HDF5 for Python

The h5py package is a Pythonic interface to the HDF5 binary data format.

HDF5 lets you store huge amounts of numerical data, and easily manipulate that data from NumPy. For example, you can slice into multi-terabyte datasets stored on disk, as if they were real NumPy arrays. Thousands of datasets can be stored in a single file, categorized and tagged however you want.

Introductory info

Quick Start Guide

Install

With Anaconda or Miniconda:

conda install h5py

With Enthought Canopy, use the GUI package manager or:

enpkg h5py

With pip or setup.py, see Installation.

Core concepts

An HDF5 file is a container for two kinds of objects: datasets, which are array-like collections of data, and groups, which are folder-like containers that hold datasets and other groups. The most fundamental thing to remember when using h5py is:

Groups work like dictionaries, and datasets work like NumPy arrays

Suppose someone has sent you a HDF5 file, mytestfile.hdf5. (To create this file, read Appendix: Creating a file.) The very first thing you’ll need to do is to open the file for reading:

>>> import h5py
>>> f = h5py.File('mytestfile.hdf5', 'r')

The File object is your starting point. What is stored in this file? Remember h5py.File acts like a Python dictionary, thus we can check the keys,

>>> list(f.keys())
['mydataset']

Based on our observation, there is one data set, mydataset in the file. Let us examine the data set as a Dataset object

>>> dset = f['mydataset']

The object we obtained isn’t an array, but an HDF5 dataset. Like NumPy arrays, datasets have both a shape and a data type:

>>> dset.shape
(100,)
>>> dset.dtype
dtype('int32')

They also support array-style slicing. This is how you read and write data from a dataset in the file:

>>> dset[...] = np.arange(100)
>>> dset[0]
0
>>> dset[10]
10
>>> dset[0:100:10]
array([ 0, 10, 20, 30, 40, 50, 60, 70, 80, 90])

For more, see File Objects and Datasets.

Appendix: Creating a file

At this point, you may wonder how mytestdata.hdf5 is created. We can create a file by setting the mode to w when the File object is initialized. Some other modes are a (for read/write/create access), and r+ (for read/write access). A full list of file access modes and their meanings is at File Objects.

>>> import h5py
>>> import numpy as np
>>> f = h5py.File("mytestfile.hdf5", "w")

The File object has a couple of methods which look interesting. One of them is create_dataset, which as the name suggests, creates a data set of given shape and dtype

>>> dset = f.create_dataset("mydataset", (100,), dtype='i')

The File object is a context manager; so the following code works too

>>> import h5py
>>> import numpy as np
>>> with h5py.File("mytestfile.hdf5", "w") as f:
>>>     dset = f.create_dataset("mydataset", (100,), dtype='i')

Groups and hierarchical organization

“HDF” stands for “Hierarchical Data Format”. Every object in an HDF5 file has a name, and they’re arranged in a POSIX-style hierarchy with /-separators:

>>> dset.name
u'/mydataset'

The “folders” in this system are called groups. The File object we created is itself a group, in this case the root group, named /:

>>> f.name
u'/'

Creating a subgroup is accomplished via the aptly-named create_group. But we need to open the file in the “append” mode first (Read/write if exists, create otherwise)

>>> f = h5py.File('mydataset.hdf5', 'a')
>>> grp = f.create_group("subgroup")

All Group objects also have the create_* methods like File:

>>> dset2 = grp.create_dataset("another_dataset", (50,), dtype='f')
>>> dset2.name
u'/subgroup/another_dataset'

By the way, you don’t have to create all the intermediate groups manually. Specifying a full path works just fine:

>>> dset3 = f.create_dataset('subgroup2/dataset_three', (10,), dtype='i')
>>> dset3.name
u'/subgroup2/dataset_three'

Groups support most of the Python dictionary-style interface. You retrieve objects in the file using the item-retrieval syntax:

>>> dataset_three = f['subgroup2/dataset_three']

Iterating over a group provides the names of its members:

>>> for name in f:
...     print name
mydataset
subgroup
subgroup2

Membership testing also uses names:

>>> "mydataset" in f
True
>>> "somethingelse" in f
False

You can even use full path names:

>>> "subgroup/another_dataset" in f
True

There are also the familiar keys(), values(), items() and iter() methods, as well as get().

Since iterating over a group only yields its directly-attached members, iterating over an entire file is accomplished with the Group methods visit() and visititems(), which take a callable:

>>> def printname(name):
...     print name
>>> f.visit(printname)
mydataset
subgroup
subgroup/another_dataset
subgroup2
subgroup2/dataset_three

For more, see Groups.

Attributes

One of the best features of HDF5 is that you can store metadata right next to the data it describes. All groups and datasets support attached named bits of data called attributes.

Attributes are accessed through the attrs proxy object, which again implements the dictionary interface:

>>> dset.attrs['temperature'] = 99.5
>>> dset.attrs['temperature']
99.5
>>> 'temperature' in dset.attrs
True

For more, see Attributes.

Installation

It is highly recommended that you use a pre-built version of h5py, either from a Python Distribution, an OS-specific package manager, or a pre-built wheel from PyPI.

Be aware however that most pre-built versions lack MPI support, and that they are built against a specific version of HDF5. If you require MPI support, or newer HDF5 features, you will need to build from source.

After installing h5py, you should run the tests to be sure that everything was installed correctly. This can be done in the python interpreter via:

import h5py
h5py.run_tests()

On Python 2.6 and 2.7, unittest2 must be installed to run the tests.

Source installation

To install h5py from source, you need three things installed:

  • A supported Python version with development headers
  • HDF5 1.8.4 or newer with development headers
  • A C compiler

On Unix platforms, you also need pkg-config unless you explicitly specify a path for HDF5 as described in Custom installation.

OS-specific instructions for installing HDF5, Python and a C compiler are in the next few sections.

Additional Python-level requirements should be installed automatically (which will require an internet connection).

The actual installation of h5py should be done via:

$ pip install --no-binary=h5py h5py

or, from a tarball or git checkout

$ pip install -v .

or

$ python setup.py install

If you are working on a development version and the underlying cython files change it may be necessary to force a full rebuild. The easiest way to achieve this is

$ git clean -xfd

from the top of your clone and then rebuilding.

Source installation on OSX/MacOS

HDF5 and Python are most likely in your package manager (e.g. Homebrew, Macports, or Fink). Be sure to install the development headers, as sometimes they are not included in the main package.

XCode comes with a C compiler (clang), and your package manager will likely have other C compilers for you to install.

Source installation on Linux/Other Unix

HDF5 and Python are most likely in your package manager. A C compiler almost definitely is, usually there is some kind of metapackage to install the default build tools, e.g. build-essential, which should be sufficient for our needs. Make sure that that you have the development headers, as they are usually not installed by default. They can usually be found in python-dev or similar and libhdf5-dev or similar.

Source installation on Windows

Installing from source on Windows is a much more difficult prospect than installing from source on other OSs, as not only are you likely to need to compile HDF5 from source, everything must be built with the correct version of Visual Studio. Additional patches are also needed to HDF5 to get HDF5 and Python to work together.

We recommend examining the appveyor build scripts, and using those to build and install HDF5 and h5py.

Custom installation

Important

Remember that pip installs wheels by default. To perform a custom installation with pip, you should use:

$ pip install --no-binary=h5py h5py

or build from a git checkout or downloaded tarball to avoid getting a pre-built version of h5py.

You can specify build options for h5py with the configure option to setup.py. Options may be given together or separately:

$ python setup.py configure --hdf5=/path/to/hdf5
$ python setup.py configure --hdf5-version=X.Y.Z
$ python setup.py configure --mpi

Note the --hdf5-version option is generally not needed, as h5py auto-detects the installed version of HDF5 (even for custom locations).

Once set, build options apply to all future builds in the source directory. You can reset to the defaults with the --reset option:

$ python setup.py configure --reset

You can also configure h5py using environment variables. This is handy when installing via pip, as you don’t have direct access to setup.py:

$ HDF5_DIR=/path/to/hdf5 pip install --no-binary=h5py h5py
$ HDF5_VERSION=X.Y.Z pip install --no-binary=h5py h5py
$ CC="mpicc" HDF5_MPI="ON" HDF5_DIR=/path/to/parallel-hdf5 pip install --no-binary=h5py h5py

Here’s a list of all the configure options currently supported:

Option Via setup.py Via environment variable
Custom path to HDF5 --hdf5=/path/to/hdf5 HDF5_DIR=/path/to/hdf5
Force HDF5 version --hdf5-version=X.Y.Z HDF5_VERSION=X.Y.Z
Enable MPI mode --mpi HDF5_MPI=ON

Building against Parallel HDF5

If you just want to build with mpicc, and don’t care about using Parallel HDF5 features in h5py itself:

$ export CC=mpicc
$ pip install --no-binary=h5py h5py

If you want access to the full Parallel HDF5 feature set in h5py (Parallel HDF5), you will further have to build in MPI mode. This can either be done with command-line options from the h5py tarball or by:

$ export HDF5_MPI="ON"

You will need a shared-library build of Parallel HDF5 (i.e. built with ./configure –enable-shared –enable-parallel).

To build in MPI mode, use the --mpi option to setup.py configure or export HDF5_MPI="ON" beforehand:

$ export CC=mpicc
$ export HDF5_MPI="ON"
$ pip install --no-binary=h5py h5py

See also Parallel HDF5.

High-level API reference

File Objects

File objects serve as your entry point into the world of HDF5. In addition to the File-specific capabilities listed here, every File instance is also an HDF5 group representing the root group of the file.

Opening & creating files

HDF5 files work generally like standard Python file objects. They support standard modes like r/w/a, and should be closed when they are no longer in use. However, there is obviously no concept of “text” vs “binary” mode.

>>> f = h5py.File('myfile.hdf5','r')

The file name may be a byte string or unicode string. Valid modes are:

r Readonly, file must exist
r+ Read/write, file must exist
w Create file, truncate if exists
w- or x Create file, fail if exists
a Read/write if exists, create otherwise (default)

File drivers

HDF5 ships with a variety of different low-level drivers, which map the logical HDF5 address space to different storage mechanisms. You can specify which driver you want to use when the file is opened:

>>> f = h5py.File('myfile.hdf5', driver=<driver name>, <driver_kwds>)

For example, the HDF5 “core” driver can be used to create a purely in-memory HDF5 file, optionally written out to disk when it is closed. Here’s a list of supported drivers and their options:

None
Strongly recommended. Use the standard HDF5 driver appropriate for the current platform. On UNIX, this is the H5FD_SEC2 driver; on Windows, it is H5FD_WINDOWS.
‘sec2’
Unbuffered, optimized I/O using standard POSIX functions.
‘stdio’
Buffered I/O using functions from stdio.h.
‘core’

Store and manipulate the data in memory, and optionally write it back out when the file is closed. Using this with an existing file and a reading mode will read the entire file into memory. Keywords:

backing_store:
If True (default), save changes to the real file at the specified path on close() or flush(). If False, any changes are discarded when the file is closed.
block_size:
Increment (in bytes) by which memory is extended. Default is 64k.
‘family’

Store the file on disk as a series of fixed-length chunks. Useful if the file system doesn’t allow large files. Note: the filename you provide must contain a printf-style integer format code (e.g. %d”), which will be replaced by the file sequence number. Keywords:

memb_size: Maximum file size (default is 2**31-1).

‘fileobj’
Store the data in a Python file-like object; see below. This is the default if a file-like object is passed to File.

Python file-like objects

New in version 2.9.

The first argument to File may be a Python file-like object, such as an io.BytesIO or tempfile.TemporaryFile instance. This is a convenient way to create temporary HDF5 files, e.g. for testing or to send over the network.

The file-like object must be open for binary I/O, and must have these methods: read() (or readinto()), write(), seek(), tell(), truncate() and flush().

>>> tf = tempfile.TemporaryFile()
>>> f = h5py.File(tf)

Accessing the File instance after the underlying file object has been closed will result in undefined behaviour.

When using an in-memory object such as io.BytesIO, the data written will take up space in memory. If you want to write large amounts of data, a better option may be to store temporary data on disk using the functions in tempfile.

"""Create an HDF5 file in memory and retrieve the raw bytes

This could be used, for instance, in a server producing small HDF5
files on demand.
"""
import io
import h5py

bio = io.BytesIO()
with h5py.File(bio) as f:
    f['dataset'] = range(10)

data = bio.getvalue() # data is a regular Python bytes object.
print("Total size:", len(data))
print("First bytes:", data[:10])

Version bounding

HDF5 has been evolving for many years now. By default, the library will write objects in the most compatible fashion possible, so that older versions will still be able to read files generated by modern programs. However, there can be feature or performance advantages if you are willing to forgo a certain level of backwards compatibility. By using the “libver” option to File, you can specify the minimum and maximum sophistication of these structures:

>>> f = h5py.File('name.hdf5', libver='earliest') # most compatible
>>> f = h5py.File('name.hdf5', libver='latest')   # most modern

Here “latest” means that HDF5 will always use the newest version of these structures without particular concern for backwards compatibility. The “earliest” option means that HDF5 will make a best effort to be backwards compatible.

The default is “earliest”.

Specifying version bounds has changed from HDF5 version 1.10.2. There are two new compatibility levels: v108 (for HDF5 1.8) and v110 (for HDF5 1.10). This change enables, for example, something like this:

>>> f = h5py.File('name.hdf5', libver=('earliest', 'v108'))

which enforces full backward compatibility up to HDF5 1.8. Using any HDF5 feature that requires a newer format will raise an error.

latest is now an alias to another bound label that represents the latest version. Because of this, the File.libver property will not use latest in its output for HDF5 1.10.2 or later.

User block

HDF5 allows the user to insert arbitrary data at the beginning of the file, in a reserved space called the user block. The length of the user block must be specified when the file is created. It can be either zero (the default) or a power of two greater than or equal to 512. You can specify the size of the user block when creating a new file, via the userblock_size keyword to File; the userblock size of an open file can likewise be queried through the File.userblock_size property.

Modifying the user block on an open file is not supported; this is a limitation of the HDF5 library. However, once the file is closed you are free to read and write data at the start of the file, provided your modifications don’t leave the user block region.

Filenames on different systems

Different operating systems (and different file systems) store filenames with different encodings. Additionally, in Python there are at least two different representations of filenames, as encoded bytes (via str on Python 2, bytes on Python 3) or as a unicode string (via unicode on Python 2 and str on Python 3). The safest bet when creating a new file is to use unicode strings on all systems.

macOS (OSX)

macOS is the simplest system to deal with, it only accepts UTF-8, so using unicode paths will just work (and should be preferred).

Linux (and non-macOS Unix)

Unix-like systems use locale settings to determine the correct encoding to use. These are set via a number of different environment variables, of which LANG and LC_ALL are the ones of most interest. Of special interest is the C locale, which Python will interpret as only allowing ASCII, meaning unicode paths should be pre-encoded. This will likely change in Python 3.7 with https://www.python.org/dev/peps/pep-0538/, but this will likely be backported by distributions to earlier versions.

To summarise, use unicode strings where possible, but be aware that sometimes using encoded bytes may be necessary to read incorrectly encoded filenames.

Windows

Windows systems have two different APIs to perform file-related operations, a ANSI (char, legacy) interface and a unicode (wchar) interface. HDF5 currently only supports the ANSI interface, which is limited in what it can encode. This means that it may not be possible to open certain files, and because External links do not specify their encoding, it is possible that opening an external link may not work. There is work being done to fix this (see https://github.com/h5py/h5py/issues/839), but it is likely there will need to be breaking changes make to allow Windows to have the same level of support for unicode filenames as other operating systems.

