API Reference

Learner{M, D, F, P, Q}

Object bundling together all information for training.

• model: Architecture optimized during training.
• data_loader: Iterable for training data.
• loss_fn: Function calculating the loss.
• optim: Optimizer state.
• quantities: Metrics computed every training step.

train!(learner, epochs, with_plots)

Train a Learner for a number of epochs, optionally with live plotting.

train_loop!(learner, epochs, channel)

Internal training loop that computes quantities and sends them to the display channel.

compute(q::AbstractQuantity, losses, back, grads, params)

Compute the value of quantity q using the provided training information.

quantity_key(q::AbstractQuantity)

Return a symbol key uniquely identifying the quantity.

AbstractQuantity

Abstract base type for all quantities tracked during training.

LossQuantity

Quantity tracking the training loss.

DistanceQuantity

Quantity tracking the L2 distance of current parameters from the initial parameters.

GradNormQuantity

Quantity tracking the norm of the model gradients.

UpdateSizeQuantity

Quantity tracking the L2 distance of parameters before and after the current update step.

NormTestQuantity

Quantity tracking the normalized gradient noise.

GradHist1dQuantity

Histogram of per-sample gradient elements.

build_dashboard(quantities) -> (fig, axes_dict)

Construct the dashboard layout for the given quantities and return the figure and axis mapping. Loss quantity is always plotted