callbacks

Callbacks for training.

Classes:

Name	Description
EarlyStopping	Initialize the EarlyStopping callback.
Scheduler	Initialize a scheduler object to encapsulate different torch schedulers.
ModelCheckpoint	Model checkpoint initialization.

Attributes:

Name	Type	Description
Callback

Callback = EarlyStopping | Scheduler | ModelCheckpoint

EarlyStopping

Initialize the EarlyStopping callback.

The given monitoring metric will be used on the validation set to evaluate the training process. If no validation set provided when calling the Trainer.train method, the monitoring metric will be used on training set and can lead to an overfitted trained model.

Note: required at least 1 pruning step to be performed to be activated. If the model was not pruned yet, the training won't stop.

Parameters:

Name	Type	Description	Default
monitoring_metric	str	Monitoring metric required for early stopping, for instance "Total loss", which is the sum between the mean explain loss and the mean loss. The specified metric is by default computed on the validation set. If no validation set is provided, the only available metric is "Total loss", the sum between the mean explain loss and the mean loss on the train set.	required
mode	Literal['maximize', 'minimize']	Whether to "maximize" or "minimize" the provided metric.	required
patience	int	Number of epochs to wait with no improvement of the monitoring value.	3.
min_delta	float	Minimum delta between two monitoring values to consider an improvement.	0.

Attributes:

Name	Type	Description
type_	str

Methods:

Name	Description
to_model	To callback model.

type_: str = field(init=False, default='EARLY_STOPPING')

mode: Literal['maximize', 'minimize'] = field(validator=(lambda self, attribute, value: value in [(str(mode_)) for mode_ in EarlyStoppingTrainCallbackTrainCallbackMode]), converter=EarlyStoppingTrainCallbackTrainCallbackMode)

monitoring_metric: str

patience: int = 3

min_delta: float = 0.0

to_model() -> EarlyStoppingTrainCallback

To callback model.

Source code in src/xpdeep/trainer/callbacks.py

def to_model(self) -> EarlyStoppingTrainCallback:
    """To callback model."""
    return EarlyStoppingTrainCallback(**_attrs_asdict(self))

Scheduler

Initialize a scheduler object to encapsulate different torch schedulers.

Parameters:

Name	Type	Description	Default
pre_scheduler	partial[LRScheduler]	Based torch lr scheduler to be instantiated. Should not contain the optimizer as xpdeep use the trainer's optimizer for the scheduler internally.	required
step_method	Literal['batch', 'epoch']	"epoch" or "batch".	required
monitoring_metric	str	Monitoring metric required for the step method, for instance "Total loss", which is the sum between the mean explain loss and the mean loss. The specified metric is by default computed on the validation set. If no validation set is provided, the only available metric is "Total loss", the sum between the mean explain loss and the mean loss on the train set.	required

Attributes:

Name	Type	Description
type_	str

Methods:

Name	Description
to_model	As SchedulerRequestBody.

type_: str = field(init=False, default='SCHEDULER')

pre_scheduler: partial[LRScheduler | ReduceLROnPlateau]

step_method: Literal['batch', 'epoch'] = field(validator=(lambda self, attribute, value: value in [(str(step_method_)) for step_method_ in SchedulerTrainCallbackInputSchedulerTrainCallbackStepMethod]), converter=SchedulerTrainCallbackInputSchedulerTrainCallbackStepMethod)

monitoring_metric: str

to_model() -> SchedulerTrainCallbackInput

As SchedulerRequestBody.

Source code in src/xpdeep/trainer/callbacks.py

def to_model(self) -> SchedulerTrainCallbackInput:
    """As SchedulerRequestBody."""
    self_as_dict = _attrs_asdict(self)

    self_as_dict["pre_scheduler"] = TrustedObjectInput.from_dict({
        "reconstructor": "partial",
        "class_": self.pre_scheduler.func.__name__,
        "module": self.pre_scheduler.func.__module__,
        "state": {
            "args": self.pre_scheduler.args,
            "kwargs": self.pre_scheduler.keywords,
        },
    })
    self_as_dict["step_method"] = SchedulerTrainCallbackInputSchedulerTrainCallbackStepMethod(self.step_method)

    return SchedulerTrainCallbackInput(**self_as_dict)

ModelCheckpoint

Model checkpoint initialization.

Parameters:

Name	Type	Description	Default
monitoring_metric	str	Monitoring metric required for the step method, for instance "Total loss", which is the sum between the mean explain loss and the mean loss. The specified metric is by default computed on the validation set. If no validation set is provided, the only available metric is "Total loss", the sum between the mean explain loss and the mean loss on the train set.	required
save_every_epoch	int	How often to save the model. If None, only save the best checkpoint.	1
mode	Literal['maximize', 'minimize']	Whether to "maximize" or "minimize" the provided metric.	required

Attributes:

Name	Type	Description
type_	str

Methods:

Name	Description
to_model	Parse the object to a RequestBody for the backend.

type_: str = field(init=False, default='MODEL_CHECKPOINT')

monitoring_metric: str

mode: Literal['maximize', 'minimize'] = field(validator=(lambda self, attribute, value: value in [(str(mode_)) for mode_ in ModelCheckpointTrainCallbackTrainCallbackMode]), converter=ModelCheckpointTrainCallbackTrainCallbackMode)

save_every_epoch: int = 1

to_model() -> ModelCheckpointTrainCallback

Parse the object to a RequestBody for the backend.

Source code in src/xpdeep/trainer/callbacks.py

def to_model(self) -> ModelCheckpointTrainCallback:
    """Parse the object to a RequestBody for the backend."""
    self_as_dict = _attrs_asdict(self)
    self_as_dict["mode"] = ModelCheckpointTrainCallbackTrainCallbackMode(self.mode)

    return ModelCheckpointTrainCallback(**self_as_dict)