The best suggestion is to use unicode strings, but to keep to ASCII for filenames to avoid possible breakage.

Chunk cache

Chunked storage allows datasets to be stored on disk in separate pieces. When a part of any one of these pieces is needed, the entire chunk is read into memory before the requested part is copied to the user’s buffer. To the extent possible those chunks are cached in memory, so that if the user requests a different part of a chunk that has already been read, the data can be copied directly from memory rather than reading the file again. The details of a given dataset’s chunks are controlled when creating the dataset, but it is possible to adjust the behavior of the chunk cache when opening the file.

The parameters controlling this behavior are prefixed by rdcc, for raw data chunk cache.

  • rdcc_nbytes sets the total size (measured in bytes) of the raw data chunk cache for each dataset. The default size is 1 MB. This should be set to the size of each chunk times the number of chunks that are likely to be needed in cache.
  • rdcc_w0 sets the policy for chunks to be removed from the cache when more space is needed. If the value is set to 0, then the library will always evict the least recently used chunk in cache. If the value is set to 1, the library will always evict the least recently used chunk which has been fully read or written, and if none have been fully read or written, it will evict the least recently used chunk. If the value is between 0 and 1, the behavior will be a blend of the two. Therefore, if the application will access the same data more than once, the value should be set closer to 0, and if the application does not, the value should be set closer to 1.
  • rdcc_nslots is the number of chunk slots in the cache for this entire file. In order to allow the chunks to be looked up quickly in cache, each chunk is assigned a unique hash value that is used to look up the chunk. The cache contains a simple array of pointers to chunks, which is called a hash table. A chunk’s hash value is simply the index into the hash table of the pointer to that chunk. While the pointer at this location might instead point to a different chunk or to nothing at all, no other locations in the hash table can contain a pointer to the chunk in question. Therefore, the library only has to check this one location in the hash table to tell if a chunk is in cache or not. This also means that if two or more chunks share the same hash value, then only one of those chunks can be in the cache at the same time. When a chunk is brought into cache and another chunk with the same hash value is already in cache, the second chunk must be evicted first. Therefore it is very important to make sure that the size of the hash table (which is determined by the rdcc_nslots parameter) is large enough to minimize the number of hash value collisions. Due to the hashing strategy, this value should ideally be a prime number. As a rule of thumb, this value should be at least 10 times the number of chunks that can fit in rdcc_nbytes bytes. For maximum performance, this value should be set approximately 100 times that number of chunks. The default value is 521.

Chunks and caching are described in greater detail in the HDF5 documentation.

Reference

Note

Unlike Python file objects, the attribute File.name gives the HDF5 name of the root group, “/”. To access the on-disk name, use File.filename.

class File(name, mode=None, driver=None, libver=None, userblock_size=None, swmr=False, rdcc_nslots=None, rdcc_nbytes=None, rdcc_w0=None, track_order=None, **kwds)

Open or create a new file.

Note that in addition to the File-specific methods and properties listed below, File objects inherit the full interface of Group.

Parameters:
  • name – Name of file (bytes or str), or an instance of h5f.FileID to bind to an existing file identifier, or a file-like object (see Python file-like objects).
  • mode – Mode in which to open file; one of (“w”, “r”, “r+”, “a”, “w-“). See Opening & creating files.
  • driver – File driver to use; see File drivers.
  • libver – Compatibility bounds; see Version bounding.
  • userblock_size – Size (in bytes) of the user block. If nonzero, must be a power of 2 and at least 512. See User block.
  • swmr – If True open the file in single-writer-multiple-reader mode. Only used when mode=”r”.
  • rdcc_nbytes – Total size of the raw data chunk cache in bytes. The default size is \(1024^2\) (1 MB) per dataset.
  • rdcc_w0 – Chunk preemption policy for all datasets. Default value is 0.75.
  • rdcc_nslots – Number of chunk slots in the raw data chunk cache for this file. Default value is 521.
  • kwds – Driver-specific keywords; see File drivers.
Track_order:

Track dataset/group/attribute creation order under root group if True. Default is h5.get_config().track_order.
close()

Close this file. All open objects will become invalid.

flush()

Request that the HDF5 library flush its buffers to disk.

id

Low-level identifier (an instance of FileID).

filename

Name of this file on disk. Generally a Unicode string; a byte string will be used if HDF5 returns a non-UTF-8 encoded string.

mode

String indicating if the file is open readonly (“r”) or read-write (“r+”). Will always be one of these two values, regardless of the mode used to open the file.

driver

String giving the driver used to open the file. Refer to File drivers for a list of drivers.

libver

2-tuple with library version settings. See Version bounding.

userblock_size

Size of user block (in bytes). Generally 0. See User block.

Groups

Groups are the container mechanism by which HDF5 files are organized. From a Python perspective, they operate somewhat like dictionaries. In this case the “keys” are the names of group members, and the “values” are the members themselves (Group and Dataset) objects.

Group objects also contain most of the machinery which makes HDF5 useful. The File object does double duty as the HDF5 root group, and serves as your entry point into the file:

>>> f = h5py.File('foo.hdf5','w')
>>> f.name
u'/'
>>> f.keys()
[]

Names of all objects in the file are all text strings (unicode on Py2, str on Py3). These will be encoded with the HDF5-approved UTF-8 encoding before being passed to the HDF5 C library. Objects may also be retrieved using byte strings, which will be passed on to HDF5 as-is.

Creating groups

New groups are easy to create:

>>> grp = f.create_group("bar")
>>> grp.name
'/bar'
>>> subgrp = grp.create_group("baz")
>>> subgrp.name
'/bar/baz'

Multiple intermediate groups can also be created implicitly:

>>> grp2 = f.create_group("/some/long/path")
>>> grp2.name
'/some/long/path'
>>> grp3 = f['/some/long']
>>> grp3.name
'/some/long'

Reference

class Group(identifier)

Generally Group objects are created by opening objects in the file, or by the method Group.create_group(). Call the constructor with a GroupID instance to create a new Group bound to an existing low-level identifier.

__iter__()

Iterate over the names of objects directly attached to the group. Use Group.visit() or Group.visititems() for recursive access to group members.

__contains__(name)

Dict-like membership testing. name may be a relative or absolute path.

__getitem__(name)

Retrieve an object. name may be a relative or absolute path, or an object or region reference. See Dict interface and links.

__setitem__(name, value)

Create a new link, or automatically create a dataset. See Dict interface and links.

keys()

Get the names of directly attached group members. On Py2, this is a list. On Py3, it’s a set-like object. Use Group.visit() or Group.visititems() for recursive access to group members.

values()

Get the objects contained in the group (Group and Dataset instances). Broken soft or external links show up as None. On Py2, this is a list. On Py3, it’s a collection or bag-like object.

items()

Get (name, value) pairs for object directly attached to this group. Values for broken soft or external links show up as None. On Py2, this is a list. On Py3, it’s a set-like object.

iterkeys()

(Py2 only) Get an iterator over key names. Exactly equivalent to iter(group). Use Group.visit() or Group.visititems() for recursive access to group members.

itervalues()

(Py2 only) Get an iterator over objects attached to the group. Broken soft and external links will show up as None.

iteritems()

(Py2 only) Get an iterator over (name, value) pairs for objects directly attached to the group. Broken soft and external link values show up as None.

get(name, default=None, getclass=False, getlink=False)

Retrieve an item, or information about an item. name and default work like the standard Python dict.get.

Parameters:
  • name – Name of the object to retrieve. May be a relative or absolute path.
  • default – If the object isn’t found, return this instead.
  • getclass – If True, return the class of object instead; Group or Dataset.
  • getlink – If true, return the type of link via a HardLink, SoftLink or ExternalLink instance. If getclass is also True, returns the corresponding Link class without instantiating it.
visit(callable)

Recursively visit all objects in this group and subgroups. You supply a callable with the signature:

callable(name) -> None or return value

name will be the name of the object relative to the current group. Return None to continue visiting until all objects are exhausted. Returning anything else will immediately stop visiting and return that value from visit:

>>> def find_foo(name):
...     """ Find first object with 'foo' anywhere in the name """
...     if 'foo' in name:
...         return name
>>> group.visit(find_foo)
u'some/subgroup/foo'
visititems(callable)

Recursively visit all objects in this group and subgroups. Like Group.visit(), except your callable should have the signature:

callable(name, object) -> None or return value

In this case object will be a Group or Dataset instance.

move(source, dest)

Move an object or link in the file. If source is a hard link, this effectively renames the object. If a soft or external link, the link itself is moved.

Parameters:
  • source (String) – Name of object or link to move.
  • dest (String) – New location for object or link.
copy(source, dest, name=None, shallow=False, expand_soft=False, expand_external=False, expand_refs=False, without_attrs=False)

Copy an object or group. The source and destination need not be in the same file. If the source is a Group object, by default all objects within that group will be copied recursively.

Parameters:
  • source – What to copy. May be a path in the file or a Group/Dataset object.
  • dest – Where to copy it. May be a path or Group object.
  • name – If the destination is a Group object, use this for the name of the copied object (default is basename).
  • shallow – Only copy immediate members of a group.
  • expand_soft – Expand soft links into new objects.
  • expand_external – Expand external links into new objects.
  • expand_refs – Copy objects which are pointed to by references.
  • without_attrs – Copy object(s) without copying HDF5 attributes.
create_group(name, track_order=None)

Create and return a new group in the file.

Parameters:name (String or None) – Name of group to create. May be an absolute or relative path. Provide None to create an anonymous group, to be linked into the file later.
Track_order:Track dataset/group/attribute creation order under this group if True. Default is h5.get_config().track_order.
Returns:The new Group object.
require_group(name)

Open a group in the file, creating it if it doesn’t exist. TypeError is raised if a conflicting object already exists. Parameters as in Group.create_group().

create_dataset(name, shape=None, dtype=None, data=None, **kwds)

Create a new dataset. Options are explained in Creating datasets.

Parameters:
  • name – Name of dataset to create. May be an absolute or relative path. Provide None to create an anonymous dataset, to be linked into the file later.
  • shape – Shape of new dataset (Tuple).
  • dtype – Data type for new dataset
  • data – Initialize dataset to this (NumPy array).
  • chunks – Chunk shape, or True to enable auto-chunking.
  • maxshape – Dataset will be resizable up to this shape (Tuple). Automatically enables chunking. Use None for the axes you want to be unlimited.
  • compression – Compression strategy. See Filter pipeline.
  • compression_opts – Parameters for compression filter.
  • scaleoffset – See Scale-Offset filter.
  • shuffle – Enable shuffle filter (T/F). See Shuffle filter.
  • fletcher32 – Enable Fletcher32 checksum (T/F). See Fletcher32 filter.
  • fillvalue – This value will be used when reading uninitialized parts of the dataset.
  • track_times – Enable dataset creation timestamps (T/F).
  • track_order – Track attribute creation order if True. Default is h5.get_config().track_order.
  • external – Store the dataset in one or more external, non-HDF5 files. This should be a list of tuples of (filename[, offset[, size]]), to store data from offset to offset + size in the specified file. The last file in the list may have size h5py.h5s.UNLIMITED to let it grow as needed.
require_dataset(name, shape=None, dtype=None, exact=None, **kwds)

Open a dataset, creating it if it doesn’t exist.

If keyword “exact” is False (default), an existing dataset must have the same shape and a conversion-compatible dtype to be returned. If True, the shape and dtype must match exactly.

Other dataset keywords (see create_dataset) may be provided, but are only used if a new dataset is to be created.

Raises TypeError if an incompatible object already exists, or if the shape or dtype don’t match according to the above rules.

Parameters:exact – Require shape and type to match exactly (T/F)
create_dataset_like(name, other, **kwds)

Create a dataset similar to other, much like numpy’s _like functions.

Parameters:
  • name – Name of the dataset (absolute or relative). Provide None to make an anonymous dataset.
  • other – The dataset whom the new dataset should mimic. All properties, such as shape, dtype, chunking, … will be taken from it, but no data or attributes are being copied.

Any dataset keywords (see create_dataset) may be provided, including shape and dtype, in which case the provided values take precedence over those from other.

create_virtual_dataset(name, layout, fillvalue=None)

Create a new virtual dataset in this group. See Virtual Datasets (VDS) for more details.

Parameters:
  • name (str) – Name of the dataset (absolute or relative).
  • layout (VirtualLayout) – Defines what source data fills which parts of the virtual dataset.
  • fillvalue – The value to use where there is no data.
attrs

Attributes for this group.

id

The groups’s low-level identifer; an instance of GroupID.

ref

An HDF5 object reference pointing to this group. See Using object references.

regionref

A proxy object allowing you to interrogate region references. See Using region references.

name

String giving the full path to this group.

file

File instance in which this group resides.

parent

Group instance containing this group.

Datasets

Datasets are very similar to NumPy arrays. They are homogeneous collections of data elements, with an immutable datatype and (hyper)rectangular shape. Unlike NumPy arrays, they support a variety of transparent storage features such as compression, error-detection, and chunked I/O.

They are represented in h5py by a thin proxy class which supports familiar NumPy operations like slicing, along with a variety of descriptive attributes:

  • shape attribute
  • size attribute
  • dtype attribute

h5py supports most NumPy dtypes, and uses the same character codes (e.g. 'f', 'i8') and dtype machinery as Numpy. See FAQ for the list of dtypes h5py supports.

Creating datasets

New datasets are created using either Group.create_dataset() or Group.require_dataset(). Existing datasets should be retrieved using the group indexing syntax (dset = group["name"]).

To make an empty dataset, all you have to do is specify a name, shape, and optionally the data type (defaults to 'f'):

>>> dset = f.create_dataset("default", (100,))
>>> dset = f.create_dataset("ints", (100,), dtype='i8')

You may initialize the dataset to an existing NumPy array:

>>> arr = np.arange(100)
>>> dset = f.create_dataset("init", data=arr)

Keywords shape and dtype may be specified along with data; if so, they will override data.shape and data.dtype. It’s required that (1) the total number of points in shape match the total number of points in data.shape, and that (2) it’s possible to cast data.dtype to the requested dtype.

Chunked storage

An HDF5 dataset created with the default settings will be contiguous; in other words, laid out on disk in traditional C order. Datasets may also be created using HDF5’s chunked storage layout. This means the dataset is divided up into regularly-sized pieces which are stored haphazardly on disk, and indexed using a B-tree.

Chunked storage makes it possible to resize datasets, and because the data is stored in fixed-size chunks, to use compression filters.

To enable chunked storage, set the keyword chunks to a tuple indicating the chunk shape:

>>> dset = f.create_dataset("chunked", (1000, 1000), chunks=(100, 100))

Data will be read and written in blocks with shape (100,100); for example, the data in dset[0:100,0:100] will be stored together in the file, as will the data points in range dset[400:500, 100:200].

Chunking has performance implications. It’s recommended to keep the total size of your chunks between 10 KiB and 1 MiB, larger for larger datasets. Also keep in mind that when any element in a chunk is accessed, the entire chunk is read from disk.

Since picking a chunk shape can be confusing, you can have h5py guess a chunk shape for you:

>>> dset = f.create_dataset("autochunk", (1000, 1000), chunks=True)

Auto-chunking is also enabled when using compression or maxshape, etc., if a chunk shape is not manually specified.

