netcdf4-python: python/numpy interface to the netCDF C library
For details on the latest updates, see the Changelog.
10/31/2021: Version 1.5.8 released. Fix Enum bug, add binary wheels for aarch64 and python 3.10.
6/22/2021: Version 1.5.7 released. Fixed OverflowError on Windows when reading data with dimension sizes greater than 2**32-1. Masked arrays no longer returned for vlens.
2/15/2021: Version 1.5.6 released. Added
Dataset.tocdl, which require
ncgenutilities to be in
$PATH. Removed python 2.7 support.
12/20/2020: Version 22.214.171.124 released. Updated binary wheels for OSX and linux that link latest netcdf-c and hdf5 libs.
12/01/2020: Version 1.5.5 released. Update license wording to be consistent with MIT license.
07/23/2020: Version 1.5.4 released. Now requires numpy >= 1.9.
09/03/2019: Version 1.5.2 released. Bugfixes, no new features.
05/02/2019: Version 126.96.36.199 released. Fixes incorrect
__version__module variable in 1.5.1 release, plus a slicing bug (issue #919)).
04/30/2019: Version 1.5.1 released. Bugfixes, no new features.
04/02/2019: Version 188.8.131.52 released. Binary wheels for macos x and linux rebuilt with netcdf-c 4.6.3 (instead of 184.108.40.206). Added read-shared capability for faster reads of NETCDF3 files (mode='rs').
03/08/2019: Version 220.127.116.11 released. Include missing membuf.pyx file in source tarball. No need to update if you installed 18.104.22.168 from a binary wheel.
03/07/2019: Version 22.214.171.124 released. Fixes bug in implementation of NETCDF4_CLASSIC parallel IO support in 1.4.3.
03/05/2019: Version 1.4.3 released. Issues with netcdf-c 4.6.2 fixed (including broken parallel IO).
set_ncstring_attrs()method added, memoryview buffer now returned when an in-memory Dataset is closed.
10/26/2018: Version 1.4.2 released. Minor bugfixes, added
08/10/2018: Version 1.4.1 released. The old slicing behavior (numpy array returned unless missing values are present, otherwise masked array returned) is re-enabled via
05/11/2018: Version 1.4.0 released. The netcdftime package is no longer included, it is now a separate package dependency. In addition to several bug fixes, there are a few important changes to the default behaviour to note: * Slicing a netCDF variable will now always return masked array by default, even if there are no masked values. The result depended on the slice before, which was too surprising. If auto-masking is turned off (with
set_auto_mask(False)) a numpy array will always be returned. *
_FillValueis no longer treated as a validmin/validmax. This was too surprising, despite the fact the thet netcdf docs attribute best practices suggests that clients should to this if
valid_rangeare not set. * Changed behavior of string attributes so that
nc.stringatt = ['foo','bar']produces an vlen string array attribute in NETCDF4, instead of concatenating into a single string (
foobar). In NETCDF3/NETCDF4CLASSIC, an IOError is now raised, instead of writing
foobar. * Retrieved compound-type variable data now returned with character array elements converted to numpy strings (issue #773). Works for assignment also. Can be disabled using `setautochartostring(False)
. Numpy structured array dtypes with'SN'
string subtypes can now be used to define netcdf compound types increateCompoundType
(they get converted to('S1',N)
character array types automatically). *validmin
are now treated as unsigned integers ifUnsigned` variable attribute is set (to mimic behaviour of netcdf-java). Conversion to unsigned type now occurs before masking and scale/offset operation (issue #794)
11/01/2017: Version 1.3.1 released. Parallel IO support with MPI! Requires that netcdf-c and hdf5 be built with MPI support, and mpi4py. To open a file for parallel access in a program running in an MPI environment using mpi4py, just use
parallel=Truewhen creating the
examples/mpi_example.pyfor a demonstration. For more info, see the tutorial section.
9/25/2017: Version 1.3.0 released. Bug fixes for
netcdftimeand optimizations for reading strided slices.
encodingkwarg added to
Dataset.filepathto deal with oddball encodings in filename paths (
sys.getfilesystemencoding()is used by default to determine encoding). Make sure numpy datatypes used to define CompoundTypes have
isalignedstructflag set to avoid segfaults - which required bumping the minimum required numpy from 1.7.0 to 1.9.0. In cases where
missing_value/valid_min/valid_max/_FillValuecannot be safely cast to the variable's dtype, they are no longer be used to automatically mask the data and a warning message is issued.
6/10/2017: Version 1.2.9 released. Fixes for auto-scaling and masking when
valid_maxattributes present. setup.py updated so that
pip installworks if cython not installed. Now requires setuptools version 18.0 or greater.
