TrainingPhase

			TrainingPhase() <: AbstractTrainingPhase

A regular training phase for supervised learning. It iterates over batches in learner.data.training and updates the model parameters using learner.optim after calculating the gradients.

Throws the following events in this order:

• EpochBegin when an epoch starts,

• StepBegin when a step starts,

• LossBegin after the forward pass but before loss calculation,

• BackwardBegin after loss calculation but before backward pass,

• BackwardEnd after the bacward pass but before the optimization step,

• StepEnd when a step ends; and

• EpochEnd when an epoch ends

It writes the following step state to learner.state, grouped by the event from which on it is available.

• StepBegin:

  • xs and ys: encoded input and target (batch)

• LossBegin:

  • ŷs: model output

• BackwardBegin:

  • loss: loss

• BackwardEnd:

  • grads: calculated gradients