Resizable datasets

In HDF5, datasets can be resized once created up to a maximum size, by calling Dataset.resize(). You specify this maximum size when creating the dataset, via the keyword maxshape:

>>> dset = f.create_dataset("resizable", (10,10), maxshape=(500, 20))

Any (or all) axes may also be marked as “unlimited”, in which case they may be increased up to the HDF5 per-axis limit of 2**64 elements. Indicate these axes using None:

>>> dset = f.create_dataset("unlimited", (10, 10), maxshape=(None, 10))

Note

Resizing an array with existing data works differently than in NumPy; if any axis shrinks, the data in the missing region is discarded. Data does not “rearrange” itself as it does when resizing a NumPy array.

Filter pipeline

Chunked data may be transformed by the HDF5 filter pipeline. The most common use is applying transparent compression. Data is compressed on the way to disk, and automatically decompressed when read. Once the dataset is created with a particular compression filter applied, data may be read and written as normal with no special steps required.

Enable compression with the compression keyword to Group.create_dataset():

>>> dset = f.create_dataset("zipped", (100, 100), compression="gzip")

Options for each filter may be specified with compression_opts:

>>> dset = f.create_dataset("zipped_max", (100, 100), compression="gzip", compression_opts=9)
Lossless compression filters
GZIP filter ("gzip")
Available with every installation of HDF5, so it’s best where portability is required. Good compression, moderate speed. compression_opts sets the compression level and may be an integer from 0 to 9, default is 4.
LZF filter ("lzf")
Available with every installation of h5py (C source code also available). Low to moderate compression, very fast. No options.
SZIP filter ("szip")
Patent-encumbered filter used in the NASA community. Not available with all installations of HDF5 due to legal reasons. Consult the HDF5 docs for filter options.
Custom compression filters

In addition to the compression filters listed above, compression filters can be dynamically loaded by the underlying HDF5 library. This is done by passing a filter number to Group.create_dataset() as the compression parameter. The compression_opts parameter will then be passed to this filter.

Note

The underlying implementation of the compression filter will have the H5Z_FLAG_OPTIONAL flag set. This indicates that if the compression filter doesn’t compress a block while writing, no error will be thrown. The filter will then be skipped when subsequently reading the block.

Scale-Offset filter

Filters enabled with the compression keywords are lossless; what comes out of the dataset is exactly what you put in. HDF5 also includes a lossy filter which trades precision for storage space.

Works with integer and floating-point data only. Enable the scale-offset filter by setting Group.create_dataset() keyword scaleoffset to an integer.

For integer data, this specifies the number of bits to retain. Set to 0 to have HDF5 automatically compute the number of bits required for lossless compression of the chunk. For floating-point data, indicates the number of digits after the decimal point to retain.

Warning

Currently the scale-offset filter does not preserve special float values (i.e. NaN, inf), see https://lists.hdfgroup.org/pipermail/hdf-forum_lists.hdfgroup.org/2015-January/008296.html for more information and follow-up.

Shuffle filter

Block-oriented compressors like GZIP or LZF work better when presented with runs of similar values. Enabling the shuffle filter rearranges the bytes in the chunk and may improve compression ratio. No significant speed penalty, lossless.

Enable by setting Group.create_dataset() keyword shuffle to True.

Fletcher32 filter

Adds a checksum to each chunk to detect data corruption. Attempts to read corrupted chunks will fail with an error. No significant speed penalty. Obviously shouldn’t be used with lossy compression filters.

Enable by setting Group.create_dataset() keyword fletcher32 to True.

Reading & writing data

HDF5 datasets re-use the NumPy slicing syntax to read and write to the file. Slice specifications are translated directly to HDF5 “hyperslab” selections, and are a fast and efficient way to access data in the file. The following slicing arguments are recognized:

  • Indices: anything that can be converted to a Python long
  • Slices (i.e. [:] or [0:10])
  • Field names, in the case of compound data
  • At most one Ellipsis (...) object

Here are a few examples (output omitted)

>>> dset = f.create_dataset("MyDataset", (10,10,10), 'f')
>>> dset[0,0,0]
>>> dset[0,2:10,1:9:3]
>>> dset[:,::2,5]
>>> dset[0]
>>> dset[1,5]
>>> dset[0,...]
>>> dset[...,6]

For compound data, you can specify multiple field names alongside the numeric slices:

>>> dset["FieldA"]
>>> dset[0,:,4:5, "FieldA", "FieldB"]
>>> dset[0, ..., "FieldC"]

To retrieve the contents of a scalar dataset, you can use the same syntax as in NumPy: result = dset[()]. In other words, index into the dataset using an empty tuple.

For simple slicing, broadcasting is supported:

>>> dset[0,:,:] = np.arange(10)  # Broadcasts to (10,10)

Broadcasting is implemented using repeated hyperslab selections, and is safe to use with very large target selections. It is supported for the above “simple” (integer, slice and ellipsis) slicing only.

Warning

Currently h5py does not support nested compound types, see GH1197 for more information.

Fancy indexing

A subset of the NumPy fancy-indexing syntax is supported. Use this with caution, as the underlying HDF5 mechanisms may have different performance than you expect.

For any axis, you can provide an explicit list of points you want; for a dataset with shape (10, 10):

>>> dset.shape
(10, 10)
>>> result = dset[0, [1,3,8]]
>>> result.shape
(3,)
>>> result = dset[1:6, [5,8,9]]
>>> result.shape
(5, 3)

The following restrictions exist:

  • Selection coordinates must be given in increasing order
  • Duplicate selections are ignored
  • Very long lists (> 1000 elements) may produce poor performance

NumPy boolean “mask” arrays can also be used to specify a selection. The result of this operation is a 1-D array with elements arranged in the standard NumPy (C-style) order. Behind the scenes, this generates a laundry list of points to select, so be careful when using it with large masks:

>>> arr = numpy.arange(100).reshape((10,10))
>>> dset = f.create_dataset("MyDataset", data=arr)
>>> result = dset[arr > 50]
>>> result.shape
(49,)

Changed in version 2.10: Selecting using an empty list is now allowed. This returns an array with length 0 in the relevant dimension.

Length and iteration

As with NumPy arrays, the len() of a dataset is the length of the first axis, and iterating over a dataset iterates over the first axis. However, modifications to the yielded data are not recorded in the file. Resizing a dataset while iterating has undefined results.

On 32-bit platforms, len(dataset) will fail if the first axis is bigger than 2**32. It’s recommended to use Dataset.len() for large datasets.

Creating and Reading Empty (or Null) datasets and attributes

HDF5 has the concept of Empty or Null datasets and attributes. These are not the same as an array with a shape of (), or a scalar dataspace in HDF5 terms. Instead, it is a dataset with an associated type, no data, and no shape. In h5py, we represent this as either a dataset with shape None, or an instance of h5py.Empty. Empty datasets and attributes cannot be sliced.

To create an empty attribute, use h5py.Empty as per Attributes:

>>> obj.attrs["EmptyAttr"] = h5py.Empty("f")

Similarly, reading an empty attribute returns h5py.Empty:

>>> obj.attrs["EmptyAttr"]
h5py.Empty(dtype="f")

Empty datasets can be created either by defining a dtype but no shape in create_dataset:

>>> grp.create_dataset("EmptyDataset", dtype="f")

or by data to an instance of h5py.Empty:

>>> grp.create_dataset("EmptyDataset", data=h5py.Empty("f"))

An empty dataset has shape defined as None, which is the best way of determining whether a dataset is empty or not. An empty dataset can be “read” in a similar way to scalar datasets, i.e. if empty_dataset is an empty dataset,:

>>> empty_dataset[()]
h5py.Empty(dtype="f")

The dtype of the dataset can be accessed via <dset>.dtype as per normal. As empty datasets cannot be sliced, some methods of datasets such as read_direct will raise an exception if used on a empty dataset.

Reference

class Dataset(identifier)

Dataset objects are typically created via Group.create_dataset(), or by retrieving existing datasets from a file. Call this constructor to create a new Dataset bound to an existing DatasetID identifier.

__getitem__(args)

NumPy-style slicing to retrieve data. See Reading & writing data.

__setitem__(args)

NumPy-style slicing to write data. See Reading & writing data.

read_direct(array, source_sel=None, dest_sel=None)

Read from an HDF5 dataset directly into a NumPy array, which can avoid making an intermediate copy as happens with slicing. The destination array must be C-contiguous and writable, and must have a datatype to which the source data may be cast. Data type conversion will be carried out on the fly by HDF5.

source_sel and dest_sel indicate the range of points in the dataset and destination array respectively. Use the output of numpy.s_[args]:

>>> dset = f.create_dataset("dset", (100,), dtype='int64')
>>> arr = np.zeros((100,), dtype='int32')
>>> dset.read_direct(arr, np.s_[0:10], np.s_[50:60])
write_direct(source, source_sel=None, dest_sel=None)

Write data directly to HDF5 from a NumPy array. The source array must be C-contiguous. Selections must be the output of numpy.s_[<args>]. Broadcasting is supported for simple indexing.

astype(dtype)

Return a context manager allowing you to read data as a particular type. Conversion is handled by HDF5 directly, on the fly:

>>> dset = f.create_dataset("bigint", (1000,), dtype='int64')
>>> with dset.astype('int16'):
...     out = dset[:]
>>> out.dtype
dtype('int16')
resize(size, axis=None)

Change the shape of a dataset. size may be a tuple giving the new dataset shape, or an integer giving the new length of the specified axis.

Datasets may be resized only up to Dataset.maxshape.

len()

Return the size of the first axis.

make_scale(name='')

Make this dataset an HDF5 dimension scale.

You can then attach it to dimensions of other datasets like this:

other_ds.dims[0].attach_scale(ds)

You can optionally pass a name to associate with this scale.

virtual_sources()

If this dataset is a virtual dataset, return a list of named tuples: (vspace, file_name, dset_name, src_space), describing which parts of the dataset map to which source datasets. The two ‘space’ members are low-level SpaceID objects.

shape

NumPy-style shape tuple giving dataset dimensions.

dtype

NumPy dtype object giving the dataset’s type.

size

Integer giving the total number of elements in the dataset.

maxshape

NumPy-style shape tuple indicating the maxiumum dimensions up to which the dataset may be resized. Axes with None are unlimited.

chunks

Tuple giving the chunk shape, or None if chunked storage is not used. See Chunked storage.

compression

String with the currently applied compression filter, or None if compression is not enabled for this dataset. See Filter pipeline.

compression_opts

Options for the compression filter. See Filter pipeline.

scaleoffset

Setting for the HDF5 scale-offset filter (integer), or None if scale-offset compression is not used for this dataset. See Scale-Offset filter.

shuffle

Whether the shuffle filter is applied (T/F). See Shuffle filter.

fletcher32

Whether Fletcher32 checksumming is enabled (T/F). See Fletcher32 filter.

fillvalue

Value used when reading uninitialized portions of the dataset, or None if no fill value has been defined, in which case HDF5 will use a type-appropriate default value. Can’t be changed after the dataset is created.

external

If this dataset is stored in one or more external files, this is a list of 3-tuples, like the external= parameter to Group.create_dataset(). Otherwise, it is None.

is_virtual

True if this dataset is a virtual dataset, otherwise False.

dims

Access to Dimension Scales.

attrs

Attributes for this dataset.

id

The dataset’s low-level identifer; an instance of DatasetID.

ref

An HDF5 object reference pointing to this dataset. See Using object references.

regionref

Proxy object for creating HDF5 region references. See Using region references.

name

String giving the full path to this dataset.

file

File instance in which this dataset resides

parent

Group instance containing this dataset.

Attributes

Attributes are a critical part of what makes HDF5 a “self-describing” format. They are small named pieces of data attached directly to Group and Dataset objects. This is the official way to store metadata in HDF5.

Each Group or Dataset has a small proxy object attached to it, at <obj>.attrs. Attributes have the following properties:

  • They may be created from any scalar or NumPy array
  • Each attribute should be small (generally < 64k)
  • There is no partial I/O (i.e. slicing); the entire attribute must be read.

The .attrs proxy objects are of class AttributeManager, below. This class supports a dictionary-style interface.

By default, attributes are iterated in alphanumeric order. However, if group or dataset is created with track_order=True, the attribute insertion order is remembered (tracked) in HDF5 file, and iteration uses that order. The latter is consistent with Python 3.7+ dictionaries.

The default track_order for all new groups and datasets can be specified globally with h5.get_config().track_order.

Reference

class AttributeManager(parent)

AttributeManager objects are created directly by h5py. You should access instances by group.attrs or dataset.attrs, not by manually creating them.

__iter__()

Get an iterator over attribute names.

__contains__(name)

Determine if attribute name is attached to this object.

__getitem__(name)

Retrieve an attribute.

__setitem__(name, val)

Create an attribute, overwriting any existing attribute. The type and shape of the attribute are determined automatically by h5py.

__delitem__(name)

Delete an attribute. KeyError if it doesn’t exist.

keys()

Get the names of all attributes attached to this object. On Py2, this is a list. On Py3, it’s a set-like object.

values()

Get the values of all attributes attached to this object. On Py2, this is a list. On Py3, it’s a collection or bag-like object.

items()

Get (name, value) tuples for all attributes attached to this object. On Py2, this is a list of tuples. On Py3, it’s a collection or set-like object.

iterkeys()

(Py2 only) Get an iterator over attribute names.

itervalues()

(Py2 only) Get an iterator over attribute values.

iteritems()

(Py2 only) Get an iterator over (name, value) pairs.

get(name, default=None)

Retrieve name, or default if no such attribute exists.

get_id(name)

Get the low-level AttrID for the named attribute.

create(name, data, shape=None, dtype=None)

Create a new attribute, with control over the shape and type. Any existing attribute will be overwritten.

Parameters:
  • name (String) – Name of the new attribute
  • data – Value of the attribute; will be put through numpy.array(data).
  • shape (Tuple) – Shape of the attribute. Overrides data.shape if both are given, in which case the total number of points must be unchanged.
  • dtype (NumPy dtype) – Data type for the attribute. Overrides data.dtype if both are given.
modify(name, value)

Change the value of an attribute while preserving its type and shape. Unlike AttributeManager.__setitem__(), if the attribute already exists, only its value will be changed. This can be useful for interacting with externally generated files, where the type and shape must not be altered.

If the attribute doesn’t exist, it will be created with a default shape and type.

Parameters:
  • name (String) – Name of attribute to modify.
  • value – New value. Will be put through numpy.array(value).

Dimension Scales

Datasets are multidimensional arrays. HDF5 provides support for labeling the dimensions and associating one or more “dimension scales” with each dimension. A dimension scale is simply another HDF5 dataset. In principle, the length of the multidimensional array along the dimension of interest should be equal to the length of the dimension scale, but HDF5 does not enforce this property.

The HDF5 library provides the H5DS API for working with dimension scales. H5py provides low-level bindings to this API in h5py.h5ds. These low-level bindings are in turn used to provide a high-level interface through the Dataset.dims property. Suppose we have the following data file:

f = File('foo.h5', 'w')
f['data'] = np.ones((4, 3, 2), 'f')

HDF5 allows the dimensions of data to be labeled, for example:

f['data'].dims[0].label = 'z'
f['data'].dims[2].label = 'x'

Note that the first dimension, which has a length of 4, has been labeled “z”, the third dimension (in this case the fastest varying dimension), has been labeled “x”, and the second dimension was given no label at all.

