espnet3.components.callbacks.default_callbacks.TrainBatchMetricsLogger

source

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

Bases: Callback

Log averaged training metrics every N steps as a single line.

This callback aggregates numeric metrics reported via Lightning’s trainer.callback_metrics during training and emits a compact, human-friendly log line at a fixed interval (log_every_n_steps). Metrics are averaged over the interval and the internal buffer is cleared after logging.

The log line is ordered as: : - Time metrics: iter_time, forward_time, backward_time, optim_step_time, train_time (when available)

• User metrics (stats from the model, train/ prefix removed)
• Learning rates (optim{idx}_lr{group})

Example

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

Example log output: : 20epoch🚋4201-4400batch: iter_time=6.212e-05, forward_time=0.145, backward_time=0.159, optim_step_time=0.015, train_time=0.562, loss_ctc=69.86, loss_att=32.128, acc=0.868, loss=46.669, optim0_lr0=2.458e-06

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.