Fine-tune a fitted torch model

Training is CPU-only and passes only explicitly trainable parameters to Adam. Frozen modules, including their batch-normalization buffers, remain in evaluation mode. Validation and prediction data must reuse frozen normalization statistics.

Usage

fineTune(
  model,
  train,
  valid = NULL,
  strategy = c("linear_probe", "partial", "full"),
  unfreeze = NULL,
  epochs = 10L,
  batch_size = 32L,
  learning_rate = 1e-04,
  weight_decay = 0,
  callbacks = NULL,
  seed = 1L,
  device = "cpu",
  num_workers = 0L,
  verbose = FALSE
)

Arguments

model

A fitted physio_dl_model or physio_transfer_model.

train, valid

Compatible targeted datasets.

strategy

Exact tuning strategy.

unfreeze

Component prefixes used only by partial tuning.

epochs, batch_size

Positive whole-number settings.

learning_rate

Positive Adam learning rate.

weight_decay

Non-negative Adam weight decay.

callbacks

Currently NULL; callback execution is owned by trainModel()'s luz workflow.

seed

Whole-number CPU seed.

device

Currently exactly "cpu".

num_workers

Currently exactly zero.

verbose

Whether to print one line per epoch.

Value

A fitted physio_dl_model with a plain transfer record.

Examples

arr <- array(stats::rnorm(50 * 2 * 6), dim = c(50, 2, 6))
pe <- PhysioExperiment::PhysioExperiment(
  assays = list(raw = arr),
  colData = S4Vectors::DataFrame(label = c("C3", "C4")),
  samplingRate = 100
)
# Needs the optional torch and luz backends.
if (requireNamespace("torch", quietly = TRUE) &&
    requireNamespace("luz", quietly = TRUE)) {
  train <- peDataset(
    pe, targets = rep(c("a", "b"), 3),
    task = "classification", class_levels = c("a", "b")
  )
  fit <- trainModel(train, model = "cnn1d", epochs = 1L)
  tuned <- fineTune(fit, train, strategy = "linear_probe", epochs = 1L)
}
#> Error: the R package `torch` is installed but LibTorch is unavailable; install it explicitly with `torch::install_torch()`