We can also use HDF5 datasets as dimension scales. For example, if we have:

f['x1'] = [1, 2]
f['x2'] = [1, 1.1]
f['y1'] = [0, 1, 2]
f['z1'] = [0, 1, 4, 9]

We are going to treat the x1, x2, y1, and z1 datasets as dimension scales:

f['x1'].make_scale()
f['x2'].make_scale('x2 name')
f['y1'].make_scale('y1 name')
f['z1'].make_scale('z1 name')

When you create a dimension scale, you may provide a name for that scale. In this case, the x1 scale was not given a name, but the others were. Now we can associate these dimension scales with the primary dataset:

f['data'].dims[0].attach_scale(f['z1'])
f['data'].dims[1].attach_scale(f['y1'])
f['data'].dims[2].attach_scale(f['x1'])
f['data'].dims[2].attach_scale(f['x2'])

Note that two dimension scales were associated with the third dimension of data. You can also detach a dimension scale:

f['data'].dims[2].detach_scale(f['x2'])

but for now, lets assume that we have both x1 and x2 still associated with the third dimension of data. You can attach a dimension scale to any number of HDF5 datasets, you can even attach it to multiple dimensions of a single HDF5 dataset.

Now that the dimensions of data have been labeled, and the dimension scales for the various axes have been specified, we have provided much more context with which data can be interpreted. For example, if you want to know the labels for the various dimensions of data:

>>> [dim.label for dim in f['data'].dims]
['z', '', 'x']

If you want the names of the dimension scales associated with the “x” axis:

>>> f['data'].dims[2].keys()
['', 'x2 name']

items() and values() methods are also provided. The dimension scales themselves can also be accessed with:

f['data'].dims[2][1]

or:

f['data'].dims[2]['x2 name']

such that:

>>> f['data'].dims[2][1] == f['x2']
True

though, beware that if you attempt to index the dimension scales with a string, the first dimension scale whose name matches the string is the one that will be returned. There is no guarantee that the name of the dimension scale is unique.

Advanced topics

Configuring h5py

Library configuration

A few library options are available to change the behavior of the library. You can get a reference to the global library configuration object via the function h5py.get_config(). This object supports the following attributes:

complex_names
Set to a 2-tuple of strings (real, imag) to control how complex numbers are saved. The default is (‘r’,’i’).
bool_names
Booleans are saved as HDF5 enums. Set this to a 2-tuple of strings (false, true) to control the names used in the enum. The default is (“FALSE”, “TRUE”).
track_order
Whether to track dataset/group/attribute creation order. If container creation order is tracked, its links and attributes are iterated in ascending creation order (consistent with dict in Python 3.7+); otherwise in ascending alphanumeric order. Global configuration value can be overriden for particular container by specifying track_order argument to h5py.File, h5py.Group.create_group(), h5py.Group.create_dataset(). The default is False.

IPython

H5py ships with a custom ipython completer, which provides object introspection and tab completion for h5py objects in an ipython session. For example, if a file contains 3 groups, “foo”, “bar”, and “baz”:

In [4]: f['b<TAB>
bar   baz

In [4]: f['f<TAB>
# Completes to:
In [4]: f['foo'

In [4]: f['foo'].<TAB>
f['foo'].attrs            f['foo'].items            f['foo'].ref
f['foo'].copy             f['foo'].iteritems        f['foo'].require_dataset
f['foo'].create_dataset   f['foo'].iterkeys         f['foo'].require_group
f['foo'].create_group     f['foo'].itervalues       f['foo'].values
f['foo'].file             f['foo'].keys             f['foo'].visit
f['foo'].get              f['foo'].name             f['foo'].visititems
f['foo'].id               f['foo'].parent

The easiest way to enable the custom completer is to do the following in an IPython session:

In  [1]: import h5py

In [2]: h5py.enable_ipython_completer()

It is also possible to configure IPython to enable the completer every time you start a new session. For >=ipython-0.11, “h5py.ipy_completer” just needs to be added to the list of extensions in your ipython config file, for example ~/.config/ipython/profile_default/ipython_config.py (if this file does not exist, you can create it by invoking ipython profile create):

c = get_config()
c.InteractiveShellApp.extensions = ['h5py.ipy_completer']

For <ipython-0.11, the completer can be enabled by adding the following lines to the main() in .ipython/ipy_user_conf.py:

def main():
    ip.ex('from h5py import ipy_completer')
    ip.ex('ipy_completer.load_ipython_extension()')

Special types

HDF5 supports a few types which have no direct NumPy equivalent. Among the most useful and widely used are variable-length (VL) types, and enumerated types. As of version 2.3, h5py fully supports HDF5 enums and VL types.

How special types are represented

Since there is no direct NumPy dtype for variable-length strings, enums or references, h5py extends the dtype system slightly to let HDF5 know how to store these types. Each type is represented by a native NumPy dtype, with a small amount of metadata attached. NumPy routines ignore the metadata, but h5py can use it to determine how to store the data.

The metadata h5py attaches to dtypes is not part of the public API, so it may change between versions. Use the functions described below to create and check for these types.

Variable-length strings

See also

Strings in HDF5

In HDF5, data in VL format is stored as arbitrary-length vectors of a base type. In particular, strings are stored C-style in null-terminated buffers. NumPy has no native mechanism to support this. Unfortunately, this is the de facto standard for representing strings in the HDF5 C API, and in many HDF5 applications.

Thankfully, NumPy has a generic pointer type in the form of the “object” (“O”) dtype. In h5py, variable-length strings are mapped to object arrays. A small amount of metadata attached to an “O” dtype tells h5py that its contents should be converted to VL strings when stored in the file.

Existing VL strings can be read and written to with no additional effort; Python strings and fixed-length NumPy strings can be auto-converted to VL data and stored.

Here’s an example showing how to create a VL array of strings:

>>> f = h5py.File('foo.hdf5')
>>> dt = h5py.string_dtype(encoding='utf-8')
>>> ds = f.create_dataset('VLDS', (100,100), dtype=dt)
>>> ds.dtype.kind
'O'
>>> h5py.check_string_dtype(ds.dtype)
string_info(encoding='utf-8', length=None)
string_dtype(encoding='utf-8', length=None)

Make a numpy dtype for HDF5 strings

Parameters:
  • encoding – 'utf-8' or 'ascii'.
  • length – None for variable-length, or an integer for fixed-length string data, giving the length in bytes.

If encoding is 'utf-8', the variable length strings should be passed as Python str objects (unicode in Python 2). For 'ascii', they should be passed as bytes.

check_string_dtype(dt)

Check if dt is a string dtype. Returns a string_info object if it is, or None if not.

class string_info

A named tuple type holding string encoding and length.

encoding

The character encoding associated with the string dtype, which can be 'utf-8' or 'ascii'.

length

For fixed-length string dtypes, the length in bytes. None for variable-length strings.

Arbitrary vlen data

Starting with h5py 2.3, variable-length types are not restricted to strings. For example, you can create a “ragged” array of integers:

>>> dt = h5py.vlen_dtype(np.dtype('int32'))
>>> dset = f.create_dataset('vlen_int', (100,), dtype=dt)
>>> dset[0] = [1,2,3]
>>> dset[1] = [1,2,3,4,5]

Single elements are read as NumPy arrays:

>>> dset[0]
array([1, 2, 3], dtype=int32)

Multidimensional selections produce an object array whose members are integer arrays:

>>> dset[0:2]
array([array([1, 2, 3], dtype=int32), array([1, 2, 3, 4, 5], dtype=int32)], dtype=object)
vlen_dtype(basetype)

Make a numpy dtype for an HDF5 variable-length datatype.

Parameters:basetype – The dtype of each element in the array.
check_vlen_dtype(dt)

Check if dt is a variable-length dtype. Returns the base type if it is, or None if not.

Enumerated types

HDF5 has the concept of an enumerated type, which is an integer datatype with a restriction to certain named values. Since NumPy has no such datatype, HDF5 ENUM types are read and written as integers.

Here’s an example of creating an enumerated type:

>>> dt = h5py.enum_dtype({"RED": 0, "GREEN": 1, "BLUE": 42}, basetype='i')
>>> h5py.check_enum_dtype(dt)
{'BLUE': 42, 'GREEN': 1, 'RED': 0}
>>> f = h5py.File('foo.hdf5','w')
>>> ds = f.create_dataset("EnumDS", (100,100), dtype=dt)
>>> ds.dtype.kind
'i'
>>> ds[0,:] = 42
>>> ds[0,0]
42
>>> ds[1,0]
0
enum_dtype(values_dict, basetype=np.uint8)

Create a NumPy representation of an HDF5 enumerated type

Parameters:
  • values_dict – Mapping of string names to integer values.
  • basetype – An appropriate integer base dtype large enough to hold the possible options.
check_enum_dtype(dt)

Check if dt represents an enumerated type. Returns the values dict if it is, or None if not.

Object and region references

References have their own section.

Older API

Before h5py 2.9, a single pair of functions was used to create and check for all of these special dtypes. These are still available for backwards compatibility, but are deprecated in favour of the functions listed above.

special_dtype(**kwds)

Create a NumPy dtype object containing type hints. Only one keyword may be specified.

Parameters:
  • vlen – Base type for HDF5 variable-length datatype.
  • enum – 2-tuple (basetype, values_dict). basetype must be an integer dtype; values_dict is a dictionary mapping string names to integer values.
  • ref – Provide class h5py.Reference or h5py.RegionReference to create a type representing object or region references respectively.
check_dtype(**kwds)

Determine if the given dtype object is a special type. Example:

>>> out = h5py.check_dtype(vlen=mydtype)
>>> if out is not None:
...     print "Vlen of type %s" % out
str
Parameters:
  • vlen – Check for an HDF5 variable-length type; returns base class
  • enum – Check for an enumerated type; returns 2-tuple (basetype, values_dict).
  • ref – Check for an HDF5 object or region reference; returns either h5py.Reference or h5py.RegionReference.

Strings in HDF5

The Most Important Thing

If you remember nothing else, remember this:

All strings in HDF5 hold encoded text.

You can’t store arbitrary binary data in HDF5 strings. Not only will this break, it will break in odd, hard-to-discover ways that will leave you confused and cursing.

How to store raw binary data

If you have a non-text blob in a Python byte string (as opposed to ASCII or UTF-8 encoded text, which is fine), you should wrap it in a void type for storage. This will map to the HDF5 OPAQUE datatype, and will prevent your blob from getting mangled by the string machinery.

Here’s an example of how to store binary data in an attribute, and then recover it:

>>> binary_blob = b"Hello\x00Hello\x00"
>>> dset.attrs["attribute_name"] = np.void(binary_blob)
>>> out = dset.attrs["attribute_name"]
>>> binary_blob = out.tostring()

How to store text strings

At the high-level interface, h5py exposes three kinds of strings. Each maps to a specific type within Python (but see Compatibility below):

  • Fixed-length ASCII (NumPy S type)
  • Variable-length ASCII (Python 2 str, Python 3 bytes)
  • Variable-length UTF-8 (Python 2 unicode, Python 3 str)

Note that h5py currently lacks support for fixed-length UTF-8.

Compatibility

If you want to write maximally-compatible files and don’t want to read the whole chapter:

  • Use numpy.string_ for scalar attributes
  • Use the NumPy S dtype for datasets and array attributes
Fixed-length ASCII

These are created when you use numpy.string_:

>>> dset.attrs["name"] = numpy.string_("Hello")

or the S dtype:

>>> dset = f.create_dataset("string_ds", (100,), dtype="S10")

In the file, these map to fixed-width ASCII strings. One byte per character is used. The representation is “null-padded”, which is the internal representation used by NumPy (and the only one which round-trips through HDF5).

Technically, these strings are supposed to store only ASCII-encoded text, although in practice anything you can store in NumPy will round-trip. But for compatibility with other programs using HDF5 (IDL, MATLAB, etc.), you should use ASCII only.

Note

This is the most-compatible way to store a string. Everything else can read it.

Variable-length ASCII

These are created when you assign a byte string to an attribute:

>>> dset.attrs["attr"] = b"Hello"

or when you create a dataset with an explicit ascii string dtype:

>>> dt = h5py.string_dtype(encoding='ascii')
>>> dset = f.create_dataset("name", (100,), dtype=dt)

Note that they’re not fully identical to Python byte strings. You can only store ASCII-encoded text, without NULL bytes:

>>> dset.attrs["name"] = b"Hello\x00there"
ValueError: VLEN strings do not support embedded NULLs

In the file, these are created as variable-length strings with character set H5T_CSET_ASCII.

Variable-length UTF-8

These are created when you assign a unicode string to an attribute:

>>> dset.attrs["name"] = u"Hello"

or if you create a dataset with an explicit string dtype:

>>> dt = h5py.string_dtype()
>>> dset = f.create_dataset("name", (100,), dtype=dt)

They can store any character a Python unicode string can store, with the exception of NULLs. In the file these are created as variable-length strings with character set H5T_CSET_UTF8.

Exceptions for Python 3

Most strings in the HDF5 world are stored in ASCII, which means they map to byte strings. But in Python 3, there’s a strict separation between data and text, which intentionally makes it painful to handle encoded strings directly.

So, when reading or writing scalar string attributes, on Python 3 they will always be returned as type str, regardless of the underlying storage mechanism. The regular rules for writing apply; to get a fixed-width ASCII string, use numpy.string_, and to get a variable-length ASCII string, use bytes.

What about NumPy’s U type?

NumPy also has a Unicode type, a UTF-32 fixed-width format (4-byte characters). HDF5 has no support for wide characters. Rather than trying to hack around this and “pretend” to support it, h5py will raise an error when attempting to create datasets or attributes of this type.

Handling of lists/tuples of strings as attributes

If you set an attribute equal to a Python list or tuple of unicode strings, such as the following:

>>> f.attrs['x'] = (u'a', u'b')

h5py will save these as arrays of variable-length strings with character set H5T_CSET_UTF8. When read back, the results will be numpy arrays of dtype 'object', as if the original data were written as:

>>> f['x'] = np.array((u'a', u'b'), dtype=h5py.string_dtype(encoding='utf-8'))

Object names

Unicode strings are used exclusively for object names in the file:

>>> f.name
u'/'

You can supply either byte or unicode strings (on both Python 2 and Python 3) when creating or retrieving objects. If a byte string is supplied, it will be used as-is; Unicode strings will be encoded down to UTF-8.

In the file, h5py uses the most-compatible representation; H5T_CSET_ASCII for characters in the ASCII range; H5T_CSET_UTF8 otherwise.

>>> grp = f.create_dataset(b"name")
>>> grp2 = f.create_dataset(u"name2")

Object and Region References

In addition to soft and external links, HDF5 supplies one more mechanism to refer to objects and data in a file. HDF5 references are low-level pointers to other objects. The great advantage of references is that they can be stored and retrieved as data; you can create an attribute or an entire dataset of reference type.

References come in two flavors, object references and region references. As the name suggests, object references point to a particular object in a file, either a dataset, group or named datatype. Region references always point to a dataset, and additionally contain information about a certain selection (dataset region) on that dataset. For example, if you have a dataset representing an image, you could specify a region of interest, and store it as an attribute on the dataset.

Using object references

It’s trivial to create a new object reference; every high-level object in h5py has a read-only property “ref”, which when accessed returns a new object reference:

>>> myfile = h5py.File('myfile.hdf5')
>>> mygroup = myfile['/some/group']
>>> ref = mygroup.ref
>>> print ref
<HDF5 object reference>

“Dereferencing” these objects is straightforward; use the same syntax as when opening any other object:

>>> mygroup2 = myfile[ref]
>>> print mygroup2
<HDF5 group "/some/group" (0 members)>

Using region references

Region references always contain a selection. You create them using the dataset property “regionref” and standard NumPy slicing syntax:

>>> myds = myfile.create_dataset('dset', (200,200))
>>> regref = myds.regionref[0:10, 0:5]
>>> print regref
<HDF5 region reference>

The reference itself can now be used in place of slicing arguments to the dataset:

>>> subset = myds[regref]

For selections which don’t conform to a regular grid, h5py copies the behavior of NumPy’s fancy indexing, which returns a 1D array. Note that for h5py release before 2.2, h5py always returns a 1D array.

