Python package to read and write a wide range of Python types to/from HDF5 formatted files. Can read/write data to the HDF5 based Matlab v7.3 MAT files.
Clone or download
frejanordsiek Fixed small bug in NumpyScalarArrayMarshallers where a list that is m…
…odified by inheriting classes was used to check if something was a handled numpy type.
Latest commit 5392751 Jul 22, 2018
Failed to load latest commit information.
doc Fixed bug where NumpyScalarArrayMarshaller could not identify dtype f… Jul 22, 2018
examples/example_hdf5storage_marshaller_plugin Fixed bugs in the previous commit. Jul 22, 2018
hdf5storage Fixed small bug in NumpyScalarArrayMarshallers where a list that is m… Jul 22, 2018
tests Changed the one test for nested 0 shaped dtypes to avoid a hard edge … Jul 22, 2018
.gitattributes Added setup.cfg to the manifest and .gitattributes Mar 28, 2016
.gitignore Added various autosaves to the gitignore list. Aug 3, 2017
.travis.yml Made travis testing load the example marshaller plugin package. Oct 15, 2017
COPYING.txt Updated copyright year. Feb 13, 2016 Added plugin marshaller support as well as the beginning of the Marsh… Oct 15, 2017
README.rst Fixed Issue #72, where python collections of ndarrays that have the s… Jul 10, 2018
THANKS.rst Added the contributors of two recent pull requests. Feb 20, 2016 Switched from distutils to setuptools for packaging so unit tests cou… Feb 2, 2014
requirements.txt Added plugin marshaller support as well as the beginning of the Marsh… Oct 15, 2017
requirements_doc.txt Updated documentation generation to newer sphinx. Jul 10, 2018
requirements_tests.txt Dropped Python 2.6 support due to it missing pkgutil.find_loader. Sep 14, 2016
setup.cfg Added default setup variables for building documentation using sphinx… Mar 28, 2016 Updated requirements list in to better elliminate versions o… Jul 10, 2018



This Python package provides high level utilities to read/write a variety of Python types to/from HDF5 (Heirarchal Data Format) formatted files. This package also provides support for MATLAB MAT v7.3 formatted files, which are just HDF5 files with a different extension and some extra meta-data.

All of this is done without pickling data. Pickling is bad for security because it allows arbitrary code to be executed in the interpreter. One wants to be able to read possibly HDF5 and MAT files from untrusted sources, so pickling is avoided in this package.

The package's documetation is found at

The package's source code is found at

The package is licensed under a 2-clause BSD license (



This package only supports Python >= 2.7. Python 2.6 support was dropped in version 0.2.

This package requires the numpy and h5py (>= 2.3) packages to run. Note that support for h5py 2.1.x and 2.2.x has been dropped in version 0.2. An optional dependency is the scipy package.

Installing by pip

This package is on PyPI. To install hdf5storage using pip, run the command:

pip install hdf5storage

Installing from Source

To install hdf5storage from source, download the package and then install the dependencies

pip install -r requirements.txt

Then to install the package, run the command with Python

python install

Running Tests

For testing, the package nose (>= 1.0) is additionally required. There are some tests that require Matlab and scipy to be installed and be in the executable path. In addition, there are some tests that require Julia with the MAT package. Not having them means that those tests cannot be run (they will be skipped) but all the other tests will run. To install all testing dependencies, other than scipy, Julia, Matlab run

pip install -r requirements_tests.txt.

To run the tests

python nosetests

Building Documentation

The documentation additionally requires sphinx (>= 1.7). The documentation dependencies can be installed by

pip install -r requirements_doc.txt

To build the documentation

python build_sphinx

Python 2

This package was designed and written for Python 3, and then backported to Python 2.x. This does mean that a few things are a little clunky in Python 2. For example, Python 2 int and long types are both mapped to the Python 3 int type. The storage format's metadata looks more familiar from a Python 3 standpoint as well.

The documentation is written in terms of Python 3 syntax and types primarily. Important Python 2 information beyond direct translations of syntax and types will be pointed out.

