2. HDF5

The NWB:N format currently uses the Hierarchical Data Format (HDF5) as the primary mechanism for data storage. HDF5 was selected for the NWB format because it met several of the project’s requirements. First, it is a mature data format standard with libraries available in multiple programming languages. Second, the format’s hierarchical structure allows data to be grouped into logical self-documenting sections. Its structure is analogous to a file system in which its “groups” and “datasets” correspond to directories and files. Groups and datasets can have attributes that provide additional details, such as authorities’ identifiers. Third, its linking feature enables data stored in one location to be transparently accessed from multiple locations in the hierarchy. The linked data can be external to the file. Fourth, HDF5 is widely supported across programming languages (e.g., C, C++, Python, MATLAB, R among others) and tools, such as, HDFView, a free, cross-platform application, can be used to open a file and browse data. Finally, ensuring the ongoing accessibility of HDF-stored data is the mission of The HDF Group, the nonprofit that is the steward of the technology.

2.1. Format Mapping

Here we describe the mapping of NWB primitives (e.g., Groups, Datasets, Attributes, Links, etc.) used by the NWB format and specification to HDF5 storage primitives. As the NWB:N format was designed with HDF5 in mind, the high-level mapping between the format specification and HDF5 is quite simple:

Table 2.1 Mapping of groups
NWB Primitive HDF5 Primitive
Group Group
Dataset Dataset
Attribute Attribute
Link Soft Link or External Link

Note

Using HDF5, NWB links are stored as HDF5 Soft Links or External Links. Hard Links are not used in NWB because the primary location and, hence, primary ownership and link path for secondary locations, cannot be determined for Hard Links.

2.2. Key Mapping

Here we describe the mapping of keys from the specification language to HDF5 storage objects:

2.2.1. Groups

Table 2.2 Mapping of groups
NWB Key HDF5
name Name of the Group in HDF5
doc HDF5 attribute doc on the HDF5 group
groups HDF5 groups within the HDF5 group
datasets HDF5 datasets within the HDF5 group
attributes HDF5 attributes on the HDF5 group
links HDF5 SoftLinks within the HDF5 group
linkable Not mapped; Stored in schema only
quantity Not mapped; Number of appearances of the dataset.
neurodata_type Attribute neurodata_type
namespace ID Attribute namespace
object ID Attribute object_id

2.2.2. Datasets

Table 2.3 Mapping of datasets
NWB Key HDF5
name Name of the dataset in HDF5
doc HDF5 attribute doc on the HDF5 dataset
dtype Data type of the HDF5 dataset (see dtype mappings table)
shape Shape of the HDF5 dataset if the shape is fixed, otherwise shape defines the maxshape
dims Not mapped
attributes HDF5 attributes on the HDF5 group
linkable Not mapped; Stored in schema only
quantity Not mapped; Number of appearances of the dataset.
neurodata_type Attribute neurodata_type
namespace ID Attribute namespace
object ID Attribute object_id

Note

  • TODO Update mapping of dims

2.2.3. Attributes

Table 2.4 Mapping of attributes
NWB Key HDF5
name Name of the attribute in HDF5
doc Not mapped; Stored in schema only
dtype Data type of the HDF5 attribute
shape Shape of the HDF5 dataset if the shape is fixed, otherwise shape defines the maxshape
dims Not mapped; Reflected by the shape of the attribute data
required Not mapped; Stored in schema only
value Data value of the attribute

2.2.5. dtype mappings

The mappings of data types is as follows

dtype spec value storage type size
  • “float”
  • “float32”
single precision floating point 32 bit
  • “double”
  • “float64”
double precision floating point 64 bit
  • “long”
  • “int64”
signed 64 bit integer 64 bit
  • “int”
  • “int32”
signed 32 bit integer 32 bit
  • “int16”
signed 16 bit integer 16 bit
  • “int8”
signed 8 bit integer 8 bit
  • “uint32”
unsigned 32 bit integer 32 bit
  • “uint16”
unsigned 16 bit integer 16 bit
  • “uint8”
unsigned 8 bit integer 8 bit
  • “bool”
boolean 8 bit
  • “text”
  • “utf”
  • “utf8”
  • “utf-8”
unicode variable
  • “ascii”
  • “str”
ascii variable
  • “ref”
  • “reference”
  • “object”
Reference to another group or dataset  
  • region
Reference to a region of another dataset  
  • compound dtype
HDF5 compound data type  
  • “isodatetime”
ASCII ISO8061 datetime string. For example 2018-09-28T14:43:54.123+02:00 variable

2.3. Caching format specifications

In practice it is useful to cache the specification a file was created with (including extensions) directly in the HDF5 file. Caching the specification in the file ensures that users can access the specification directly if necessary without requiring external resources. However, the mechanisms for caching format specifications is likely different for different storage backends and is not part of the NWB:N format specification itself. For the HDF5 backend, caching of the schema is implemented as follows.

The HDF5 backend adds the reserved top-level group /specifications in which all format specifications (including extensions) are cached. The /specifications group contains for each specification namespace a subgroup /specifications/<namespace-name>/<version> in which the specification for a particular version of a namespace are stored (e.g., /specifications/core/2.0.1 in the case of the NWB:N core namespace at version 2.0.1). The actual specification data is then stored as a JSON string in scalar datasets with a binary, variable-length string data type (e.g., dtype=special_dtype(vlen=binary_type) in Python). The specification of the namespace is stored in /specifications/<namespace-name>/<version>/namespace while additional source files are stored in /specifications/<namespace-name>/<version>/<source-filename>. Here <source-filename> refers to the main name of the source-file without file extension (e.g., the core namespace defines nwb.ephys.yaml as source which would be stored in /specifications/core/2.0.1/nwb.ecephys).