In addition to storing a selection, region references inherit from object references, and can be used anywhere an object reference is accepted. In this case the object they point to is the dataset used to create them.

Storing references in a dataset

HDF5 treats object and region references as data. Consequently, there is a special HDF5 type to represent them. However, NumPy has no equivalent type. Rather than implement a special “reference type” for NumPy, references are handled at the Python layer as plain, ordinary python objects. To NumPy they are represented with the “object” dtype (kind ‘O’). A small amount of metadata attached to the dtype tells h5py to interpret the data as containing reference objects.

These dtypes are available from h5py for references and region references:

>>> type(h5py.ref_dtype)
<type 'numpy.dtype'>
>>> ref_dtype.kind
'O'
>>> type(h5py.regionref_dtype)
<type 'numpy.dtype'>

The types accepted by this “ref=” keyword argument are h5py.Reference (for object references) and h5py.RegionReference (for region references).

To create an array of references, use this dtype as you normally would:

>>> ref_dataset = myfile.create_dataset("MyRefs", (100,), dtype=h5py.ref_dtype)

You can read from and write to the array as normal:

>>> ref_dataset[0] = myfile.ref
>>> print ref_dataset[0]
<HDF5 object reference>

Storing references in an attribute

Simply assign the reference to a name; h5py will figure it out and store it with the correct type:

>>> myref = myfile.ref
>>> myfile.attrs["Root group reference"] = myref

Null references

When you create a dataset of reference type, the uninitialized elements are “null” references. H5py uses the truth value of a reference object to indicate whether or not it is null:

>>> print bool(myfile.ref)
True
>>> nullref = ref_dataset[50]
>>> print bool(nullref)
False

Parallel HDF5

Starting with version 2.2.0, h5py includes support for Parallel HDF5. This is the “native” way to use HDF5 in a parallel computing environment.

How does Parallel HDF5 work?

Parallel HDF5 is a configuration of the HDF5 library which lets you share open files across multiple parallel processes. It uses the MPI (Message Passing Interface) standard for interprocess communication. Consequently, when using Parallel HDF5 from Python, your application will also have to use the MPI library.

This is accomplished through the mpi4py Python package, which provides excellent, complete Python bindings for MPI. Here’s an example “Hello World” using mpi4py:

>>> from mpi4py import MPI
>>> print "Hello World (from process %d)" % MPI.COMM_WORLD.Get_rank()

To run an MPI-based parallel program, use the mpiexec program to launch several parallel instances of Python:

$ mpiexec -n 4 python demo.py
Hello World (from process 1)
Hello World (from process 2)
Hello World (from process 3)
Hello World (from process 0)

The mpi4py package includes all kinds of mechanisms to share data between processes, synchronize, etc. It’s a different flavor of parallelism than, say, threads or multiprocessing, but easy to get used to.

Check out the mpi4py web site for more information and a great tutorial.

Building against Parallel HDF5

HDF5 must be built with at least the following options:

$./configure --enable-parallel --enable-shared

Note that --enable-shared is required.

Often, a “parallel” version of HDF5 will be available through your package manager. You can check to see what build options were used by using the program h5cc:

$ h5cc -showconfig

Once you’ve got a Parallel-enabled build of HDF5, h5py has to be compiled in “MPI mode”. This is simple; set your default compiler to the mpicc wrapper and build h5py with the --mpi option:

$ export CC=mpicc
$ python setup.py configure --mpi [--hdf5=/path/to/parallel/hdf5]
$ python setup.py build

Using Parallel HDF5 from h5py

The parallel features of HDF5 are mostly transparent. To open a file shared across multiple processes, use the mpio file driver. Here’s an example program which opens a file, creates a single dataset and fills it with the process ID:

from mpi4py import MPI
import h5py

rank = MPI.COMM_WORLD.rank  # The process ID (integer 0-3 for 4-process run)

f = h5py.File('parallel_test.hdf5', 'w', driver='mpio', comm=MPI.COMM_WORLD)

dset = f.create_dataset('test', (4,), dtype='i')
dset[rank] = rank

f.close()

Run the program:

$ mpiexec -n 4 python demo2.py

Looking at the file with h5dump:

$ h5dump parallel_test.hdf5
HDF5 "parallel_test.hdf5" {
GROUP "/" {
   DATASET "test" {
      DATATYPE  H5T_STD_I32LE
      DATASPACE  SIMPLE { ( 4 ) / ( 4 ) }
      DATA {
      (0): 0, 1, 2, 3
      }
   }
}
}

Collective versus independent operations

MPI-based programs work by launching many instances of the Python interpreter, each of which runs your script. There are certain requirements imposed on what each process can do. Certain operations in HDF5, for example, anything which modifies the file metadata, must be performed by all processes. Other operations, for example, writing data to a dataset, can be performed by some processes and not others.

These two classes are called collective and independent operations. Anything which modifies the structure or metadata of a file must be done collectively. For example, when creating a group, each process must participate:

>>> grp = f.create_group('x')  # right

>>> if rank == 1:
...     grp = f.create_group('x')   # wrong; all processes must do this

On the other hand, writing data to a dataset can be done independently:

>>> if rank > 2:
...     dset[rank] = 42   # this is fine

MPI atomic mode

HDF5 versions 1.8.9+ support the MPI “atomic” file access mode, which trades speed for more stringent consistency requirements. Once you’ve opened a file with the mpio driver, you can place it in atomic mode using the settable atomic property:

>>> f = h5py.File('parallel_test.hdf5', 'w', driver='mpio', comm=MPI.COMM_WORLD)
>>> f.atomic = True

More information

Parallel HDF5 is a new feature in h5py. If you have any questions, feel free to ask on the mailing list (h5py at google groups). We welcome bug reports, enhancements and general inquiries.

Single Writer Multiple Reader (SWMR)

Starting with version 2.5.0, h5py includes support for the HDF5 SWMR features.

What is SWMR?

The SWMR features allow simple concurrent reading of a HDF5 file while it is being written from another process. Prior to this feature addition it was not possible to do this as the file data and meta-data would not be synchronised and attempts to read a file which was open for writing would fail or result in garbage data.

A file which is being written to in SWMR mode is guaranteed to always be in a valid (non-corrupt) state for reading. This has the added benefit of leaving a file in a valid state even if the writing application crashes before closing the file properly.

This feature has been implemented to work with independent writer and reader processes. No synchronisation is required between processes and it is up to the user to implement either a file polling mechanism, inotify or any other IPC mechanism to notify when data has been written.

The SWMR functionality requires use of the latest HDF5 file format: v110. In practice this implies using at least HDF5 1.10 (this can be checked via h5py.version.info) and setting the libver bounding to “latest” when opening or creating the file.

Warning

New v110 format files are not compatible with v18 format. So, files written in SWMR mode with libver=’latest’ cannot be opened with older versions of the HDF5 library (basically any version older than the SWMR feature).

The HDF Group has documented the SWMR features in details on the website: Single-Writer/Multiple-Reader (SWMR) Documentation. This is highly recommended reading for anyone intending to use the SWMR feature even through h5py. For production systems in particular pay attention to the file system requirements regarding POSIX I/O semantics.

Using the SWMR feature from h5py

The following basic steps are typically required by writer and reader processes:

  • Writer process creates the target file and all groups, datasets and attributes.
  • Writer process switches file into SWMR mode.
  • Reader process can open the file with swmr=True.
  • Writer writes and/or appends data to existing datasets (new groups and datasets cannot be created when in SWMR mode).
  • Writer regularly flushes the target dataset to make it visible to reader processes.
  • Reader refreshes target dataset before reading new meta-data and/or main data.
  • Writer eventually completes and close the file as normal.
  • Reader can finish and close file as normal whenever it is convenient.

The following snippet demonstrate a SWMR writer appending to a single dataset:

f = h5py.File("swmr.h5", 'w', libver='latest')
arr = np.array([1,2,3,4])
dset = f.create_dataset("data", chunks=(2,), maxshape=(None,), data=arr)
f.swmr_mode = True
# Now it is safe for the reader to open the swmr.h5 file
for i in range(5):
    new_shape = ((i+1) * len(arr), )
    dset.resize( new_shape )
    dset[i*len(arr):] = arr
    dset.flush()
    # Notify the reader process that new data has been written

The following snippet demonstrate how to monitor a dataset as a SWMR reader:

f = h5py.File("swmr.h5", 'r', libver='latest', swmr=True)
dset = f["data"]
while True:
    dset.id.refresh()
    shape = dset.shape
    print( shape )

Examples

In addition to the above example snippets, a few more complete examples can be found in the examples folder. These examples are described in the following sections.

Dataset monitor with inotify

The inotify example demonstrates how to use SWMR in a reading application which monitors live progress as a dataset is being written by another process. This example uses the the linux inotify (pyinotify python bindings) to receive a signal each time the target file has been updated.


"""
    Demonstrate the use of h5py in SWMR mode to monitor the growth of a dataset
    on nofication of file modifications.

    This demo uses pyinotify as a wrapper of Linux inotify.
    https://pypi.python.org/pypi/pyinotify

    Usage:
            swmr_inotify_example.py [FILENAME [DATASETNAME]]

              FILENAME:    name of file to monitor. Default: swmr.h5
              DATASETNAME: name of dataset to monitor in DATAFILE. Default: data

    This script will open the file in SWMR mode and monitor the shape of the
    dataset on every write event (from inotify). If another application is
    concurrently writing data to the file, the writer must have have switched
    the file into SWMR mode before this script can open the file.
"""
import asyncore
import pyinotify
import sys
import h5py
import logging

#assert h5py.version.hdf5_version_tuple >= (1,9,178), "SWMR requires HDF5 version >= 1.9.178"

class EventHandler(pyinotify.ProcessEvent):

    def monitor_dataset(self, filename, datasetname):
        logging.info("Opening file %s", filename)
        self.f = h5py.File(filename, 'r', libver='latest', swmr=True)
        logging.debug("Looking up dataset %s"%datasetname)
        self.dset = self.f[datasetname]

        self.get_dset_shape()

    def get_dset_shape(self):
        logging.debug("Refreshing dataset")
        self.dset.refresh()

        logging.debug("Getting shape")
        shape = self.dset.shape
        logging.info("Read data shape: %s"%str(shape))
        return shape

    def read_dataset(self, latest):
        logging.info("Reading out dataset [%d]"%latest)
        self.dset[latest:]

    def process_IN_MODIFY(self, event):
        logging.debug("File modified!")
        shape = self.get_dset_shape()
        self.read_dataset(shape[0])

    def process_IN_CLOSE_WRITE(self, event):
        logging.info("File writer closed file")
        self.get_dset_shape()
        logging.debug("Good bye!")
        sys.exit(0)


if __name__ == "__main__":
    logging.basicConfig(format='%(asctime)s  %(levelname)s\t%(message)s',level=logging.INFO)

    file_name = "swmr.h5"
    if len(sys.argv) > 1:
        file_name = sys.argv[1]
    dataset_name = "data"
    if len(sys.argv) > 2:
        dataset_name = sys.argv[2]


    wm = pyinotify.WatchManager()  # Watch Manager
    mask = pyinotify.IN_MODIFY | pyinotify.IN_CLOSE_WRITE
    evh = EventHandler()
    evh.monitor_dataset( file_name, dataset_name )

    notifier = pyinotify.AsyncNotifier(wm, evh)
    wdd = wm.add_watch(file_name, mask, rec=False)

    # Sit in this loop() until the file writer closes the file
    # or the user hits ctrl-c
    asyncore.loop()
Multiprocess concurrent write and read

The SWMR multiprocess example starts two concurrent child processes: a writer and a reader. The writer process first creates the target file and dataset. Then it switches the file into SWMR mode and the reader process is notified (with a multiprocessing.Event) that it is safe to open the file for reading.

The writer process then continue to append chunks to the dataset. After each write it notifies the reader that new data has been written. Whether the new data is visible in the file at this point is subject to OS and file system latencies.

The reader first waits for the initial “SWMR mode” notification from the writer, upon which it goes into a loop where it waits for further notifications from the writer. The reader may drop some notifications, but for each one received it will refresh the dataset and read the dimensions. After a time-out it will drop out of the loop and exit.

"""
    Demonstrate the use of h5py in SWMR mode to write to a dataset (appending)
    from one process while monitoring the growing dataset from another process.

    Usage:
            swmr_multiprocess.py [FILENAME [DATASETNAME]]

              FILENAME:    name of file to monitor. Default: swmrmp.h5
              DATASETNAME: name of dataset to monitor in DATAFILE. Default: data

    This script will start up two processes: a writer and a reader. The writer
    will open/create the file (FILENAME) in SWMR mode, create a dataset and start
    appending data to it. After each append the dataset is flushed and an event
    sent to the reader process. Meanwhile the reader process will wait for events
    from the writer and when triggered it will refresh the dataset and read the
    current shape of it.
"""

import sys, time
import h5py
import numpy as np
import logging
from multiprocessing import Process, Event

class SwmrReader(Process):
    def __init__(self, event, fname, dsetname, timeout = 2.0):
        super(SwmrReader, self).__init__()
        self._event = event
        self._fname = fname
        self._dsetname = dsetname
        self._timeout = timeout

    def run(self):
        self.log = logging.getLogger('reader')
        self.log.info("Waiting for initial event")
        assert self._event.wait( self._timeout )
        self._event.clear()

        self.log.info("Opening file %s", self._fname)
        f = h5py.File(self._fname, 'r', libver='latest', swmr=True)
        assert f.swmr_mode
        dset = f[self._dsetname]
        try:
            # monitor and read loop
            while self._event.wait( self._timeout ):
                self._event.clear()
                self.log.debug("Refreshing dataset")
                dset.refresh()

                shape = dset.shape
                self.log.info("Read dset shape: %s"%str(shape))
        finally:
            f.close()

class SwmrWriter(Process):
    def __init__(self, event, fname, dsetname):
        super(SwmrWriter, self).__init__()
        self._event = event
        self._fname = fname
        self._dsetname = dsetname

    def run(self):
        self.log = logging.getLogger('writer')
        self.log.info("Creating file %s", self._fname)
        f = h5py.File(self._fname, 'w', libver='latest')
        try:
            arr = np.array([1,2,3,4])
            dset = f.create_dataset(self._dsetname, chunks=(2,), maxshape=(None,), data=arr)
            assert not f.swmr_mode

            self.log.info("SWMR mode")
            f.swmr_mode = True
            assert f.swmr_mode
            self.log.debug("Sending initial event")
            self._event.set()

            # Write loop
            for i in range(5):
                new_shape = ((i+1) * len(arr), )
                self.log.info("Resizing dset shape: %s"%str(new_shape))
                dset.resize( new_shape )
                self.log.debug("Writing data")
                dset[i*len(arr):] = arr
                #dset.write_direct( arr, np.s_[:], np.s_[i*len(arr):] )
                self.log.debug("Flushing data")
                dset.flush()
                self.log.info("Sending event")
                self._event.set()
        finally:
            f.close()


if __name__ == "__main__":
    logging.basicConfig(format='%(levelname)10s  %(asctime)s  %(name)10s  %(message)s',level=logging.INFO)
    fname = 'swmrmp.h5'
    dsetname = 'data'
    if len(sys.argv) > 1:
        fname = sys.argv[1]
    if len(sys.argv) > 2:
        dsetname = sys.argv[2]

    event = Event()
    reader = SwmrReader(event, fname, dsetname)
    writer = SwmrWriter(event, fname, dsetname)

    logging.info("Starting reader")
    reader.start()
    logging.info("Starting reader")
    writer.start()

    logging.info("Waiting for writer to finish")
    writer.join()
    logging.info("Waiting for reader to finish")
    reader.join()