Hierarchal Data Format 5 (HDF5)

HDF5 files (see are a commonly used file format for exchange of numerical data. It has built in support for a large variety of number formats (un/signed integers, floating point numbers, strings, etc.) as scalars and arrays, enums and compound types. It also handles differences in data representation on different hardware platforms (endianness, different floating point formats, etc.). As can be imagined from the name, data is represented in an HDF5 file in a hierarchal form modelling a Unix filesystem (Datasets are equivalent to files, Groups are equivalent to directories, and links are supported).

This package interfaces HDF5 files using the h5py package ( as opposed to the PyTables package (

MATLAB MAT v7.3 file support

MATLAB ( MAT files version 7.3 and later are HDF5 files with a different file extension (.mat) and a very specific set of meta-data and storage conventions. This package provides read and write support for a limited set of Python and MATLAB types.

SciPy ( has functions to read and write the older MAT file formats. This package has functions modeled after the and functions, that have the same names and similar arguments. The dispatch to the SciPy versions if the MAT file format is not an HDF5 based one.

Supported Types

The supported Python and MATLAB types are given in the tables below. The tables assume that one has imported collections and numpy as:

import collections as cl
import numpy as np

The table gives which Python types can be read and written, the first version of this package to support it, the numpy type it gets converted to for storage (if type information is not written, that will be what it is read back as) the MATLAB class it becomes if targetting a MAT file, and the first version of this package to support writing it so MATlAB can read it.

Python MATLAB Notes
Type Version Converted to Class Version  
bool 0.1 np.bool_ or np.uint8 logical 0.1 [1]
None 0.1 np.float64([]) [] 0.1  
int 0.1 np.int64 or np.bytes_ int64 0.1 [2] [3]
long 0.1 np.int64 or np.bytes_ int64 0.1 [3] [4]
float 0.1 np.float64 double 0.1  
complex 0.1 np.complex128 double 0.1  
str 0.1 np.uint32/16 char 0.1 [5]
bytes 0.1 np.bytes_ or np.uint16 char 0.1 [6]
bytearray 0.1 np.bytes_ or np.uint16 char 0.1 [6]
list 0.1 np.object_ cell 0.1  
tuple 0.1 np.object_ cell 0.1  
set 0.1 np.object_ cell 0.1  
frozenset 0.1 np.object_ cell 0.1  
cl.deque 0.1 np.object_ cell 0.1  
dict 0.1   struct 0.1 [7]
cl.OrderedDict 0.2   struct 0.2 [7]
np.bool_ 0.1   logical 0.1  
np.void 0.1        
np.uint8 0.1   uint8 0.1  
np.uint16 0.1   uint16 0.1  
np.uint32 0.1   uint32 0.1  
np.uint64 0.1   uint64 0.1  
np.uint8 0.1   int8 0.1  
np.int16 0.1   int16 0.1  
np.int32 0.1   int32 0.1  
np.int64 0.1   int64 0.1  
np.float16 0.1       [8]
np.float32 0.1   single 0.1  
np.float64 0.1   double 0.1  
np.complex64 0.1   single 0.1  
np.complex128 0.1   double 0.1  
np.str_ 0.1 np.uint32/16 char/uint32 0.1 [5]
np.bytes_ 0.1 np.bytes_ or np.uint16 char 0.1 [6]
np.object_ 0.1   cell 0.1  
np.ndarray 0.1 see notes see notes 0.1 [9] [10] [11]
np.matrix 0.1 see notes see notes 0.1 [9]
np.chararray 0.1 see notes see notes 0.1 [9]
np.recarray 0.1 structured np.ndarray see notes 0.1 [9] [10]
[1]Depends on the selected options. Always np.uint8 when doing MATLAB compatiblity, or if the option is explicitly set.
[2]In Python 2.x, it may be read back as a long if it can't fit in the size of an int.
[3](1, 2) Stored as a np.int64 if it is small enough to fit. Otherwise its decimal string representation is stored as an np.bytes_ for hdf5storage >= 0.2 (error in earlier versions).
[4]Type found only in Python 2.x. Python 2.x's long and int are unified into a single int type in Python 3.x. Read as an int in Python 3.x.
[5](1, 2) Depends on the selected options and whether it can be converted to UTF-16 without using doublets. If the option is explicity set (or implicitly when doing MATLAB compatibility) and it can be converted to UTF-16 without losing any characters that can't be represented in UTF-16 or using UTF-16 doublets (MATLAB doesn't support them), then it is written as np.uint16 in UTF-16 encoding. Otherwise, it is stored at np.uint32 in UTF-32 encoding.
[6](1, 2, 3) Depends on the selected options. If the option is explicitly set (or implicitly when doing MATLAB compatibility), it will be stored as np.uint16 in UTF-16 encoding unless it has non-ASCII characters in which case a NotImplementedError is thrown). Otherwise, it is just written as np.bytes_.
[7](1, 2) Stored either as each key-value as their own Dataset or as two Datasets, one for keys and one for values. The former is used if all keys can be converted to str in Python 3 or unicode in Python 2 and they don't have null characters ('\x00') or forward slashes ('/') in them. Otherwise, the latter format is used.
[8]np.float16 are not supported for h5py versions before 2.2. Version 2.3 or higher is required for this package since version 0.2.
[9](1, 2, 3, 4) Container types are only supported if their underlying dtype is supported. Data conversions are done based on its dtype.
[10](1, 2) Structured np.ndarray s (have fields in their dtypes) can be written as an HDF5 COMPOUND type or as an HDF5 Group with Datasets holding its fields (either the values directly, or as an HDF5 Reference array to the values for the different elements of the data). Can only be written as an HDF5 COMPOUND type if none of its field are of dtype 'object'. Field names cannot have null characters ('\x00') and, when writing as an HDF5 GROUP, forward slashes ('/') in them.
[11]Structured np.ndarray s with no elements, when written like a structure, will not be read back with the right dtypes for their fields (will all become 'object').

