plotTrainingSchedule

Visualise per-stage base learning rates, an optional LR scheduler trace, and deep supervision weight decay without running a training loop.

Usage

plotTrainingSchedule(
  stageLRs,
  epochs,
  initialWeights = NULL,
  decay = "linear",
  primary_only_frac = 0.1,
  stageNames = NULL,
  scheduler = NULL
)

Arguments

stageLRs

Numeric vector of per-stage base learning rates matching the stageLRs argument passed to the model. A single value produces a one-bar chart.

epochs

Total number of training epochs.

initialWeights

Optional numeric vector of initial deep supervision loss weights as passed to make_ds_weights. When NULL the DS weight panel is omitted.

decay

Character string, either "linear" (default) or "cosine", matching the decay argument passed to callback_ds_weight_decay.

primary_only_frac

Numeric value in (0, 1) specifying the fraction of total epochs reserved for primary-loss-only training at the end of the run, matching the primary_only_frac argument passed to callback_ds_weight_decay. Defaults to 0.1.

stageNames

Optional character vector of stage labels for the learning rate chart. Defaults to "Stage 1", "Stage 2", ...

scheduler

Optional named list specifying an LR scheduler to visualise. Must contain at least type ("cosine_annealing", "one_cycle", or "step"). See Details for per-type parameters.

Value

Invisibly returns a named list with element lr (a data frame of stage names and learning rates), lr_trace (a data frame of epoch-by-LR values when scheduler is supplied), and weights (a data frame of epoch-by-weight values when initialWeights is supplied).

Details

Produces up to three panels depending on inputs. The left panel always shows a bar chart of per-stage base learning rates. When scheduler is supplied a centre panel shows the effective learning rate curve over all epochs. When initialWeights is supplied a right panel shows how each deep supervision weight evolves over epochs, replicating the exact decay applied by callback_ds_weight_decay.

The scheduler list accepts three types, each mirroring a torch R scheduler:

"cosine_annealing"

Mirrors lr_cosine_annealing. Named elements: base_lr (starting LR; defaults to max(stageLRs)), T_max (epochs per half-cycle; defaults to epochs), eta_min (minimum LR, default 0). The LR oscillates between base_lr and eta_min with period 2 * T_max.

"one_cycle"

Mirrors lr_one_cycle. Named elements: max_lr (peak LR; defaults to max(stageLRs)), pct_start (warmup fraction, default 0.3), div_factor (initial_lr = max_lr / div_factor, default 25), final_div_factor (min_lr = initial_lr / final_div_factor, default 1e4), anneal_strategy ("cos" or "linear", default "cos").

"step"

Mirrors lr_step. Named elements: base_lr (starting LR; defaults to max(stageLRs)), step_size (epochs between LR drops; defaults to max(1, floor(epochs / 10))), gamma (multiplicative decay factor per drop, default 0.1).