espnet3.components.modeling.optimization_spec.OptimizationStep

source

class espnet3.components.modeling.optimization_spec.OptimizationStep(loss: Tensor, name: str)

Bases: object

Describe one optimizer update emitted by the model.

ESPnet3 keeps the model return contract as (loss, stats, weight). In the single-optimizer path, loss is a plain tensor. In the multiple-optimizer path, loss becomes either one OptimizationStep or a list of them so the training loop knows which named optimizer should consume each loss.

Example

OptimizationStep(loss=g_loss, name=”generator”)

means “apply g_loss to the optimizer and scheduler pair named generator.” The name must match the keys configured under optimizers.<name> and schedulers.<name>.

loss : Tensor

name : str