerrado-moco-v1")
  ds   <- readRDS("tools/pretrain/cerrado_dataset.rds")
  patches <- array(rnorm(300 * ds$n_channels * 16 * 16),
                   dim = c(300, ds$n_channels, 16, 16))
  soc <- rnorm(300, mean = 15, sd = 6)
  fit <- foundation_fit_regressor(
    moco, patches, soc,
    freeze_backbone = TRUE, head = "linear",
    epochs = 40L, device = "mps", seed = 1L
  )
  predict(fit, patches[1:10, , , , drop = FALSE])
} # }