6/1/2017: Version 1.2.8 released. From Changelog: * recognize
_Unsignedattribute used by netcdf-java to designate unsigned integer data stored with a signed integer type in netcdf-3 issue #656. * add Dataset init memory parameter to allow loading a file from memory pull request #652, issue #406 and issue #295. * fix for negative times in num2date issue #659. * fix for failing tests in numpy 1.13 due to changes in
numpy.maissue #662. * Checking for
NC_STRINGvariables, otherwise use 'utf-8'. 'utf-8' is used everywhere else, 'defaultencoding' global module variable is no longer used. getncattr method now takes optional kwarg 'encoding' (default 'utf-8') so encoding of attributes can be specified if desired. If `Encoding
is specified for anNCCHAR
) variable, the chartostring utility function is used to convert the array of characters to an array of strings with one less dimension (the last dimension is interpreted as the length of each string) when reading the data. When writing the data, stringtochar is used to convert a numpy array of fixed length strings to an array of characters with one more dimension. chartostring and stringtochar now also have an 'encoding' kwarg. Automatic conversion to/from character to string arrays can be turned off via a newsetautochartostring
Dataset and Variable method (default isTrue
). Addresses [issue #654](https://github.com/Unidata/netcdf4-python/issues/654) * [Cython](http://cython.org) >= 0.19 now required,netCDF4.c
and_netcdftime.c` removed from repository.
1/8/2017: Version 1.2.7 released. Python 3.6 compatibility, and fix for vector missing_values.
12/10/2016: Version 1.2.6 released. Bug fixes for Enum data type, and FillValue/missingvalue usage when data is stored in non-native endian format. Add getvariablesby_attributes to MFDataset. Support for python 2.6 removed.
4/15/2016: Version 1.2.4 released. Bugs in handling of variables with specified non-native "endian-ness" (byte-order) fixed (issue #554). Build instructions updated and warning issued to deal with potential backwards incompatibility introduced when using HDF5 1.10.x (see Unidata/netcdf-c/issue#250).
3/10/2016: Version 1.2.3 released. Various bug fixes. All text attributes in
NETCDF4formatted files are now written as type
NC_CHAR, unless they contain unicode characters that cannot be encoded in ascii, in which case they are written as
NC_STRING. Previously, all unicode strings were written as
NC_STRING. This change preserves compatibility with clients, like Matlab, that can't deal with
setncattr_stringmethod was added to force attributes to be written as
1/1/2016: Version 1.2.2 released. Mostly bugfixes, but with two new features.
support for the new
NETCDF3_64BIT_DATAformat introduced in netcdf-c 4.4.0. Similar to
NETCDF3_64BIT_OFFSET), but includes 64 bit dimension sizes (> 2 billion), plus unsigned and 64 bit integer data types. Uses the classic (netcdf-3) data model, and does not use HDF5 as the underlying storage format.
Dimension objects now have a
sizeattribute, which is the current length of the dimension (same as invoking
lenon the Dimension instance).
The minimum required python version has now been increased from 2.5 to 2.6.
10/15/2015: Version 1.2.1 released. Adds the ability to slice Variables with unsorted integer sequences, and integer sequences with duplicates.
9/23/2015: Version 1.2.0 released. New features:
Groupmethod for retrieving variables that have matching attributes.
Support for Enum data types.
7/28/2015: Version 1.1.9 bugfix release.
5/14/2015: Version 1.1.8 released. Unix-like paths can now be used in
python v = nc.createVariable('/path/to/var1', ('xdim', 'ydim'), float)will create a variable named 'var1', while also creating the groups 'path' and 'path/to' if they do not already exist.
python g = nc.createGroup('/path/to')now acts like
mkdir -pin unix, creating groups 'path' and '/path/to', if they don't already exist. Users who relied on
nc.createGroup(groupname)failing when the group already exists will have to modify their code, since
nc.createGroupwill now return the existing group instance.
Dataset.__getitem__was also added.
nc['/path/to']now returns a group instance, and
nc['/path/to/var1']now returns a variable instance.
3/19/2015: Version 1.1.7 released. Global Interpreter Lock (GIL) now released when extension module calls C library for read operations. This speeds up concurrent reads when using threads. Users who wish to use netcdf4-python inside threads should read http://www.hdfgroup.org/hdf5-quest.html#gconc regarding thread-safety in the HDF5 C library. Fixes to
setup.pynow ensure that
pip install netCDF4with
export USE_NCCONFIG=0will use environment variables to find paths to libraries and include files, instead of relying exclusively on the nc-config utility.
The easiest way to install is through pip:
pip install netCDF4
or, if you are a user of the Conda package manager,
conda install -c conda-forge netCDF4
Clone GitHub repository (
git clone https://github.com/Unidata/netcdf4-python.git)
Make sure HDF5 and netcdf-4 are installed, and the
nc-configutility is in your Unix PATH.
python setup.py build, then
pip install -e ..
To run all the tests, execute
cd test && python run_all.py.
See the online docs for more details.