H5PY Changelog

What's new in H5PY 2.2.1

Dec 10, 2013

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.
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)
Region references preserve shape (issue #295)
Committed types can be linked to datasets and attributes

New in H5PY 2.2.0 Beta 1 (Jul 18, 2013)

New in H5PY 2.1.3 (Apr 23, 2013)

New in H5PY 2.1.1 (Jan 7, 2013)

New in H5PY 2.1.0 (Oct 2, 2012)

New in H5PY 2.0.1 (Feb 11, 2012)

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 extendible 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HDF5 1.6.X is no longer supported on any platform; following the release of 1.6.10 some time ago, this branch is no longer maintained by The HDF Group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 a 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