ng_training_configuration

Training configuration.

`NGTrainingConfig` #

Bases: BaseModel

Parameters related to the training.

Attributes:

Name	Type	Description
`trainer_params`	`dict`	Parameters passed to the PyTorch Lightning Trainer class.
`logger`	`Literal['wandb', 'tensorboard'] \| None`	Additional Logger to use during training. If None, no logger will be used. Note that the `CAREamist` uses the `csv` logger regardless of the value of this field.
`checkpoint_params`	`dict[str, Any]`	Checkpoint callback parameters, following PyTorch Lightning Checkpoint callback.
`early_stopping_params`	`dict[str, Any] \| None`	Early stopping callback parameters, following PyTorch Lightning Checkpoint callback.

Source code in src/careamics/config/ng_configs/ng_training_configuration.py

class NGTrainingConfig(BaseModel):
    """
    Parameters related to the training.

    Attributes
    ----------
    trainer_params : dict
        Parameters passed to the PyTorch Lightning Trainer class.
    logger : Literal["wandb", "tensorboard"] | None
        Additional Logger to use during training. If None, no logger will be used.
        Note that the `CAREamist` uses the `csv` logger regardless of the value of this
        field.
    checkpoint_params : dict[str, Any]
        Checkpoint callback parameters, following PyTorch Lightning Checkpoint
        callback.
    early_stopping_params : dict[str, Any] | None
        Early stopping callback parameters, following PyTorch Lightning Checkpoint
        callback.
    """

    # Pydantic class configuration
    model_config = ConfigDict(
        validate_assignment=True,
    )

    trainer_params: dict = Field(default={})
    """Parameters passed to the PyTorch Lightning Trainer class"""

    logger: Literal["wandb", "tensorboard"] | None = None
    """Logger to use during training. If None, no logger will be used. Available
    loggers are defined in SupportedLogger."""

    # Only basic callbacks - they may have different defaults for different algorithms
    checkpoint_params: dict[str, Any] = Field(default_factory=dict)
    """Checkpoint callback parameters, following PyTorch Lightning Checkpoint
    callback."""

    early_stopping_params: dict[str, Any] | None = Field(default_factory=dict)
    """Early stopping callback parameters, following PyTorch Lightning Checkpoint
    callback."""

    def __str__(self) -> str:
        """Pretty string reprensenting the configuration.

        Returns
        -------
        str
            Pretty string.
        """
        return pformat(self.model_dump())

`checkpoint_params = Field(default_factory=dict)` `class-attribute` `instance-attribute` #

Checkpoint callback parameters, following PyTorch Lightning Checkpoint callback.

`early_stopping_params = Field(default_factory=dict)` `class-attribute` `instance-attribute` #

Early stopping callback parameters, following PyTorch Lightning Checkpoint callback.

`logger = None` `class-attribute` `instance-attribute` #

Logger to use during training. If None, no logger will be used. Available loggers are defined in SupportedLogger.

`trainer_params = Field(default={})` `class-attribute` `instance-attribute` #

Parameters passed to the PyTorch Lightning Trainer class

`str()` #

Pretty string reprensenting the configuration.

Returns:

Type	Description
`str`	Pretty string.

Source code in src/careamics/config/ng_configs/ng_training_configuration.py

def __str__(self) -> str:
    """Pretty string reprensenting the configuration.

    Returns
    -------
    str
        Pretty string.
    """
    return pformat(self.model_dump())

`SelfSupervisedCheckpointing` `dataclass` #

Presets for checkpointing Noise2Noise and Noise2Void.

Because self-supervised algorithms are evaluating the loss against noisy pixels, its value is not a good measure of performances after a few epochs. Therefore, it cannot be used to evaluate the best performing models.

This presets saves checkpoints every 10 epochs, as well as the last one.

Source code in src/careamics/config/ng_configs/ng_training_configuration.py

@dataclass
class SelfSupervisedCheckpointing:
    """Presets for checkpointing Noise2Noise and Noise2Void.

    Because self-supervised algorithms are evaluating the loss against noisy pixels,
    its value is not a good measure of performances after a few epochs. Therefore, it
    cannot be used to evaluate the best performing models.

    This presets saves checkpoints every 10 epochs, as well as the last one.
    """

    every_n_epochs: int = 10
    """Save a checkpoint every 10 epochs."""

    save_top_k: int = -1
    """Save all checkpoints. Checkpoints are checked `every_n_epochs`."""

    save_last: bool = True
    """Save the last checkpoint."""

    auto_insert_metric_name: bool = False
    """Do not insert the monitored value in the checkpoint name."""

`auto_insert_metric_name = False` `class-attribute` `instance-attribute` #

Do not insert the monitored value in the checkpoint name.

`every_n_epochs = 10` `class-attribute` `instance-attribute` #

Save a checkpoint every 10 epochs.

`save_last = True` `class-attribute` `instance-attribute` #

Save the last checkpoint.

`save_top_k = -1` `class-attribute` `instance-attribute` #

Save all checkpoints. Checkpoints are checked every_n_epochs.

`SupervisedCheckpointing` `dataclass` #

Presets for checkpointing CARE.

This preset saves the top 3 best performing checkpoints based on val_loss, as well as the last one.

Source code in src/careamics/config/ng_configs/ng_training_configuration.py

@dataclass
class SupervisedCheckpointing:
    """Presets for checkpointing CARE.

