sakura.model_controllers.extractor_controller.ExtractorController

class sakura.model_controllers.extractor_controller.ExtractorController(model: Extractor, config: dict | None = None, pheno_config: dict | None = None, signature_config: dict | None = None, verbose=False)

Bases: object

Central controller class managing the complete training lifecycle of the Extractor model

This class provides the core controller for the Extractor model. It handles multi-component loss calculation, dynamic weight adjustment, distributed regularization, and training workflow management.

Parameters:

• model (models.Extractor) – Extractor model

• config (dict, optional) – experiment configuration dict

• pheno_config (dict, optional) – phenotype supervision configuration

• signature_config (dict, optional) – signature supervision configuration

• verbose (bool, optional) – Whether to enable verbose console logging, defaults to False

Methods

eval	Evaluate the model using the specified batch of data.
load_checkpoint	Restore training state of the extractor from given checkpoint.
loss	Calculate composite loss across all active components.
next_epoch	Handle inter epoch loss weight progressions and tick epoch counters.
print_weight_projection	Print the projected weights for main latent, phenotype, and signature groups over the expected epochs.
regularize	Handle regularization to the given tensor based on the specified configuration.
reset	Reset the trainer's state.
save_checkpoint	Save the current training state of the extractor.
select_item_dict	Utility method to handle selections from internal phenotype/signature configs.
select_loss_dict	Select loss keys from the internal dict based on selection setting.
setup_optimizer	Configure optimizer with filtered parameters based on exclusion rules.
tick	Increment the current tick counter.
train	Train the model using the spectrain(ified batch of data.