This table gives the MATLAB classes that can be read from a MAT file, the first version of this package that can read them, and the Python type they are read as.

MATLAB Class Version Python Type
logical 0.1 np.bool_
single 0.1 np.float32 or np.complex64 [12]
double 0.1 np.float64 or np.complex128 [12]
uint8 0.1 np.uint8
uint16 0.1 np.uint16
uint32 0.1 np.uint32
uint64 0.1 np.uint64
int8 0.1 np.int8
int16 0.1 np.int16
int32 0.1 np.int32
int64 0.1 np.int64
char 0.1 np.str_
struct 0.1 structured np.ndarray
cell 0.1 np.object_
canonical empty 0.1 np.float64([])
[12](1, 2) Depends on whether there is a complex part or not.


0.2. Feature release adding/changing the following, including some API breaking changes.
  • Issue #50. Python 2.6 support dropped. The pkgutil.find_loader function is required, and it is not present in Python 2.6.
  • Issue #53. h5py 2.1.x and 2.2.x support dropped.
  • Issue #60. Platform label in the MAT file header changed to hdf5storage VERSION from CPython VERSION.
  • Issue #61. User provided marshallers must inherit from Marshallers.TypeMarshaller. Before, they just had to provide the same interface.
  • Issue #62. The priority ordering between builtin, plugin, and user provided marshallers can be selected. The default is now builtin, plugin, user; as opposed to user, builtin in the 0.1.x branch.
  • Issue #65. Added the ability to load marshallers from other python packages via plugin using the 'hdf5storage.marshallers.plugins' entry point in their files. Third party marshallers are not loaded into the default initial MarshallerCollection. Users who want to use them must call make_new_default_MarshallerCollection with the load_plugins option set to True.
  • Issue #66. A version Marshaller API has been added to make it easier for developers to write plugin marshallers without having to do extensive checking of the hdf5storage package version. The Marshaller API version will advance separately from the package version. The initial version is '1.0'.
  • Issue #27. Added of paths with null characters and slashes. It is used for the field names of structured numpy ndarrays as well as the keys of dict like objects when writing their values to individual Datasets.
  • Issue #49. Changed marshaller types and their handling code to support marshallers that handle types in modules that may not be available or should not be imported until needed. If the the required modules are not available, an approximate version of the data is read using the read_approximate method of the marshaller instead of the read method. The required modules, if available, can either be imported immediately upon the creation of the MarshallerCollection or they can be imported only when the marshaller is needed for actual use (lazy loading).
  • Issue #52. Added the usage of a default MarshallerCollection which is used whenever creating a new Options without a MarshallerCollection specified. The default can be obtained using get_default_MarshallerCollection and a new default can be generated using make_new_default_MarshallerCollection. This is useful if one wants to override the default lazy loading behavior.
  • Issue #42. read and write functions moved from the lowlevel and Marshallers modules to the utilities module and the lowlevel module renamed to exceptions since that is all that remains in it.
  • Ability to write Python 3.x int and Python 2.x long that are too large to fit into np.int64. Doing so no longer raises an exception.
  • Ability to write np.bytes_ with non-ASCII characters in them. Doing so no longer raises an exception.
  • Issue #24 and #25. Added support for writing dict like objects with keys that are not all str without null and '/' characters. Two new options, 'dict_like_keys_name' and 'dict_like_values_name' control how they are stored if the keys are not string like, can't be converted to Python 3.x str or Python 2.x unicode, or have null or '/' characters.
  • Issue #38. Added support for cl.OrderedDict. It was added to the Marshallers.PythonDictMarshaller.
  • Issue #40. Made it so that tests use tempfiles instead of using hardcoded filenames in the local directory.
  • Issue #41. Added tests using the Julia MAT package to check interop with Matlab v7.3 MAT files.
  • Issue #39. Documentation now uses the napoleon extension in Sphinx >= 1.3 as a replacement for numpydoc package.