The example output below (from a virtual Ubuntu machine) illustrate some latency between the writer and reader:

python examples/swmr_multiprocess.py
  INFO  2015-02-26 18:05:03,195        root  Starting reader
  INFO  2015-02-26 18:05:03,196        root  Starting reader
  INFO  2015-02-26 18:05:03,197      reader  Waiting for initial event
  INFO  2015-02-26 18:05:03,197        root  Waiting for writer to finish
  INFO  2015-02-26 18:05:03,198      writer  Creating file swmrmp.h5
  INFO  2015-02-26 18:05:03,203      writer  SWMR mode
  INFO  2015-02-26 18:05:03,205      reader  Opening file swmrmp.h5
  INFO  2015-02-26 18:05:03,210      writer  Resizing dset shape: (4,)
  INFO  2015-02-26 18:05:03,212      writer  Sending event
  INFO  2015-02-26 18:05:03,213      reader  Read dset shape: (4,)
  INFO  2015-02-26 18:05:03,214      writer  Resizing dset shape: (8,)
  INFO  2015-02-26 18:05:03,214      writer  Sending event
  INFO  2015-02-26 18:05:03,215      writer  Resizing dset shape: (12,)
  INFO  2015-02-26 18:05:03,215      writer  Sending event
  INFO  2015-02-26 18:05:03,215      writer  Resizing dset shape: (16,)
  INFO  2015-02-26 18:05:03,215      reader  Read dset shape: (12,)
  INFO  2015-02-26 18:05:03,216      writer  Sending event
  INFO  2015-02-26 18:05:03,216      writer  Resizing dset shape: (20,)
  INFO  2015-02-26 18:05:03,216      reader  Read dset shape: (16,)
  INFO  2015-02-26 18:05:03,217      writer  Sending event
  INFO  2015-02-26 18:05:03,217      reader  Read dset shape: (20,)
  INFO  2015-02-26 18:05:03,218      reader  Read dset shape: (20,)
  INFO  2015-02-26 18:05:03,219        root  Waiting for reader to finish

Virtual Datasets (VDS)

Starting with version 2.9, h5py includes high-level support for HDF5 ‘virtual datasets’. The VDS feature is available in version 1.10 of the HDF5 library; h5py must be built with a new enough version of HDF5 to create or read virtual datasets.

What are virtual datasets?

Virtual datasets allow a number of real datasets to be mapped together into a single, sliceable dataset via an interface layer. The mapping can be made ahead of time, before the parent files are written, and is transparent to the parent dataset characteristics (SWMR, chunking, compression etc…). The datasets can be meshed in arbitrary combinations, and even the data type converted.

Once a virtual dataset has been created, it can be read just like any other HDF5 dataset.

Warning

Virtual dataset files cannot be opened with versions of the hdf5 library older than 1.10.

The HDF Group has documented the VDS features in detail on the website: Virtual Datasets (VDS) Documentation.

Creating virtual datasets in h5py

To make a virtual dataset using h5py, you need to:

  1. Create a VirtualLayout object representing the dimensions and data type of the virtual dataset.
  2. Create a number of VirtualSource objects, representing the datasets the array will be built from. These objects can be created either from an h5py Dataset, or from a filename, dataset name and shape. This can be done even before the source file exists.
  3. Map slices from the sources into the layout.
  4. Convert the VirtualLayout object into a virtual dataset in an HDF5 file.

The following snippet creates a virtual dataset to stack together four 1D datasets from separate files into a 2D dataset:

layout = h5py.VirtualLayout(shape=(4, 100), dtype='i4')

for n in range(1, 5):
    filename = "{}.h5".format(n)
    vsource = h5py.VirtualSource(filename, 'data', shape=(100,))
    layout[n - 1] = vsource

# Add virtual dataset to output file
with h5py.File("VDS.h5", 'w', libver='latest') as f:
    f.create_virtual_dataset('data', layout, fillvalue=-5)

This is an extract from the vds_simple.py example in the examples folder.

Examples

In addition to the above example snippet, a few more complete examples can be found in the examples folder:

Reference

class VirtualLayout(shape, dtype=None, maxshape=None)

Object for building a virtual dataset.

Instantiate this class to define a virtual dataset, assign VirtualSource objects to slices of it, and then pass it to Group.create_virtual_dataset() to add the virtual dataset to a file.

This class does not allow access to the data; the virtual dataset must be created in a file before it can be used.

Parameters:
  • shape (tuple) – The full shape of the virtual dataset.
  • dtype – Numpy dtype or string.
  • maxshape (tuple) – The virtual dataset is resizable up to this shape. Use None for axes you want to be unlimited.
class VirtualSource(path_or_dataset, name=None, shape=None, dtype=None, maxshape=None)

Source definition for virtual data sets.

Instantiate this class to represent an entire source dataset, and then slice it to indicate which regions should be used in the virtual dataset.

Parameters:
  • path_or_dataset – The path to a file, or a Dataset object. If a dataset is given, no other parameters are allowed, as the relevant values are taken from the dataset instead.
  • name (str) – The name of the source dataset within the file.
  • shape (tuple) – The full shape of the source dataset.
  • dtype – Numpy dtype or string.
  • maxshape (tuple) – The source dataset is resizable up to this shape. Use None for axes you want to be unlimited.

Meta-info about the h5py project

“What’s new” documents

These document the changes between minor (or major) versions of h5py.

What’s new in h5py 2.9

New features
  • A convenient high-level API for creating virtual datasets, HDF5 objects which act as a view over one or more real datasets (GH1060, GH1126). See Virtual Datasets (VDS) for details.
  • File can now be constructed with a Python file-like object, making it easy to create an HDF5 file in memory using io.BytesIO (GH1061, GH1105, GH1116). See Python file-like objects for details.
  • File now accepts parameters to control the chunk cache (GH1008). See Chunk cache for details.
  • New options to record the order of insertion for attributes and group entries. Iterating over these collections now follows insertion order if it was recorded, or name order if not (GH1098).
  • A new method Group.create_dataset_like() to create a new dataset with similar properties to an existing one (GH1085).
  • Datasets can now be created with storage backed by external non-HDF5 files (GH1000).
  • Lists or tuples of unicode strings can now be stored as HDF5 attributes (GH1032).
  • Inspecting the view returned by .keys() now shows the key names, for convenient interactive use (GH1049).
Exposing HDF5 functions
  • H5LTopen_file_image as h5py.h5f.open_file_image() (GH1075).
  • External dataset storage functions H5Pset_external, H5Pget_external and H5Pget_external_count as methods on h5py.h5p.PropDCID (GH1000).
Bugfixes
  • Fix reading/writing of float128 data (GH1114).
  • Converting data to float16 when creating a dataset (GH1115).
Support for old Python

Support for Python 3.3 has been dropped.

Support for Python 2.6 has been dropped.

What’s new in h5py 2.8

This is the first release of the h5py 2.8 series. Note that the 2.8 series is the last series of h5py to support Python 2.6 and 3.3. Users should look to moving to Python 2.7 or (preferably) Python 3.4 or higher, as earlier releases are now outside of security support.

API changes
  • Deprecation of h5t.available_ftypes. This is no longer used internally and will be removed in the future. There is no replacement public API. See GH926 for how to add addition floating point types to h5py.
  • Do not sort fields in compound types (GH970 by James Tocknell). This is to account for changes in numpy 1.14.
  • Minimum required version of Cython is now 0.23.
Features
  • Allow registration of new file drivers (GH956 by Joe Jevnik).
  • Add option to track object creation order to Group.create_group (GH968 by Chen Yufei)
Bug fixes
  • Support slices with stop < start as empty slices (GH924 by Joe Jevnik)
  • Avoid crashing the IPython auto-completer when missing key (GH885, GH958 by James Tocknell). The auto-completer currently only works on older versions of IPython, see GH1022 for what’s needed to support newer versions of IPython/jupyter (PRs welcome!)
  • Set libver default to ‘earliest’ (a.k.a LIBVER_EARLIEST) as previously documented (GH933, GH936 by James Tocknell)
  • Fix conflict between fletcher32 and szip compression when using the float64 dtype (GH953, GH989, by Paul Müller).
  • Update floating point type handling as flagged by numpy deprecation warning (GH985, by Eric Larson)
  • Allow ExternalLinks to use non-ASCII hdf5 paths (GH333, GH952 by James Tocknell)
  • Prefer custom HDF5 over pkg-config/fallback paths when building/installing (GH946, GH947 by Lars Viklund)
  • Fix compatibility with Python 3 in document generation (GH921 by Ghislain Antony Vaillant)
  • Fix spelling and grammar in documentation (GH931 by Michael V. DePalatis, GH950 by Christian Sachs, GH1015 by Mikhail)
  • Add minor changes to documentation in order to improve clarity and warn about potential problems (GH528, GH783, GH829, GH849, GH911, GH959, by James Tocknell)
  • Add license field to setup.py metadata (GH999 by Nils Werner).
  • Use system encoding for errors, not utf-8 (GH1016, GH1025 by James Tocknell)
  • Add write_direct to the documentation (GH1028 by Sajid Ali and Thomas A Caswell)
Wheels HDF5 Version
  • Wheels uploaded to PyPI will now be built against the HDF5 1.10 series as opposed to the 1.8 series (h5py 2.8 is built against HDF5 1.10.2).
CI/Testing improvements and fixes

There were a number of improvements to testing and CI systems of h5py, including running the CI against multiple versions of HDF5, improving reliability and speed of the CIs, and simplifying the tox file. See GH857, GH894, GH922, GH954 and GH962 by Thomas A Caswell and James Tocknell for more details.

Other changes
  • Emphasise reading from HDF5 files rather than writing to files in Quickguide (GH609, GH610 by Yu Feng). Note these changes were in the 2.5 branch, but never got merged into master. The h5py 2.8 release now actually includes these changes.
  • Use lazy-loading of run_tests to avoid strong dependency on unittest2 (GH1013, GH1014 by Thomas VINCENT)
  • Correctly handle with multiple float types of the same size (GH926 by James Tocknell)
Acknowledgements and Thanks

The h5py developers thank Nathan Goldbaum, Matthew Brett, and Christoph Gohlke for building the wheels that appear on PyPI.

What’s new in h5py 2.7.1

2.7.1 is the first bug-fix release in the 2.7.x series.

Bug fixes
  • GH903 Fixed critical issue with cyclic gc which resulted in segfaults
  • GH904 Avoid unaligned access fixing h5py on sparc64
  • GH883 Fixed compilation issues for some library locations
  • GH868 Fix deadlock between phil and the import lock in py2
  • GH841 Improve windows handling if filenames
  • GH874 Allow close to be called on file multiple times
  • GH867, GH872 Warn on loaded vs complied hdf5 version issues
  • GH902 Fix overflow computing size of dataset on windows
  • GH912 Do not mangle capitalization of filenames in error messages
  • GH842 Fix longdouble on ppc64le
  • GH862, GH916 Fix compounds structs with variable-size members
Fix h5py segfaulting on some Python 3 versions

Through an intersection of Python Issue 30484 and GH888, it was possible for the Python Garbage Collector to activate when closing h5py objects, which due to how dictionaries were iterated over in Python could cause a segfault. GH903 fixes the Garbage Collector activating whilst closing, whilst Python Issue 30484 had been fixed upstream (and backported to Python 3.3 onwards).

Avoid unaligned memory access in conversion functions

Some architectures (e.g. SPARC64) do not allow unaligned memory access, which can come up when copying packed structs. GH904 (by James Clarke) uses memcpy to avoid said unaligned memory access.

What’s new in h5py 2.7

Python 3.2 is no longer supported

h5py 2.7 drops Python 3.2 support, and testing is not longer preformed on Python 3.2. The latest versions of pip, virtualenv, setuptools and numpy do not support Python 3.2, and dropping 3.2 allows both u and b prefixes to be used for strings. A clean up of some of the legacy code was done in #675 by Andrew Collette.

Additionally, support for Python 2.6 is soon to be dropped for pip (See https://github.com/pypa/pip/issues/3955) and setuptools (See https://github.com/pypa/setuptools/issues/878), and numpy has dropped Python 2.6 also in the latest release. While h5py has not dropped Python 2.6 this release, users are strongly encouraged to move to Python 2.7 where possible.

Improved testing support

There has been a major increase in the number of configurations h5py is automatically tested in, with Windows CI support added via Appveyor (#795, #798, #799 and #801 by James Tocknell) and testing of minimum requirements to ensure we still satisfy them (#703 by James Tocknell). Additionally, tox was used to ensure that we don’t run tests on Python versions which our dependencies have dropped or do not support (#662, #700 and #733). Thanks to to the Appveyor support, unicode tests were made more robust (#788, #800 and #804 by James Tocknell). Finally, other tests were improved or added where needed (#724 by Matthew Brett, #789, #794 and #802 by James Tocknell).

Improved python compatibility

The ipython/jupyter completion support now has Python 3 support (#715 by Joseph Kleinhenz). h5py now supports pathlib filenames (#716 by James Tocknell).

Documentation improvements

An update to the installation instructions and some whitespace cleanup was done in #808 by Thomas A Caswell, and mistake in the quickstart was fixed by Joydeep Bhattacharjee in #708.

setup.py improvements

Support for detecting the version of HDF5 via pkgconfig was added by Axel Huebl in #734, and support for specifying the path to MPI-supported HDF5 was added by Axel Huebl in #721. h5py's classifiers were updated to include supported python version and interpreters in #811 by James Tocknell.

Support for additional HDF5 features added

Low-level support for HDF5 Direct Chunk Write was added in #691 by Simon Gregor Ebner. Minimal support for HDF5 File Image Operations was added by Andrea Bedini in #680. Ideas and opinions for further support for both HDF5 Direct Chunk Write and HDF5 File Image Operations are welcome. High-level support for reading and writing null dataspaces was added in #664 by James Tocknell.

Improvements to type system

Reading and writing of compound datatypes has improved, with support for different orderings and alignments (#701 by Jonah Bernhard, #702 by Caleb Morse #738 by @smutch, #765 by Nathan Goldbaum and #793 by James Tocknell). Support for reading extended precision and non-standard floating point numbers has also been added (#749, #812 by Thomas A Caswell, #787 by James Tocknell and #781 by Martin Raspaud). Finally, compatibility improvements to Cython annotations of HDF5 types were added in #692 and #693 by Aleksandar Jelenak.

Other changes
  • Fix deprecation of - for numpy boolean arrays (#683 by James Tocknell)
  • Check for duplicates in fancy index validation (#739 by Sam Toyer)
  • Avoid potential race condition (#754 by James Tocknell)
  • Fix inconsistency when slicing with numpy.array of shape (1,) (#772 by Artsiom)
  • Use size_t to store Python object id (#773 by Christoph Gohlke)
  • Avoid errors when the Python GC runs during nonlocal_close() (#776 by Antoine Pitrou)
  • Move from six.PY3 to six.PY2 (#686 by James Tocknell)
Acknowledgements

What’s new in h5py 2.6

Support for HDF5 Virtual Dataset API

Initial support for the HDF5 Virtual Dataset API, which was introduced in HDF5 1.10, was added to the low-level API. Ideas and input for how this should work as part of the high-level interface are welcome.

This work was added in #663 by Aleksandar Jelenak.

Add MPI Collective I/O Support

Support for using MPI Collective I/O in both low-level and high-level code has been added. See the collective_io.py example for a simple demonstration of how to use MPI Collective I/O with the high level API.

