HDF5 Tensor File Format

Definition

Metavision SDK allows to store pre-computed event data in HDF5, a standard hierarchical self-describing file format that can be leveraged by our Machine Learning module.

Warning

HDF5 event files should not be confused with HDF5 event files that store events that are non-preprocessed.

HDF5 Tensor Files Usage

To create HDF5 tensor files, you can use the sample Generate HDF5 Tensor Files that allows you to convert one or multiple RAW or DAT files to precomputed tensor features following a preprocessing function of your choice.

The available preprocessing methods are presented in the Event Preprocessing Tutorial and the preprocessing SDK tools is also shown in the Precomputing features as HDF5 tensor files tutorial.

To visualize the content of an HDF5 tensor file, take a look at the sample Visualization of Preprocessed HDF5 Tensor Files.

Note

You can download some HDF5 tensor files in our Sample Recordings page in the section Precomputed Datasets.

Format of HDF5 tensor files

The structure of an HDF5 tensor file is organized as follows:

• precomputed event-based tensors are stored in an HDF5 dataset named data.

• metadata and parameters used during preprocessing are stored as associated attributes of data

Description and possible extensions of the format are shown in the section Exploring the format of HDF5 tensor files of our Precomputing features as HDF5 tensor files tutorial