minerva.pipelines.lightning_pipeline

Classes

SimpleLightningPipeline

Simple pipeline to train, test, predict and evaluate models using Pytorch

Functions

main()

Module Contents

class minerva.pipelines.lightning_pipeline.SimpleLightningPipeline(model, trainer, log_dir=None, save_run_status=True, classification_metrics=None, regression_metrics=None, apply_metrics_per_sample=False)

Bases: minerva.pipelines.base.Pipeline

Simple pipeline to train, test, predict and evaluate models using Pytorch Lightning. This class is intended to be seamlessly integrated with jsonargparse CLI.

Train/test/predict/evaluate a Pytorch Lightning model.

It provides 4 tasks: fit, test, predict and evaluate. The fit task trains the model, the test task evaluates the model on the test set, the predict task generates predictions for the predict set and the evaluate task evaluates the model on the predict set and returns the metrics.

The evaluate task can calculate classification and regression metrics, which is passed as arguments. The metrics are calculated per sample if apply_metrics_per_sample is True (that generate a metric for each), otherwise the metrics are calculated for the whole dataset (single metric). The last option is the default.

Parameters

modelL.LightningModule

The LightningModule to be used.

trainerL.Trainer

The Lightning Trainer to be used.

log_dirPathLike, optional

The default logging directory where all related pipeline files should be saved. By default None (uses current working directory)

save_run_statusbool, optional

If True, save the status of each run in a YAML file. This file will be saved in the working directory with the name run_{pipeline_id}.yaml. By default True.

classification_metricsDict[str, Metric], optional

The classification metrics to be used in the evaluate task. This dictionary should have the metric name as key and the torchmetrics.Metric-like object as value. The metric should be able to receive two tensors (y_true, y_pred) and return a tensor with the metric value. If None, no classification metrics will be calculated. Different from regression, the torch.argmax will be applied to the predictions before calculating the metrics. By default None.

regression_metricsDict[str, Metric], optional

The regression metrics to be used in the evaluate task. This dictionary should have the metric name as key and the torchmetrics.Metric-like object as value. The metric should be able to receive two tensors (y_true, y_pred) and return a tensor with the metric value. If None, no regression metrics will be calculated. By default None.

apply_metrics_per_samplebool, optional

Apply the metrics per sample. If True, the metrics will be calculated for each sample and the results will be a list of metrics. If False, the metrics will be calculated for the whole dataset and the results will be a single metric (single-element list). By default False

_calculate_metrics(metrics, y_hat, y)

Calculate the metrics for the given predictions and targets.

Parameters

metricsDict[str, Metric]

The metrics to be calculated. The dictionary should have the metric name as key and the torchmetrics.Metric-like object as value.

y_hattorch.Tensor

The predictions tensor.

ytorch.Tensor

The targets tensor.

Returns

Dict[str, Any]

A dictionary with the metric name as key and the list of metric values as value. The list will have a single element if apply_metrics_per_sample is False, otherwise it will have a value.

Parameters:

  • metrics (Dict[str, torchmetrics.Metric])

  • y_hat (torch.Tensor)

  • y (torch.Tensor)

Return type:

Dict[str, Any]

_evaluate(data, ckpt_path=None)

Evaluate the model and calculate regression and/or classification metrics.

Parameters

dataL.LightningDataModule

The data module to be used. The data module should have the predict_dataloader method implemented.

ckpt_pathPathLike

The checkpoint path to be used. If None, no checkpoint will be used.

Returns

Dict[str, Dict[str, Any]

A dictionary with metrics.

Parameters:

  • data (lightning.LightningDataModule)

  • ckpt_path (minerva.utils.typing.PathLike)

Return type:

Dict[str, Any]

_fit(data, ckpt_path)

Fit the model using the given data.

Parameters

dataL.LightningDataModule

The data module to be used. The data module should have the train_dataloader method implemented.

ckpt_pathPathLike

The checkpoint path to be used. If None, no checkpoint will be used.

Parameters:

  • data (lightning.LightningDataModule)

  • ckpt_path (minerva.utils.typing.PathLike)

_predict(data, ckpt_path=None)

Predict using the given data.

Parameters

dataL.LightningDataModule

The data module to be used. The data module should have the predict_dataloader method implemented.

ckpt_pathPathLike

The checkpoint path to be used. If None, no checkpoint will be used.

Returns

torch.Tensor

The predictions tensor.

Parameters:

  • data (lightning.LightningDataModule)

  • ckpt_path (minerva.utils.typing.PathLike)

Return type:

torch.Tensor

_run(data, task, ckpt_path=None)

Run the specified task on the given data.

Parameters

dataL.LightningDataModule

The LightningDataModule object containing the data for the task.

taskLiteral[“fit”, “test”, “predict”, “evaluate”], optional

The task to be performed. Valid options are “fit”, “test”, “predict”, and “evaluate”.

ckpt_pathPathLike, optional

The path to the checkpoint file to be used for resuming training or performing inference. Defaults to None.

Returns

Any

The result of the specified task.

Raises

ValueError

If an unknown task is provided.

Parameters:

  • data (lightning.LightningDataModule)

  • task (Literal['fit', 'test', 'predict', 'evaluate'])

  • ckpt_path (minerva.utils.typing.PathLike)

_test(data, ckpt_path)

Test the model using the given data.

Parameters

dataL.LightningDataModule

The data module to be used. The data module should have the test_dataloader method implemented.

ckpt_pathPathLike

The checkpoint path to be used. If None, no checkpoint will be used.

Parameters:

  • data (lightning.LightningDataModule)

  • ckpt_path (minerva.utils.typing.PathLike)

property data: lightning.LightningDataModule

The LightningDataModule used in the last run of the pipeline.

Returns

L.LightningDataModule

The data used in the last run of the pipeline.

Return type:

lightning.LightningDataModule

property model: lightning.LightningModule

The LightningModule used in the pipeline.

Returns

L.LightningModule

The model used in the pipeline.

Return type:

lightning.LightningModule

property trainer: lightning.Trainer

The Lightning Trainer used in the pipeline.

Returns

L.Trainer

The trainer used in the pipeline.

Return type:

lightning.Trainer

Parameters:

  • model (lightning.LightningModule)

  • trainer (lightning.Trainer)

  • log_dir (minerva.utils.typing.PathLike)

  • save_run_status (bool)

  • classification_metrics (Dict[str, torchmetrics.Metric])

  • regression_metrics (Dict[str, torchmetrics.Metric])

  • apply_metrics_per_sample (bool)

minerva.pipelines.lightning_pipeline.main()