  • Issue #55. Major performance increases by reducing the overhead involved with reading and writing each Dataset and Group.
0.1.15. Bugfix release that fixed the following bugs.
  • Issue #68. Fixed bug where str and numpy.unicode_ strings (but not ndarrays of them) were saved in uint32 format regardless of the value of Options.convert_numpy_bytes_to_utf16.
  • Issue #70. Updated and requirements.txt to specify the maximum versions of numpy and h5py that can be used for specific python versions (avoid version with dropped support).
  • Issue #71. Fixed bug where the 'python_fields' attribute wouldn't always be written when doing python metadata for data written in a struct-like fashion. The bug caused the field order to not be preserved when writing and reading.
  • Fixed an assertion in the tests to handle field re-ordering when no metadata is used for structured dtypes that only worked on older versions of numpy.
  • Issue #72. Fixed bug where python collections filled with ndarrays that all have the same shape were converted to multi-dimensional object ndarrays instead of a 1D object ndarray of the elements.
0.1.14. Bugfix release that also added a couple features.
  • Issue #45. Fixed syntax errors in unicode strings for Python 3.0 to 3.2.
  • Issues #44 and #47. Fixed bugs in testing of conversion and storage of string types.
  • Issue #46. Fixed raising of RuntimeWarnings in tests due to signalling NaNs.
  • Added requirements files for building documentation and running tests.
  • Made it so that Matlab compatability tests are skipped if Matlab is not found, instead of raising errors.
0.1.13. Bugfix release fixing the following bug.
  • Issue #36. Fixed bugs in writing int and long to HDF5 and their tests on 32 bit systems.
0.1.12. Bugfix release fixing the following bugs. In addition, copyright years were also updated and notices put in the Matlab files used for testing.
  • Issue #32. Fixed transposing before reshaping np.ndarray when reading from HDF5 files where python metadata was stored but not Matlab metadata.
  • Issue #33. Fixed the loss of the number of characters when reading empty numpy string arrays.
  • Issue #34. Fixed a conversion error when np.chararray are written with Matlab metadata.
0.1.11. Bugfix release fixing the following.
  • Issue #30. Fixed loadmat not opening files in read mode.
0.1.10. Minor feature/performance fix release doing the following.
  • Issue #29. Added writes and reads functions to write and read more than one piece of data at a time and made savemat and loadmat use them to increase performance. Previously, the HDF5 file was being opened and closed for each piece of data, which impacted performance, especially for large files.
0.1.9. Bugfix and minor feature release doing the following.
  • Issue #23. Fixed bug where a structured np.ndarray with a field name of 'O' could never be written as an HDF5 COMPOUND Dataset (falsely thought a field's dtype was object).
  • Issue #6. Added optional data compression and the storage of data checksums. Controlled by several new options.
0.1.8. Bugfix release fixing the following two bugs.
  • Issue #21. Fixed bug where the 'MATLAB_class' Attribute is not set when writing dict types when writing MATLAB metadata.
  • Issue #22. Fixed bug where null characters ('\x00') and forward slashes ('/') were allowed in dict keys and the field names of structured np.ndarray (except that forward slashes are allowed when the structured_numpy_ndarray_as_struct is not set as is the case when the matlab_compatible option is set). These cause problems for the h5py package and the HDF5 library. NotImplementedError is now thrown in these cases.
0.1.7. Bugfix release with an added compatibility option and some added test code. Did the following.
  • Fixed an issue reading variables larger than 2 GB in MATLAB MAT v7.3 files when no explicit variable names to read are given to hdf5storage.loadmat. Fix also reduces memory consumption and processing time a little bit by removing an unneeded memory copy.
  • Options now will accept any additional keyword arguments it doesn't support, ignoring them, to be API compatible with future package versions with added options.
  • Added tests for reading data that has been compressed or had other HDF5 filters applied.

