Re: [h5md-user] Unit attribute versus non-dimensionless quantities

[Felix Höfling]

Am 01.08.2013, 09:56 Uhr, schrieb Felix Höfling <address@hidden>:

> Am 31.07.2013, 22:40 Uhr, schrieb Peter Colberg  
> <address@hidden>:
>
> > Hi all,
> >
> > I have discovered the perfect implementation of units in H5MD :-).
> >
> > Please marvel at the output of the attached example:
> >
> >   HDF5 "h5md_units.h5" {
> >   GROUP "/" {
> >      ATTRIBUTE "data" {
> >         DATATYPE  "/units/acceleration"
> >         DATASPACE  SCALAR
> >         DATA {
> >         (0): 9.81
> >         }
> >      }
> >      GROUP "units" {
> >         DATATYPE "acceleration" H5T_IEEE_F64LE;
> >            ATTRIBUTE "unit" {
> >               DATATYPE  H5T_STRING {
> >                  STRSIZE H5T_VARIABLE;
> >                  STRPAD H5T_STR_NULLTERM;
> >                  CSET H5T_CSET_UTF8;
> >                  CTYPE H5T_C_S1;
> >               }
> >               DATASPACE  SCALAR
> >               DATA {
> >               (0): "meter per square second"
> >               }
> >            }
> >      }
> >   }
> >   }
> >
> > I will prepare a complete H5MD units proposal for your consideration.
> >
> > Peter
>
> Hi Peter,
>
> Thanks for working out the example. It is an interesting solution, and I  
> consider it superior over the compound type way.
>
> I played a bit with h5py, which seems to ignore user-defined types  
> unfortunately. The type is immediately converted to a NumPy type, and it  
> is not possible (for me) to retrieve the actual HDF5 type of the  
> attribute (and the 'unit' attribute of this type):
>
> import h5py
> f = h5py.File('h5md_units.h5')
> t = f['units/acceleration']
> print f.attrs['data'].dtype
> f.create_dataset('dataset', dtype=t, data=(1,2))
> print type(f['dataset'])
> print f['dataset'].dtype
> f.close()
>
> The output is:
>
> float64
> <class 'h5py._hl.dataset.Dataset'>
> float64
>
> And the new dataset is shown by h5dump as:
>
>     DATASET "dataset" {
>        DATATYPE  H5T_IEEE_F64LE
>        DATASPACE  SIMPLE { ( 2 ) / ( 2 ) }
>        DATA {
>        (0): 1, 2
>        }
>     }
>
> Maybe there is still a trick with h5py. But a solution that works only  
> with low-level APIs would not be practical either :-(
>
> Cheers,
>
> Felix

h5py has also a low-level API, of course. But I think it is not  
interchangeable with the high-level one and one would have to stick to it  
everywhere. I was not even able to open a file with h5py.h5f.open() ...

My conclusion so far is that custom datatypes are something for  
experienced users, but are not suited to become part of the H5MD  
specification.

Felix