qumphy.trainer module

File: qumphy/trainer.py Project: 22HLT01 QUMPHY Contact: oskar.pfeffer@ptb.de Gitlab: https://gitlab.com/qumphy Description: Lightning Trainer.

class qumphy.trainer.Trainer(config)

Bases: object

Training pipeline wrapper for QuMPhy experiments.

This class loads the trainer, data module, feature extractor, model, and optional ensemble configuration from a configuration dictionary.

add_parameters_to_experiment()

base_config()

best_model_path(trainer)

Get the best checkpoint path from a trainer.

Parameters:

trainer (lightning.Trainer) – Lightning trainer whose callbacks are searched.

Returns:

Best model checkpoint path if a ModelCheckpoint callback is found.

Return type:

str or None

extract_features()

Extract features from the train, validation, and test datasets.

Returns:

The function replaces each dataset’s data with extracted features and keeps the corresponding labels.

Return type:

None

find_lr()

fit()

Train the model or ensemble members.

Returns:

The function starts the Lightning training loop.

Return type:

None

load_data_module()

Load data module from config.

load_ensemble_model()

Load trained ensemble members into an ensemble model.

Returns:

The function loads the best checkpoint for each ensemble member and stores the combined ensemble model.

Return type:

None

load_feature_extractor()

load_model()

load_trainer()

load_tuner()

Load the Lightning tuner.

Returns:

The function creates a tuner if it does not already exist.

Return type:

None

predict()

Predict using the best model and save the predictions to “best_predictions.pt”.

seed_everything()

Set random seeds for reproducibility.

Returns:

The function calls the Lightning seed utility.

Return type:

None

set_save_dir()

The save directory and project name are passed to the config of the model_checkpoint and logger.

set_sweep_parameters()

Set sweep parameters inside the nested configuration.

Returns:

The function modifies the configuration dictionary in place.

Return type:

None

test()