API Reference¶

Core¶

`Data`	Create dataloaders for brain-modeling experiments.
`Experiment`	Brain-modeling experiment with support for loading pretrained weights.
`BenchmarkAggregator`	Orchestrate multiple `Experiment` runs and visualise results.

get_default_dataloaders

Return the train/val/test DataLoaders for a task's default data config.

BrainModule

Pytorch-lightning module for M/EEG model training.

Running the benchmark¶

Three entry points onto the same experiments – the neuralbench command, the same selections from Python, and one model you built over a selection. They are compared in the quickstart.

`run_benchmark_cli`	CLI entry point for `neuralbench`.
`run_benchmark`	Run one or more NeuralBench experiments from Python.
`check_model`	Push synthetic batches of a selection's shapes through model.
`evaluate_model`	Run model on a selection of tasks and return the results.

Events Transforms¶

`TextPreprocessor`	Clean and filter text-related events.
`SklearnSplit`	Perform train/val/test split using sklearn's `train_test_split`.
`SimilaritySplit`	Perform train/val/test split based on similarity of sentence events.
`PredefinedSplit`	Assign train/test labels based on a predefined split, and optionally split train into validation as well.
`CropSleepRecordings`	Keep up to max_wake_duration_min mins of wake (W) time before and after the first and last sleep events.
`CropTimelines`	Crop neuro timelines.
`AddDefaultEvents`	Add default events to a timeline to fill out its duration.
`OffsetEvents`	Offset selected events by specified amounts.

Callbacks¶

`TestFullRetrievalMetrics`	Accumulate predictions on entire test set before evaluating metrics.
`RecordingLevelEval`	Callback to evaluate average prediction over each recording (timeline).
`PlotConfusionMatrix`
`PlotRegressionVectors`	Visualize predictions vs ground truth for multi-dimensional regression tasks.
`plot_confusion_matrix`	Plot confusion matrix/matrices with a shared colorbar.

Utilities¶

`TrainerConfig`	Joint configuration for Trainer and some callbacks.
`load_checkpoint`	Load checkpoint through state_dicts.
`compute_class_weights_from_dataset`	Compute class weights from training dataset for handling class imbalance.
`make_weighted_sampler`	Create a weighted random sampler for the given dataset to handle class imbalance.

Modules¶

`DownstreamWrapper`	Configuration for wrapping a (pretrained) model for downstream fine-tuning or linear probing.
`DownstreamWrapperModel`	Wrapper for downstream evaluation of pretrained models.

Configuration¶

`setup_config`	Set up neuralbench configuration.
`load_config`	Load neuralbench configuration.
`get_config`	Get the current configuration, loading it if necessary.