tf.keras.callbacks.OrbaxCheckpoint

Callback to save and load model state using Orbax with a similar API to ModelCheckpoint.

Inherits From: Callback

tf.keras.callbacks.OrbaxCheckpoint(
    directory,
    monitor='val_loss',
    verbose=0,
    save_best_only=False,
    mode='auto',
    save_freq='epoch',
    initial_value_threshold=None,
    max_to_keep=1,
    save_on_background=True,
    save_weights_only=False
)

This callback saves the model's weights and optimizer state asynchronously using Orbax, allowing training to continue without blocking for I/O.

Multi-host Support: When running in a multi-host distributed training environment with JAX backend, this callback automatically coordinates checkpointing across all hosts to ensure consistency and proper synchronization. Multi-host checkpointing is only supported on JAX.

Example:

model.compile(loss=..., optimizer=..., metrics=['accuracy'])

EPOCHS = 10
checkpoint_dir = '/tmp/ckpt'
orbax_checkpoint_callback = keras.callbacks.OrbaxCheckpoint(
    directory=checkpoint_dir,
    monitor='val_accuracy',
    mode='max',
    save_best_only=True)

# Model is saved at the end of every epoch, if it's the best seen so far.
model.fit(epochs=EPOCHS, callbacks=[orbax_checkpoint_callback])

# Alternatively, save checkpoints every N batches -
orbax_checkpoint_callback = keras.callbacks.OrbaxCheckpoint(
    directory=checkpoint_dir,
    save_freq=100)  # Save every 100 batches

model.fit(epochs=EPOCHS, callbacks=[orbax_checkpoint_callback])

Args
directory	path to the directory where to save the checkpoints.
monitor	The metric name to monitor (e.g., 'val_loss').
verbose	Verbosity mode, 0 or 1.
save_best_only	if save_best_only=True, it only saves when the model is considered the "best" based on the monitored quantity.
mode	one of {'auto', 'min', 'max'}. Used with save_best_only.
save_freq	'epoch' or integer. Frequency to save checkpoints.
max_to_keep	Integer, maximum number of recent checkpoints to keep. If None, keeps all. Defaults to 1.
save_on_background	Boolean, whether to save asynchronously in the background. Defaults to True.
initial_value_threshold	Floating point initial "best" value for the monitor, used with save_best_only.

Attributes
model

Methods

is_multihost_enabled

View source

is_multihost_enabled()

Return True if multi-host checkpointing is enabled and initialized.

This method can be used to check if the callback is operating in a multi-host distributed training environment. Multi-host checkpointing is only supported on JAX backend.

Returns
bool	True if multi-host support is active, False otherwise.

is_primary_host

View source

is_primary_host()

Return True if this process is the primary host in multi-host setup.

In multi-host environments, only the primary host typically handles logging and coordination tasks. Multi-host checkpointing is only supported on JAX backend.

Returns
bool	True if this is the primary host, False otherwise. Always returns True in single-host environments.

on_batch_begin

View source

on_batch_begin(
    batch, logs=None
)

A backwards compatibility alias for on_train_batch_begin.

on_batch_end

View source

on_batch_end(
    batch, logs=None
)

A backwards compatibility alias for on_train_batch_end.

on_epoch_begin

View source

on_epoch_begin(
    epoch, logs=None
)

Called at the start of an epoch.

Subclasses should override for any actions to run. This function should only be called during TRAIN mode.

Args
epoch	Integer, index of epoch.
logs	Dict. Currently no data is passed to this argument for this method but that may change in the future.

on_epoch_end

View source

on_epoch_end(
    epoch, logs=None
)

Called at the end of an epoch.

Subclasses should override for any actions to run. This function should only be called during TRAIN mode.

Args
epoch	Integer, index of epoch.
logs	Dict, metric results for this training epoch, and for the validation epoch if validation is performed. Validation result keys are prefixed with val_. For training epoch, the values of the Model's metrics are returned. Example: {'loss': 0.2, 'accuracy': 0.7}.

on_predict_batch_begin

View source

on_predict_batch_begin(
    batch, logs=None
)

Called at the beginning of a batch in predict methods.