This work was added in #648 by Jialin Liu.

Numerous build/testing/CI improvements

There were a number of improvements to the setup.py file, which should mean that pip install h5py should work in most places. Work was also done to clean up the current testing system, using tox is the recommended way of testing h5py across different Python versions. See #576 by Jakob Lombacher, #640 by Lawrence Mitchell, and #650, #651 and #658 by James Tocknell.

Cleanup of codebase based on pylint

There was a large cleanup of pylint-identified problems by Andrew Collette (#578, #579).

Fixes to low-level API

Fixes to the typing of functions were added in #597 by Ulrik Kofoed Pedersen, #589 by Peter Chang, and #625 by Spaghetti Sort. A fix for variable-length arrays was added in #621 by Sam Mason. Fixes to compound types were added in #639 by @nevion and #606 by Yu Feng. Finally, a fix to type conversion was added in #614 by Andrew Collette.

Documentation improvements
  • Updates to FAQ by Dan Guest (#608) and Peter Hill (#607).
  • Updates MPI-related documentation by Jens Timmerman (#604) and Matthias König (#572).
  • Fixes to documentation building by Ghislain Antony Vaillant (#562, #561).
  • Update PyTables link (#574 by Dominik Kriegner)
  • Add File opening modes to docstring (#563 by Antony Lee)
Other changes
  • Add Dataset.ndim (#649, #660 by @jakirkham, #661 by James Tocknell)
  • Fix import errors in IPython completer (#605 by Niru Maheswaranathan)
  • Turn off error printing in new threads (#583 by Andrew Collette)
  • Use item value in KeyError instead of error message (#642 by Matthias Geier)
Acknowledgements

What’s new in h5py 2.5

Experimental support for Single Writer Multiple Reader (SWMR)

This release introduces experimental support for the highly-anticipated “Single Writer Multiple Reader” (SWMR) feature in the upcoming HDF5 1.10 release. SWMR allows sharing of a single HDF5 file between multiple processes without the complexity of MPI or multiprocessing-based solutions.

This is an experimental feature that should NOT be used in production code. We are interested in getting feedback from the broader community with respect to performance and the API design.

For more details, check out the h5py user guide: http://docs.h5py.org/en/latest/swmr.html

SWMR support was contributed by Ulrik Pedersen (#551).

Other changes
  • Use system Cython as a fallback if cythonize() fails (#541 by Ulrik Pedersen).
  • Use pkg-config for building/linking against hdf5 (#505 by James Tocknell).
  • Disable building Cython on Travis (#513 by Andrew Collette).
  • Improvements to release tarball (#555, #560 by Ghislain Antony Vaillant).
  • h5py now has one codebase for both Python 2 and 3; 2to3 removed from setup.py (#508 by James Tocknell).
  • Add python 3.4 to tox (#507 by James Tocknell).
  • Warn when importing from inside install dir (#558 by Andrew Collette).
  • Tweak installation docs with reference to Anaconda and other Python package managers (#546 by Andrew Collette).
  • Fix incompatible function pointer types (#526, #524 by Peter H. Li).
  • Add explicit vlen is not None check to work around https://github.com/numpy/numpy/issues/2190 (#538 by Will Parkin).
  • Group and AttributeManager classes now inherit from the appropriate ABCs (#527 by James Tocknell).
  • Don’t strip metadata from special dtypes on read (#512 by Antony Lee).
  • Add ‘x’ mode as an alias for ‘w-‘ (#510 by Antony Lee).
  • Support dynamical loading of LZF filter plugin (#506 by Peter Colberg).
  • Fix accessing attributes with array type (#501 by Andrew Collette).
  • Don’t leak types in enum converter (#503 by Andrew Collette).
Acknowledgements

What’s new in h5py 2.4

Build system changes

The setup.py-based build system has been reworked to be more maintainable, and to fix certain long-standing bugs. As a consequence, the options to setup.py have changed; a new top-level “configure” command handles options like --hdf5=/path/to/hdf5 and --mpi. Setup.py now works correctly under Python 3 when these options are used.

Cython (0.17+) is now required when building from source on all platforms; the .c files are no longer shipped in the UNIX release. The minimum NumPy version is now 1.6.1.

Files will now auto-close

Files are now automatically closed when all objects within them are unreachable. Previously, if File.close() was not explicitly called, files would remain open and “leaks” were possible if the File object was lost.

Thread safety improvements

Access to all APIs, high- and low-level, are now protected by a global lock. The entire API is now believed to be thread-safe. Feedback and real-world testing is welcome.

Thanks to

Many people, but especially:

  • Matthieu Brucher
  • Laurence Hole
  • John Tyree
  • Pierre de Buyl
  • Matthew Brett

What’s new in h5py 2.3

Support for arbitrary vlen data

Variable-length data is no longer restricted to strings. You can use this feature to produce “ragged” arrays, whose members are 1D arrays of variable length.

The implementation of special types was changed to use the NumPy dtype “metadata” field. This change should be transparent, as access to special types is handled through h5py.special_dtype and h5py.check_dtype.

Improved exception messages

H5py has historically suffered from low-detail exception messages generated automatically by HDF5. While the exception types in 2.3 remain identical to those in 2.2, the messages have been substantially improved to provide more information as to the source of the error.

Examples:

ValueError: Unable to set extend dataset (Dimension cannot exceed the existing maximal size (new: 100 max: 1))

IOError: Unable to open file (Unable to open file: name = 'x3', errno = 2, error message = 'no such file or directory', flags = 0, o_flags = 0)

KeyError: "Unable to open object (Object 'foo' doesn't exist)"
Improved setuptools support

setup.py now uses setup_requires to make installation via pip friendlier.

Multiple low-level additions

Improved support for opening datasets via the low-level interface, by adding H5Dopen2 and many new property-list functions.

Improved support for MPI features

Added support for retrieving the MPI communicator and info objects from an open file. Added boilerplate code to allow compiling cleanly against newer versions of mpi4py.

Readonly files can now be opened in default mode

When opening a read-only file with no mode flags, now defaults to opening the file on RO mode rather than raising an exception.

Single-step build for HDF5 on Windows

Building h5py on windows has typically been hamstrung by the need to build a compatible version of HDF5 first. A new Paver-based system located in the “windows” distribution directory allows single-step compilation of HDF5 with settings that are known to work with h5py.

For more, see:

https://github.com/h5py/h5py/tree/master/windows

Thanks to
  • Martin Teichmann
  • Florian Rathgerber
  • Pierre de Buyl
  • Thomas Caswell
  • Andy Salnikov
  • Darren Dale
  • Robert David Grant
  • Toon Verstraelen
  • Many others who contributed bug reports

What’s new in h5py 2.2

Support for Parallel HDF5

On UNIX platforms, you can now take advantage of MPI and Parallel HDF5. Cython, mpi4py and an MPI-enabled build of HDF5 are required.. See Parallel HDF5 in the documentation for details.

Support for Python 3.3

Python 3.3 is now officially supported.

Mini float support (issue #141)

Two-byte floats (NumPy float16) are supported.

HDF5 scale/offset filter

The Scale/Offset filter added in HDF5 1.8 is now available.

Field indexing is now allowed when writing to a dataset (issue #42)

H5py has long supported reading only certain fields from a dataset:

>>> dset = f.create_dataset('x', (100,), dtype=np.dtype([('a', 'f'), ('b', 'i')]))
>>> out = dset['a', 0:100:10]
>>> out.dtype
dtype('float32')

Now, field names are also allowed when writing to a dataset:

>>> dset['a', 20:50] = 1.0
Region references preserve shape (issue #295)

Previously, region references always resulted in a 1D selection, even when 2D slicing was used:

>>> dset = f.create_dataset('x', (10, 10))
>>> ref = dset.regionref[0:5,0:5]
>>> out = dset[ref]
>>> out.shape
(25,)

Shape is now preserved:

>>> out = dset[ref]
>>> out.shape
(5, 5)

Additionally, the shape of both the target dataspace and the selection shape can be determined via new methods on the regionref proxy (now available on both datasets and groups):

>>> f.regionref.shape(ref)
(10, 10)
>>> f.regionref.selection(ref)
(5, 5)
Committed types can be linked to datasets and attributes

HDF5 supports “shared” named types stored in the file:

>>> f['name'] = np.dtype("int64")

You can now use these types when creating a new dataset or attribute, and HDF5 will “link” the dataset type to the named type:

>>> dset = f.create_dataset('int dataset', (10,), dtype=f['name'])
>>> f.attrs.create('int scalar attribute', shape=(), dtype=f['name'])
move method on Group objects

It’s no longer necessary to move objects in a file by manually re-linking them:

>>> f.create_group('a')
>>> f['b'] = f['a']
>>> del f['a']

The method Group.move allows this to be performed in one step:

>>> f.move('a', 'b')

Both the source and destination must be in the same file.

What’s new in h5py 2.1

Dimension scales

H5py now supports the Dimension Scales feature of HDF5! Thanks to Darren Dale for implementing this. You can find more information on using scales in the Dimension Scales section of the docs.

Unicode strings allowed in attributes

Group, dataset and attribute names in h5py 2.X can all be given as unicode. Now, you can also store (scalar) unicode data in attribute values as well:

>>> myfile.attrs['x'] = u"I'm a Unicode string!"

Storing Unicode strings in datasets or as members of compound types is not yet implemented.

Dataset size property

Dataset objects now expose a .size property which provides the total number of elements in the dataspace.

Dataset.value property is now deprecated.

The property Dataset.value, which dates back to h5py 1.0, is deprecated and will be removed in a later release. This property dumps the entire dataset into a NumPy array. Code using .value should be updated to use NumPy indexing, using mydataset[...] or mydataset[()] as appropriate.

Bug fixes
  • Object and region references were sometimes incorrectly wrapped wrapped in a numpy.object_ instance (issue 202)
  • H5py now ignores old versions of Cython (<0.13) when building (issue 221)
  • Link access property lists weren’t being properly tracked in the high level interface (issue 212)
  • Race condition fixed in identifier tracking which led to Python crashes (issue 151)
  • Highlevel objects will now complain if you try to bind them to the wrong HDF5 object types (issue 191)
  • Unit tests can now be run after installation (issue 201)

What’s new in h5py 2.0

HDF5 for Python (h5py) 2.0 represents the first major refactoring of the h5py codebase since the project’s launch in 2008. Many of the most important changes are behind the scenes, and include changes to the way h5py interacts with the HDF5 library and Python. These changes have substantially improved h5py’s stability, and make it possible to use more modern versions of HDF5 without compatibility concerns. It is now also possible to use h5py with Python 3.

Enhancements unlikely to affect compatibility
  • HDF5 1.8.3 through 1.8.7 now work correctly and are officially supported.
  • Python 3.2 is officially supported by h5py! Thanks especially to Darren Dale for getting this working.
  • Fill values can now be specified when creating a dataset. The fill time is H5D_FILL_TIME_IFSET for contiguous datasets, and H5D_FILL_TIME_ALLOC for chunked datasets.
  • On Python 3, dictionary-style methods like Group.keys() and Group.values() return view-like objects instead of lists.
  • Object and region references now work correctly in compound types.
  • Zero-length dimensions for extendable axes are now allowed.
  • H5py no longer attempts to auto-import ipython on startup.
  • File format bounds can now be given when opening a high-level File object (keyword “libver”).
Changes which may break existing code
Supported HDF5/Python versions
  • HDF5 1.6.X is no longer supported on any platform; following the release of 1.6.10 some time ago, this branch is no longer maintained by The HDF Group.
  • Python 2.6 or later is now required to run h5py. This is a consequence of the numerous changes made to h5py for Python 3 compatibility.
  • On Python 2.6, unittest2 is now required to run the test suite.
Group, Dataset and Datatype constructors have changed

In h5py 2.0, it is no longer possible to create new groups, datasets or named datatypes by passing names and settings to the constructors directly. Instead, you should use the standard Group methods create_group and create_dataset.

The File constructor remains unchanged and is still the correct mechanism for opening and creating files.

Code which manually creates Group, Dataset or Datatype objects will have to be modified to use create_group or create_dataset. File-resident datatypes can be created by assigning a NumPy dtype to a name (e.g. mygroup[“name”] = numpy.dtype(‘S10’)).

Unicode is now used for object names

Older versions of h5py used byte strings to represent names in the file. Starting with version 2.0, you may use either byte or unicode strings to create objects, but object names (obj.name, etc) will generally be returned as Unicode.

Code which may be affected:

  • Anything which uses “isinstance” or explicit type checks on names, expecting “str” objects. Such checks should be removed, or changed to compare to “basestring” instead.
  • In Python 2.X, other parts of your application may complain if they are handed Unicode data which can’t be encoded down to ascii. This is a general problem in Python 2.
File objects must be manually closed

With h5py 1.3, when File objects (or low-level FileID) objects went out of scope, the corresponding HDF5 file was closed. This led to surprising behavior, especially when files were opened with the H5F_CLOSE_STRONG flag; “losing” the original File object meant that all open groups and datasets suddenly became invalid.

Beginning with h5py 2.0, files must be manually closed, by calling the “close” method or by using the file object as a context manager. If you forget to close a file, the HDF5 library will try to close it for you when the application exits.

Please note that opening the same file multiple times (i.e. without closing it first) continues to result in undefined behavior.

Changes to scalar slicing code

When a scalar dataset was accessed with the syntax dataset[()], h5py incorrectly returned an ndarray. H5py now correctly returns an array scalar. Using dataset[...] on a scalar dataset still returns an ndarray.

Array scalars now always returned when indexing a dataset

When using datasets of compound type, retrieving a single element incorrectly returned a tuple of values, rather than an instance of numpy.void_ with the proper fields populated. Among other things, this meant you couldn’t do things like dataset[index][field]. H5py now always returns an array scalar, except in the case of object dtypes (references, vlen strings).

Reading object-like data strips special type information

In the past, reading multiple data points from dataset with vlen or reference type returned a Numpy array with a “special dtype” (such as those created by h5py.special_dtype()). In h5py 2.0, all such arrays now have a generic Numpy object dtype (numpy.dtype('O')). To get a copy of the dataset’s dtype, always use the dataset’s dtype property directly (mydataset.dtype).

The selections module has been removed

Only numpy-style slicing arguments remain supported in the high level interface. Existing code which uses the selections module should be refactored to use numpy slicing (and numpy.s_ as appropriate), or the standard C-style HDF5 dataspace machinery.

The H5Error exception class has been removed (along with h5py.h5e)

All h5py exceptions are now native Python exceptions, no longer inheriting from H5Error. RuntimeError is raised if h5py can’t figure out what exception is appropriate… every instance of this behavior is considered a bug. If you see h5py raising RuntimeError please report it so we can add the correct mapping!

The old errors module (h5py.h5e) has also been removed. There is no public error-management API.

File .mode property is now either ‘r’ or ‘r+

Files can be opened using the same mode arguments as before, but now the property File.mode will always return ‘r’ (read-only) or ‘r+’ (read-write).

Long-deprecated dict methods have been removed

Certain ancient aliases for Group/AttributeManager methods (e.g. listnames) have been removed. Please use the standard Python dict interface (Python 2 or Python 3 as appropriate) to interact with these objects.

Known issues
  • Thread support has been improved in h5py 2.0. However, we still recommend that for your own sanity you use locking to serialize access to files.
  • There are reports of crashes related to storing object and region references. If this happens to you, please post on the mailing list or contact the h5py author directly.

Bug Reports & Contributions

Contributions and bug reports are welcome from anyone! Some of the best features in h5py, including thread support, dimension scales, and the scale-offset filter, came from user code contributions.

