Organizing Data with References, Types, and Dimension Scales - Python and HDF5 - page 102

>>> f = h5py . File ( 'refs_demo.hdf5' , 'w' )
>>> grp1 = f . create_group ( 'group1' )
>>> grp2 = f . create_group ( 'group2' )
>>> dset = f . create_dataset ( 'mydata' , shape = ( 100 ,))
Looking at the group grp1 , we notice an interesting property called ref :
>>> grp1 . ref
<HDF5 object reference>
The object returned from accessing .ref is an HDF5 object reference . These are basically
pointers to objects in the file. You can “dereference” them by using the same syntax as
we used for string names:
>>> out = f [ grp1 . ref ]
>>> out == grp1
True
By the way, the Python type for these objects is available at h5py.Reference , in case you
want to use isinstance :
>>> isinstance ( grp1 . ref , h5py . Reference )
True
Since the reference is an “absolute” way of locating an object, you can use any group in
the file for dereferencing, not just the root group:
>>> out = grp2 [ grp1 . ref ]
>>> out == grp1
True
But keep in mind they're local to the file. Trying to dereference them in the context of
another file will fail:
>>> with h5py . File ( 'anotherfile.hdf5' , 'w' ) as f2 :
... out = f2 [ grp1 . ref ]
ValueError: unable dereference object
References as “Unbreakable” Links
So far there seems to be no improvement over using links. But there's an important
difference: you can store them as data , and they're independent of later renaming of the
objects involved.
Here's an example: suppose we wanted to add an attribute on one of our groups pointing
to the dataset mydata . We could simply record the name as an attribute:
>>> grp1 . attrs [ 'dataset' ] = dset . name
>>> grp1 . attrs [ 'dataset' ]
u'/mydata'
>>> out = f [ grp1 . attrs [ 'dataset' ]]
>>> out == dset
True