peal.architectures.interfaces.TaskConfig¶

class peal.architectures.interfaces.TaskConfig[source]¶

Bases: BaseModel

A dict of critirion names (that have to be implemented in peal.training.criterions) mapped to the weight. Like this the loss function can be post_hoc attached without changing the code.

Methods

Attributes

`bce_mse_mix`	The mix between binary cross entropy loss and mse loss.
`class_restriction`	Gives the option to only use samples from one or more specified classes
`config_name`	The type of config.
`criterions`	The criterions used for training the model.
`focal_alpha`	The alpha parameter of the focal loss.
`focal_gamma`	The gamma parameter of the focal loss.
`focal_mse_mix`	The mix between focal loss and mse loss.
`mixed_bce_pos_weight`	The pos_weight parameter of the BCEWithLogitsLoss used for mixed BCE and MSE loss.
`output_channels`	The output_channels that can be at most the output_channels of the dataset, but if a subtask is chosen the output_channels has also be adapted accordingly
`output_type`	The output_type that either can just be the output_type of the dataset or could be some possible subtype.
`x_selection`	Gives the option to select a subset of the input variables.
`y_selection`	Gives the option to select a subset of the output_variables.

config_name: str = 'TaskConfig'¶: The type of config. This is necessary to find config class from yaml config

criterions: dict = {'ce': 1.0, 'l2': 100.0}¶: The criterions used for training the model.

focal_alpha: float = 0.25¶: The alpha parameter of the focal loss.

focal_gamma: float = 1.0¶: The gamma parameter of the focal loss.

focal_mse_mix: float = 1.0¶: The mix between focal loss and mse loss.

mixed_bce_pos_weight: float = 1.0¶: The pos_weight parameter of the BCEWithLogitsLoss used for mixed BCE and MSE loss.

bce_mse_mix: float = 1.0¶: The mix between binary cross entropy loss and mse loss.

output_type: str = 'singleclass'¶: The output_type that either can just be the output_type of the dataset or could be some possible subtype. E.g. when having a binary multiclass dataset one could use as task binary single class classification for one of the output variables.

output_channels: pydantic.PositiveInt = 2¶: The output_channels that can be at most the output_channels of the dataset, but if a subtask is chosen the output_channels has also be adapted accordingly

x_selection: list[str] | None = None¶: Gives the option to select a subset of the input variables. Only works for symbolic data.

y_selection: list[str] | None = None¶: Gives the option to select a subset of the output_variables. Can be used e.g. to transform binary multiclass into the subtask of predicting one of the binary variables with single class classification.

class_restriction: int | list[int] | None = None¶: Gives the option to only use samples from one or more specified classes