Since we use GitHub, the workflow will be familiar to many people. If you have questions about the process or about the details of implementing your feature, always feel free to ask on the Google Groups list, either by emailing:

or via the web interface at:

Anyone can post to this list. Your first message will be approved by a moderator, so don’t worry if there’s a brief delay.

This guide is divided into three sections. The first describes how to file a bug report.

The second describes the mechanics of how to submit a contribution to the h5py project; for example, how to create a pull request, which branch to base your work on, etc. We assume you’re are familiar with Git, the version control system used by h5py. If not, here’s a great place to start.

Finally, we describe the various subsystems inside h5py, and give technical guidance as to how to implement your changes.

How to File a Bug Report

Bug reports are always welcome! The issue tracker is at:

If you’re unsure whether you’ve found a bug

Always feel free to ask on the mailing list (h5py at Google Groups). Discussions there are seen by lots of people and are archived by Google. Even if the issue you’re having turns out not to be a bug in the end, other people can benefit from a record of the conversation.

By the way, nobody will get mad if you file a bug and it turns out to be something else. That’s just how software development goes.

What to include

When filing a bug, there are two things you should include. The first is the output of h5py.version.info:

>>> import h5py
>>> print h5py.version.info

The second is a detailed explanation of what went wrong. Unless the bug is really trivial, include code if you can, either via GitHub’s inline markup:

```
    import h5py
    h5py.explode()    # Destroyed my computer!
```

or by uploading a code sample to Github Gist.

How to Get Your Code into h5py

This section describes how to contribute changes to the h5py code base. Before you start, be sure to read the h5py license and contributor agreement in “license.txt”. You can find this in the source distribution, or view it online at the main h5py repository at GitHub.

The basic workflow is to clone h5py with git, make your changes in a topic branch, and then create a pull request at GitHub asking to merge the changes into the main h5py project.

Here are some tips to getting your pull requests accepted:

  1. Let people know you’re working on something. This could mean posting a comment in an open issue, or sending an email to the mailing list. There’s nothing wrong with just opening a pull request, but it might save you time if you ask for advice first.
  2. Keep your changes focused. If you’re fixing multiple issues, file multiple pull requests. Try to keep the amount of reformatting clutter small so the maintainers can easily see what you’ve changed in a diff.
  3. Unit tests are mandatory for new features. This doesn’t mean hundreds (or even dozens) of tests! Just enough to make sure the feature works as advertised. The maintainers will let you know if more are needed.
Clone the h5py repository

The best way to do this is by signing in to GitHub and cloning the h5py project directly. You’ll end up with a new repository under your account; for example, if your username is yourname, the repository would be at http://github.com/yourname/h5py.

Then, clone your new copy of h5py to your local machine:

$ git clone http://github.com/yourname/h5py
Create a topic branch for your feature

Check out a new branch for the bugfix or feature you’re writing:

$ git checkout -b newfeature master

The exact name of the branch can be anything you want. For bug fixes, one approach is to put the issue number in the branch name.

We develop all changes against the master branch. If we’re making a bugfix release, a bot will backport merged pull requests.

Implement the feature!

You can implement the feature as a number of small changes, or as one big commit; there’s no project policy. Double-check to make sure you’ve included all your files; run git status and check the output.

Push your changes back and open a pull request

Push your topic branch back up to your GitHub clone:

$ git push origin newfeature

Then, create a pull request based on your topic branch.

Work with the maintainers

Your pull request might be accepted right away. More commonly, the maintainers will post comments asking you to fix minor things, like add a few tests, clean up the style to be PEP-8 compliant, etc.

The pull request page also shows the results of building and testing the modified code on Travis and Appveyor CI. Check back after about 30 minutes to see if the build succeeded, and if not, try to modify your changes to make it work.

When making changes after creating your pull request, just add commits to your topic branch and push them to your GitHub repository. Don’t try to rebase or open a new pull request! We don’t mind having a few extra commits in the history, and it’s helpful to keep all the history together in one place.

How to Modify h5py

This section is a little more involved, and provides tips on how to modify h5py. The h5py package is built in layers. Starting from the bottom, they are:

  1. The HDF5 C API (provided by libhdf5)
  2. Auto-generated Cython wrappers for the C API (api_gen.py)
  3. Low-level interface, written in Cython, using the wrappers from (2)
  4. High-level interface, written in Python, with things like h5py.File.
  5. Unit test code

Rather than talk about the layers in an abstract way, the parts below are guides to adding specific functionality to various parts of h5py. Most sections span at least two or three of these layers.

Adding a function from the HDF5 C API

This is one of the most common contributed changes. The example below shows how one would add the function H5Dget_storage_size, which determines the space on disk used by an HDF5 dataset. This function is already partially wrapped in h5py, so you can see how it works.

It’s recommended that you follow along, if not by actually adding the feature then by at least opening the various files as we work through the example.

First, get ahold of the function signature; the easiest place for this is at the online HDF5 Reference Manual. Then, add the function’s C signature to the file api_functions.txt:

hsize_t   H5Dget_storage_size(hid_t dset_id)

This particular signature uses types (hsize_t, hid_t) which are already defined elsewhere. But if the function you’re adding needs a struct or enum definition, you can add it using Cython code to the file api_types_hdf5.pxd.

The next step is to add a Cython function or method which calls the function you added. The h5py modules follow the naming convention of the C API; functions starting with H5D are wrapped in h5d.pyx.

Opening h5d.pyx, we notice that since this function takes a dataset identifier as the first argument, it belongs as a method on the DatasetID object. We write a wrapper method:

def get_storage_size(self):
    """ () => LONG storage_size

        Determine the amount of file space required for a dataset.  Note
        this only counts the space which has actually been allocated; it
        may even be zero.
    """
    return H5Dget_storage_size(self.id)

The first line of the docstring gives the method signature. This is necessary because Cython will use a “generic” signature like method(*args, **kwds) when the file is compiled. The h5py documentation system will extract the first line and use it as the signature.

Next, we decide whether we want to add access to this function to the high-level interface. That means users of the top-level h5py.Dataset object will be able to see how much space on disk their files use. The high-level interface is implemented in the subpackage h5py._hl, and the Dataset object is in module dataset.py. Opening it up, we add a property on the Dataset object:

@property
def storagesize(self):
    """ Size (in bytes) of this dataset on disk. """
    return self.id.get_storage_size()

You’ll see that the low-level DatasetID object is available on the high-level Dataset object as obj.id. This is true of all the high-level objects, like File and Group as well.

Finally (and don’t skip this step), we write unit tests for this feature. Since the feature is ultimately exposed at the high-level interface, it’s OK to write tests for the Dataset.storagesize property only. Unit tests for the high-level interface are located in the “tests” subfolder, right near dataset.py.

It looks like the right file is test_dataset.py. Unit tests are implemented as methods on custom unittest.UnitTest subclasses; each new feature should be tested by its own new class. In the test_dataset module, we see there’s already a subclass called BaseDataset, which implements some simple set-up and cleanup methods and provides a h5py.File object as obj.f. We’ll base our test class on that:

class TestStorageSize(BaseDataset):

    """
        Feature: Dataset.storagesize indicates how much space is used.
    """

    def test_empty(self):
        """ Empty datasets take no space on disk """
        dset = self.f.create_dataset("x", (100,100))
        self.assertEqual(dset.storagesize, 0)

    def test_data(self):
        """ Storage size is correct for non-empty datasets """
        dset = self.f.create_dataset("x", (100,), dtype='uint8')
        dset[...] = 42
        self.assertEqual(dset.storagesize, 100)

This set of tests would be adequate to get a pull request approved. We don’t test every combination under the sun (different ranks, datasets with more than 2**32 elements, datasets with the string “kumquat” in the name…), but the basic, commonly encountered set of conditions.

To build and test our changes, we have to do a few things. First of all, run the file api_gen.py to re-generate the Cython wrappers from api_functions.txt:

$ python api_gen.py

Then build the project, which recompiles h5d.pyx:

$ python setup.py build

Finally, run the test suite, which includes the two methods we just wrote:

$ python setup.py test

If the tests pass, the feature is ready for a pull request.

Adding a function only available in certain versions of HDF5

At the moment, h5py must be backwards-compatible all the way back to HDF5 1.8.4. Starting with h5py 2.2.0, it’s possible to conditionally include functions which only appear in newer versions of HDF5. It’s also possible to mark functions which require Parallel HDF5. For example, the function H5Fset_mpi_atomicity was introduced in HDF5 1.8.9 and requires Parallel HDF5. Specifiers before the signature in api_functions.txt communicate this:

MPI 1.8.9 herr_t H5Fset_mpi_atomicity(hid_t file_id, hbool_t flag)

You can specify either, both or none of “MPI” or a version number in “X.Y.Z” format.

In the Cython code, these show up as “preprocessor” defines MPI and HDF5_VERSION. So the low-level implementation (as a method on h5py.h5f.FileID) looks like this:

IF MPI and HDF5_VERSION >= (1, 8, 9):

    def set_mpi_atomicity(self, bint atomicity):
        """ (BOOL atomicity)

        For MPI-IO driver, set to atomic (True), which guarantees sequential
        I/O semantics, or non-atomic (False), which improves  performance.

        Default is False.

        Feature requires: 1.8.9 and Parallel HDF5
        """
        H5Fset_mpi_atomicity(self.id, <hbool_t>atomicity)

High-level code can check the version of the HDF5 library, or check to see if the method is present on FileID objects.

FAQ

What datatypes are supported?

Below is a complete list of types for which h5py supports reading, writing and creating datasets. Each type is mapped to a native NumPy type.

Fully supported types:

Type Precisions Notes
Integer 1, 2, 4 or 8 byte, BE/LE, signed/unsigned  
Float 2, 4, 8, 12, 16 byte, BE/LE  
Complex 8 or 16 byte, BE/LE Stored as HDF5 struct
Compound Arbitrary names and offsets  
Strings (fixed-length) Any length  
Strings (variable-length) Any length, ASCII or Unicode  
Opaque (kind ‘V’) Any length  
Boolean NumPy 1-byte bool Stored as HDF5 enum
Array Any supported type  
Enumeration Any NumPy integer type Read/write as integers
References Region and object  
Variable length array Any supported type See Special Types

Unsupported types:

Type Status
HDF5 “time” type  
NumPy “U” strings No HDF5 equivalent
NumPy generic “O” Not planned

What compression/processing filters are supported?

Filter Function Availability
DEFLATE/GZIP Standard HDF5 compression All platforms
SHUFFLE Increase compression ratio All platforms
FLETCHER32 Error detection All platforms
Scale-offset Integer/float scaling and truncation All platforms
SZIP Fast, patented compression for int/float
  • UNIX: if supplied with HDF5.
  • Windows: read-only
LZF Very fast compression, all types Ships with h5py, C source available

What file drivers are available?

A number of different HDF5 “drivers”, which provide different modes of access to the filesystem, are accessible in h5py via the high-level interface. The currently supported drivers are:

Driver Purpose Notes
sec2 Standard optimized driver Default on UNIX/Windows
stdio Buffered I/O using stdio.h  
core In-memory file (optionally backed to disk)  
family Multi-file driver  
mpio Parallel HDF5 file access  

What’s the difference between h5py and PyTables?

The two projects have different design goals. PyTables presents a database-like approach to data storage, providing features like indexing and fast “in-kernel” queries on dataset contents. It also has a custom system to represent data types.

In contrast, h5py is an attempt to map the HDF5 feature set to NumPy as closely as possible. For example, the high-level type system uses NumPy dtype objects exclusively, and method and attribute naming follows Python and NumPy conventions for dictionary and array access (i.e. “.dtype” and “.shape” attributes for datasets, group[name] indexing syntax for groups, etc).

Underneath the “high-level” interface to h5py (i.e. NumPy-array-like objects; what you’ll typically be using) is a large Cython layer which calls into C. This “low-level” interface provides access to nearly all of the HDF5 C API. This layer is object-oriented with respect to HDF5 identifiers, supports reference counting, automatic translation between NumPy and HDF5 type objects, translation between the HDF5 error stack and Python exceptions, and more.

This greatly simplifies the design of the complicated high-level interface, by relying on the “Pythonicity” of the C API wrapping.

There’s also a PyTables perspective on this question at the PyTables FAQ.

Does h5py support Parallel HDF5?

Starting with version 2.2, h5py supports Parallel HDF5 on UNIX platforms. mpi4py is required, as well as an MPIO-enabled build of HDF5. Check out Parallel HDF5 for details.

Variable-length (VLEN) data

Starting with version 2.3, all supported types can be stored in variable-length arrays (previously only variable-length byte and unicode strings were supported) See Special Types for use details. Please note that since strings in HDF5 are encoded as ASCII or UTF-8, NUL bytes are not allowed in strings.

Enumerated types

HDF5 enumerated types are supported. As NumPy has no native enum type, they are treated on the Python side as integers with a small amount of metadata attached to the dtype.

NumPy object types

Storage of generic objects (NumPy dtype “O”) is not implemented and not planned to be implemented, as the design goal for h5py is to expose the HDF5 feature set, not add to it. However, objects picked to the “plain-text” protocol (protocol 0) can be stored in HDF5 as strings.

Appending data to a dataset

The short response is that h5py is NumPy-like, not database-like. Unlike the HDF5 packet-table interface (and PyTables), there is no concept of appending rows. Rather, you can expand the shape of the dataset to fit your needs. For example, if I have a series of time traces 1024 points long, I can create an extendable dataset to store them:

>>> dset = myfile.create_dataset("MyDataset", (10, 1024), maxshape=(None, 1024))
>>> dset.shape
(10,1024)

The keyword argument “maxshape” tells HDF5 that the first dimension of the dataset can be expanded to any size, while the second dimension is limited to a maximum size of 1024. We create the dataset with room for an initial ensemble of 10 time traces. If we later want to store 10 more time traces, the dataset can be expanded along the first axis:

>>> dset.resize(20, axis=0)   # or dset.resize((20,1024))
>>> dset.shape
(20, 1024)

Each axis can be resized up to the maximum values in “maxshape”. Things to note:

  • Unlike NumPy arrays, when you resize a dataset the indices of existing data do not change; each axis grows or shrinks independently
  • The dataset rank (number of dimensions) is fixed when it is created

Unicode

As of h5py 2.0.0, Unicode is supported for file names as well as for objects in the file. When object names are read, they are returned as Unicode by default.

However, HDF5 has no predefined datatype to represent fixed-width UTF-16 or UTF-32 (NumPy format) strings. Therefore, the NumPy ‘U’ datatype is not supported.

Exceptions

h5py tries to map the error codes from hdf5 to the corresponding Exception class on the Python side. However the HDF5 group does not consider the error codes to be public API so we can not guarantee type stability of the exceptions raised.

Development

Building from Git

We moved to GitHub in December of 2012 (http://github.com/h5py/h5py).

We use the following conventions for branches and tags:

  • master: integration branch for the next minor (or major) version
  • 2.0, 2.1, 2.2, etc: bugfix branches for released versions
  • tags 2.0.0, 2.0.1, etc: Released bugfix versions
To build from a Git checkout:

Clone the project:

$ git clone https://github.com/h5py/h5py.git
$ cd h5py

(Optional) Choose which branch to build from (e.g. a stable branch):

$ git checkout 2.1

Build the project. If given, /path/to/hdf5 should point to a directory containing a compiled, shared-library build of HDF5 (containing things like “include” and “lib”):

$ python setup.py build [--hdf5=/path/to/hdf5]

(Optional) Run the unit tests:

$ python setup.py test

Report any failing tests to the mailing list (h5py at googlegroups), or by filing a bug report at GitHub.