torch_brain.dataset

Base classes for creating PyTorch datasets

Module Overview

This module contains base classes to ease creation of PyTorch datasets for your data.

• The Dataset class is inherited by all datasets. These handle opening and accessing single datasets.

• The NestedDataset class is for opening and accessing multiple datasets through a unified interface.

• Mixin classes are provided to add modality-specific functionalities to the Dataset classes.

Dataset

torch_brain’s Dataset class (and its sub-classes) allow you to sample time-slices of your data. This is a major deviation from the standard torch.utils.data.Dataset, which is indexed by integers. To achieve arbitrary time-slice based access, our Dataset class is indexed by three things:

1. The recording id from which you want the slice,

2. Start time of the slice, and

3. End time of the slice

These are put into a DatasetIndex object, which is then used to index the Dataset. Since different machine learning applications require different ways of sampling, we provide a collection of samplers which are responsible for creating these DatasetIndex objects.

NestedDataset

The Dataset class is designed to operate on a single dataset. However, many modern ML methods perform training over multiple datasets. For this, we provide NestedDataset that allows users to open and index through multiple datasets.

Base Classes

Dataset	PyTorch Dataset for loading time-slices of neural data recordings from HDF5 files.
DatasetIndex	Index for accessing a specific time interval of a recording within a Dataset.
NestedDataset	Dataset that composes multiple Dataset instances under a single interface.

Mixins

SpikingDatasetMixin	Mixin class for torch_brain.dataset.Dataset subclasses containing spiking data.
CalciumImagingDatasetMixin	Mixin class for torch_brain.dataset.Dataset subclasses containing calcium imaging data.
MultiChannelDatasetMixin	Mixin class for torch_brain.dataset.Dataset subclasses containing multi-channel recordings (e.g., EEG, ECoG, EMG, sEEG, etc).