Callback that terminates training when a NaN loss is encountered.

This callback monitors the loss during training and terminates training when a NaN or Inf loss is detected. By default, training stops gracefully by setting the model's stop_training flag, which allows callback cleanup methods such as on_train_end() to run.

Set raise_error = TRUE to raise an error immediately when a NaN or Inf is detected. In this mode, on_train_end() is not called on other callbacks. This can preserve backup states or prevent unintended cleanup after a training failure.

Usage

callback_terminate_on_nan(raise_error = FALSE)

Arguments

raise_error

If FALSE, stop training gracefully. If TRUE, raise an error immediately when a NaN or Inf loss is detected, bypassing callback cleanup methods.

Value

A Callback instance that can be passed to fit.keras.src.models.model.Model().

Examples

# Graceful termination (default)
callback <- callback_terminate_on_nan()
model |> fit(x, y, callbacks = list(callback))

# Immediate error
callback <- callback_terminate_on_nan(raise_error = TRUE)
model |> fit(x, y, callbacks = list(callback))

See also

• https://keras.io/api/callbacks/terminate_on_nan#terminateonnan-class

Other callbacks:
Callback()
callback_backup_and_restore()
callback_csv_logger()
callback_early_stopping()
callback_lambda()
callback_learning_rate_scheduler()
callback_model_checkpoint()
callback_orbax_checkpoint()
callback_reduce_lr_on_plateau()
callback_remote_monitor()
callback_swap_ema_weights()
callback_tensorboard()