0.1.6. Bugfix release fixing a bug with determining the maximum size of a Python 2.x int on a 32-bit system.

0.1.5. Bugfix release fixing the following bug.
  • Fixed bug where an int could be stored that is too big to fit into an int when read back in Python 2.x. When it is too big, it is converted to a long.
  • Fixed a bug where an int or long that is too big to big to fit into an np.int64 raised the wrong exception.
  • Fixed bug where fields names for structured np.ndarray with non-ASCII characters (assumed to be UTF-8 encoded in Python 2.x) can't be read or written properly.
  • Fixed bug where np.bytes_ with non-ASCII characters can were converted incorrectly to UTF-16 when that option is set (set implicitly when doing MATLAB compatibility). Now, it throws a NotImplementedError.
0.1.4. Bugfix release fixing the following bugs. Thanks goes to mrdomino for writing the bug fixes.
  • Fixed bug where dtype is used as a keyword parameter of np.ndarray.astype when it is a positional argument.
  • Fixed error caused by h5py.__version__ being absent on Ubuntu 12.04.
0.1.3. Bugfix release fixing the following bug.
  • Fixed broken ability to correctly read and write empty structured np.ndarray (has fields).
0.1.2. Bugfix release fixing the following bugs.
  • Removed mistaken support for np.float16 for h5py versions before 2.2 since that was when support for it was introduced.
  • Structured np.ndarray where one or more fields is of the 'object' dtype can now be written without an error when the structured_numpy_ndarray_as_struct option is not set. They are written as an HDF5 Group, as if the option was set.
  • Support for the 'MATLAB_fields' Attribute for data types that are structures in MATLAB has been added for when the version of the h5py package being used is 2.3 or greater. Support is still missing for earlier versions (this package requires a minimum version of 2.1).
  • The check for non-unicode string keys (str in Python 3 and unicode in Python 2) in the type dict is done right before any changes are made to the HDF5 file instead of in the middle so that no changes are applied if an invalid key is present.
  • HDF5 userblock set with the proper metadata for MATLAB support right at the beginning of when data is being written to an HDF5 file instead of at the end, meaning the writing can crash and the file will still be a valid MATLAB file.
0.1.1. Bugfix release fixing the following bugs.
  • str is now written like numpy.str_ instead of numpy.bytes_.
  • Complex numbers where the real or imaginary part are nan but the other part are not are now read correctly as opposed to setting both parts to nan.
  • Fixed bugs in string conversions on Python 2 resulting from str.decode() and unicode.encode() not taking the same keyword arguments as in Python 3.
  • MATLAB structure arrays can now be read without producing an error on Python 2.
  • numpy.str_ now written as numpy.uint16 on Python 2 if the convert_numpy_str_to_utf16 option is set and the conversion can be done without using UTF-16 doublets, instead of always writing them as numpy.uint32.

0.1. Initial version.