espnet3.components.callbacks.default_callbacks.MetricsLogger

source

class espnet3.components.callbacks.default_callbacks.MetricsLogger(log_every_n_steps: int = 500)

Bases: Callback

Log compact train and validation metric summaries.

This callback owns the human-readable metric logging for the default ESPnet3 training loop. It handles three reporting points in one place.

• interval-based training batch summaries
• end-of-epoch training summaries
• end-of-epoch validation summaries

The goal is to keep logging responsibility in a single callback instead of splitting train and validation reporting across separate callback classes.

• Parameters:log_every_n_steps – Number of training steps between batch-summary log lines.
• Returns: None.
• Raises:None. –

Notes

• Training summaries remove the train/ prefix from logged metrics.
• Validation summaries remove the valid/ prefix from logged metrics.
• Validation sanity-check runs are ignored to avoid noisy startup logs.

Examples

>>> cb = MetricsLogger(log_every_n_steps=200)
>>> trainer = Trainer(callbacks=[cb, ...])

Example train log output. 20epoch🚋4201-4400batch: iter_time=6.212e-05, loss=46.669

Example validation log output. epoch_summary:20epoch:valid: valid_time=1.42, acc=0.91, loss=0.83

Initialize the logger with a reporting interval.

on_after_backward(trainer, pl_module)

Record backward time.

on_after_optimizer_step(trainer, pl_module)

Record optimizer step time.

on_before_backward(trainer, pl_module, loss)

Mark the end of forward pass and start backward timing.

on_before_optimizer_step(trainer, pl_module, optimizer)

Start optimizer step timing.

on_train_batch_end(trainer, pl_module, outputs, batch, batch_idx)

Aggregate metrics and emit a summary line when due.

on_train_batch_start(trainer, pl_module, batch, batch_idx)

Start timers for a training batch.

on_train_epoch_end(trainer, pl_module)

Reset buffers at the end of each training epoch.

on_validation_epoch_end(trainer, pl_module)

Emit one log line with aggregated validation metrics.

on_validation_epoch_start(trainer, pl_module)

Start wall-clock timing for one validation epoch.