A Python interface for the netCDF4 file-format that reads and writes HDF5 files API directly via h5py, without relying on the Unidata netCDF library.
- We've seen occasional reports of better performance with h5py than netCDF4-python, though in many cases performance is identical. For one workflow, h5netcdf was reported to be almost 4x faster than netCDF4-python.
- It has one less massive binary dependency (netCDF C). If you already have h5py installed, reading netCDF4 with h5netcdf may be much easier than installing netCDF4-Python.
- Anecdotally, HDF5 users seem to be unexcited about switching to netCDF -- hopefully this will convince them that the netCDF4 is actually quite sane!
- Finally, side-stepping the netCDF C library (and Cython bindings to it) gives us an easier way to identify the source of performance issues and bugs.
Ensure you have a recent version of h5py installed (I recommend using conda).
At least version 2.1 is required (for dimension scales); versions 2.3 and newer
have been verified to work, though some tests only pass on h5py 2.6. Then:
pip install h5netcdf
h5netcdf has two APIs, a new API and a legacy API. Both interfaces currently reproduce most of the features of the netCDF interface, with the noteable exceptions of:
- support for operations the rename or delete existing objects.
- suport for creating unlimited dimensions.
We simply haven't gotten around to implementing these features yet. Patches would be very welcome.
A new feature is the lazy opening of datasets. Unlike the netCDF interface where attributes for all variables and groups are read at the creation of the Dataset object, h5netcdf loads these attributes only when accessed.
The new API supports direct hierarchical access of variables and groups. Its design is an adaptation of h5py to the netCDF data model. For example:
import h5netcdf
import numpy as np
with h5netcdf.File('mydata.nc', 'w') as f:
# set dimensions with a dictionary
f.dimensions = {'x': 5}
# and update them with a dict-like interface
# f.dimensions['x'] = 5
# f.dimensions.update({'x': 5})
v = f.create_variable('hello', ('x',), float)
v[:] = np.ones(5)
# you don't need to create groups first
# you also don't need to create dimensions first if you supply data
# with the new variable
v = f.create_variable('/grouped/data', ('y',), data=np.arange(10))
# access and modify attributes with a dict-like interface
v.attrs['foo'] = 'bar'
# you can access variables and groups directly using a hierarchical
# keys like h5py
print(f['/grouped/data'])
The legacy API is designed for compatibility with netCDF4-python. To use it, import
h5netcdf.legacyapi
:
import h5netcdf.legacyapi as netCDF4
# everything here would also work with this instead:
# import netCDF4
import numpy as np
with netCDF4.Dataset('mydata.nc', 'w') as ds:
ds.createDimension('x', 5)
v = ds.createVariable('hello', float, ('x',))
v[:] = np.ones(5)
g = ds.createGroup('grouped')
g.createDimension('y', 10)
g.createVariable('data', 'i8', ('y',))
v = g['data']
v[:] = np.arange(10)
v.foo = 'bar'
print(ds.groups['grouped'].variables['data'])
The legacy API is designed to be easy to try-out for netCDF4-python users, but it is not an exact match. Here is an incomplete list of functionality we don't include:
- Utility functions
chartostring
,num2date
, etc., that are not directly necessary for writing netCDF files. - We don't support the
endian
argument tocreateVariable
. The h5py API does not appear to offer this feature. - h5netcdf variables do not support automatic masking or scaling (e.g., of values matching
the
_FillValue
attribute). We prefer to leave this functionality to client libraries (e.g., xarray), which can implement their exact desired scaling behavior.
Version 0.3.0:
- Datasets are now loaded lazily. This should increase performance when opening files with a large number of groups and/or variables.
- Support for writing arrays of variable length unicode strings with dtype=str via the legacy API.
- h5netcdf now writes the _NCProperties attribute for identifying netCDF4 files.