    This preset saves the top 3 best performing checkpoints based on `val_loss`, as well
    as the last one.
    """

    monitor: str = "val_loss"
    """Monitor `val_loss`."""

    mode: str = "min"
    """Top checkpoints are selected by minimum `val_loss`."""

    save_top_k: int = 3
    """Save the top 3 best performing checkpoints."""

    save_last: bool = True
    """Save the last checkpoint."""

    auto_insert_metric_name: bool = False
    """Do not insert the monitored value in the checkpoint name."""

`auto_insert_metric_name = False` `class-attribute` `instance-attribute` #

Do not insert the monitored value in the checkpoint name.

`mode = 'min'` `class-attribute` `instance-attribute` #

Top checkpoints are selected by minimum val_loss.

`monitor = 'val_loss'` `class-attribute` `instance-attribute` #

Monitor val_loss.

`save_last = True` `class-attribute` `instance-attribute` #

Save the last checkpoint.

`save_top_k = 3` `class-attribute` `instance-attribute` #

Save the top 3 best performing checkpoints.

`default_training_dict(algorithm, trainer_params=None, logger='none', checkpoint_params=None, monitor_metric='val_loss')` #

Default training configuration constructor.

This function sets default training parameters based on the algorithm configuration. If the user provides any of the parameters, they will take precedence over the defaults.

Parameters:

Name	Type	Description	Default
`algorithm`	`('care', 'n2n', 'n2v')`	Algorithm type, used to select the default checkpointing preset.	`"care"`
`trainer_params`	`dict`	Parameters for Lightning Trainer class, by default None.	`None`
`logger`	`('wandb', 'tensorboard', 'none')`	Logger to use, by default "none".	`"wandb"`
`checkpoint_params`	`dict`	Parameters for the checkpoint callback, by default None. If None, then default parameters are applied based on the algorithm.	`None`
`monitor_metric`	`str`	Metric to monitor for early stopping, by default "val_loss".	`'val_loss'`

Returns:

Type	Description
`dict`	Training configuration dictionary with the specified parameters.

Source code in src/careamics/config/ng_configs/ng_training_configuration.py

def default_training_dict(
    algorithm: Literal["care", "n2n", "n2v"],
    trainer_params: dict[str, Any] | None = None,
    logger: Literal["wandb", "tensorboard", "none"] = "none",
    checkpoint_params: dict[str, Any] | None = None,
    monitor_metric: str = "val_loss",
) -> dict:
    """Default training configuration constructor.

    This function sets default training parameters based on the algorithm configuration.
    If the user provides any of the parameters, they will take precedence over the
    defaults.

    Parameters
    ----------
    algorithm : {"care", "n2n", "n2v"}
        Algorithm type, used to select the default checkpointing preset.
    trainer_params : dict, optional
        Parameters for Lightning Trainer class, by default None.
    logger : {"wandb", "tensorboard", "none"}, optional
        Logger to use, by default "none".
    checkpoint_params : dict, optional
        Parameters for the checkpoint callback, by default None. If None, then default
        parameters are applied based on the algorithm.
    monitor_metric : str, optional
        Metric to monitor for early stopping, by default "val_loss".

    Returns
    -------
    dict
        Training configuration dictionary with the specified parameters.
    """
    # user parameters take precedence over defaults
    # since resulting checkpointing behaviour depends on complex interactions between
    # parameters, we keep either user defined or the defaults
    if checkpoint_params is None:
        # select default checkpointing preset based on algorithm
        default_preset = (
            SupervisedCheckpointing
            if algorithm == "care"
            # since Noise2Noise is comparing noisy pixels to other noisy pixels, it
            # cannot be monitored based on a metric, we use the self-supervised preset
            else SelfSupervisedCheckpointing
        )
        default_checkpoint = asdict(default_preset())
        checkpoint_params = default_checkpoint

    return {
        "trainer_params": {} if trainer_params is None else trainer_params,
        "logger": None if logger == "none" else logger,
        "checkpoint_params": checkpoint_params,
        "early_stopping_params": {
            "monitor": monitor_metric,
            "mode": "min",
        },
    }

`default_training_factory(validated_dict)` #

Default training configuration constructor.

Parameters:

Name	Type	Description	Default
`validated_dict`	`dict`	Validated configuration dictionary, used to set default training parameters based on the algorithm configuration. This is expected to be passed by Pydantic when calling the default constructor.	required

Returns:

Type	Description
`NGTrainingConfig`	Training configuration with the specified parameters.

Source code in src/careamics/config/ng_configs/ng_training_configuration.py

def default_training_factory(validated_dict: dict[str, Any]) -> NGTrainingConfig:
    """Default training configuration constructor.

    Parameters
    ----------
    validated_dict : dict
        Validated configuration dictionary, used to set default training parameters
        based on the algorithm configuration. This is expected to be passed by Pydantic
        when calling the default constructor.

    Returns
    -------
    NGTrainingConfig
        Training configuration with the specified parameters.
    """
    key = "algorithm_config"

    if key not in validated_dict:
        raise ValueError(
            "Algorithm configuration is required to set default training parameters."
        )
    algorithm = validated_dict[key].algorithm

    # N2V algorithm has a monitor_metric parameter
    monitor_metric = getattr(validated_dict[key], "monitor_metric", "val_loss")

    return NGTrainingConfig(
        **default_training_dict(
            algorithm=algorithm,
            monitor_metric=monitor_metric,
        )
    )

ng_training_configuration