Subclasses should override for any actions to run.

Note that if the steps_per_execution argument to compile in Model is set to N, this method will only be called every N batches.

Args
batch	Integer, index of batch within the current epoch.
logs	Dict. Currently no data is passed to this argument for this method but that may change in the future.

on_predict_batch_end

View source

on_predict_batch_end(
    batch, logs=None
)

Called at the end of a batch in predict methods.

Subclasses should override for any actions to run.

Note that if the steps_per_execution argument to compile in Model is set to N, this method will only be called every N batches.

Args
batch	Integer, index of batch within the current epoch.
logs	Dict. Aggregated metric results up until this batch.

on_predict_begin

View source

on_predict_begin(
    logs=None
)

Called at the beginning of prediction.

Subclasses should override for any actions to run.

Args
logs	Dict. Currently no data is passed to this argument for this method but that may change in the future.

on_predict_end

View source

on_predict_end(
    logs=None
)

Called at the end of prediction.

Subclasses should override for any actions to run.

Args
logs	Dict. Currently no data is passed to this argument for this method but that may change in the future.

on_test_batch_begin

View source

on_test_batch_begin(
    batch, logs=None
)

Called at the beginning of a batch in evaluate methods.

Also called at the beginning of a validation batch in the fit methods, if validation data is provided.

Subclasses should override for any actions to run.

Note that if the steps_per_execution argument to compile in Model is set to N, this method will only be called every N batches.

Args
batch	Integer, index of batch within the current epoch.
logs	Dict. Currently no data is passed to this argument for this method but that may change in the future.

on_test_batch_end

View source

on_test_batch_end(
    batch, logs=None
)

Called at the end of a batch in evaluate methods.

Also called at the end of a validation batch in the fit methods, if validation data is provided.

Subclasses should override for any actions to run.

Note that if the steps_per_execution argument to compile in Model is set to N, this method will only be called every N batches.

Args
batch	Integer, index of batch within the current epoch.
logs	Dict. Aggregated metric results up until this batch.

on_test_begin

View source

on_test_begin(
    logs=None
)

Called at the beginning of evaluation or validation.

Subclasses should override for any actions to run.

Args
logs	Dict. Currently no data is passed to this argument for this method but that may change in the future.

on_test_end

View source

on_test_end(
    logs=None
)

Called at the end of evaluation or validation.

Subclasses should override for any actions to run.

Args
logs	Dict. Currently the output of the last call to on_test_batch_end() is passed to this argument for this method but that may change in the future.

on_train_batch_begin

View source

on_train_batch_begin(
    batch, logs=None
)

Called at the beginning of a training batch in fit methods.

Subclasses should override for any actions to run.

Note that if the steps_per_execution argument to compile in Model is set to N, this method will only be called every N batches.

Args
batch	Integer, index of batch within the current epoch.
logs	Dict. Currently no data is passed to this argument for this method but that may change in the future.

on_train_batch_end

View source

on_train_batch_end(
    batch, logs=None
)

Called at the end of a training batch in fit methods.

Subclasses should override for any actions to run.

Note that if the steps_per_execution argument to compile in Model is set to N, this method will only be called every N batches.

Args
batch	Integer, index of batch within the current epoch.
logs	Dict. Aggregated metric results up until this batch.

on_train_begin

View source

on_train_begin(
    logs=None
)

Called at the beginning of training.

Subclasses should override for any actions to run.

Args
logs	Dict. Currently no data is passed to this argument for this method but that may change in the future.

on_train_end

View source

on_train_end(
    logs=None
)

Called at the end of training.

Subclasses should override for any actions to run.

Args
logs	Dict. Currently the output of the last call to on_epoch_end() is passed to this argument for this method but that may change in the future.

set_model

View source

set_model(
    model
)

set_params

View source

set_params(
    params
)

wait_until_finished

View source

wait_until_finished()

Wait for any in-progress checkpoint operations to complete. This method blocks until all asynchronous checkpoint save operations have completed across all hosts in a multi-host setup.