A Python library to read from and write to FITS files.
This is a Python extension written in C and Python. Data are read into numerical Python arrays.
A version of cfitsio is bundled with this package, there is no need to install
your own, nor will this conflict with a version you have installed.
- Read from and write to image, binary, and ASCII table extensions.
- Read arbitrary subsets of table columns and rows without loading all the data to memory.
- Read image subsets without reading the whole image.
- Write subsets to existing images.
- Write and read variable length table columns.
- Read images and tables using slice notation similar to
numpyarrays. (This is like a more powerfulmemmap, since it is column-aware for tables.) - Append rows to an existing table.
- Delete row sets and row ranges, resize tables, or insert rows.
- Query the columns and rows in a table.
- Read and write header keywords.
- Read and write images in tile-compressed format (
RICE,GZIP,PLIO,HCOMPRESS). - Read/write
GZIPfiles directly. - Read unix compress (
.Z,.zip) and bzip2 (.bz2) files. TDIMinformation is used to return array columns in the correct shape.- Write and read string table columns, including array columns of arbitrary shape.
- Read and write complex, bool (logical), unsigned integer, signed bytes types.
- Write checksums into the header and verify them.
- Insert new columns into tables in-place.
- Iterate over rows in a table. Data are buffered for efficiency.
- Python 3 support, including Python 3 strings.
import fitsio
from fitsio import FITS,FITSHDR
# Often you just want to quickly read or write data without bothering to
# create a FITS object. In that case, you can use the read and write
# convienience functions.
# read all data from the first hdu that has data
filename='data.fits'
data = fitsio.read(filename)
# read a subset of rows and columns from a table
data = fitsio.read(filename, rows=[35,1001], columns=['x','y'], ext=2)
# read the header
h = fitsio.read_header(filename)
# read both data and header
data,h = fitsio.read(filename, header=True)
# open the file and write a new binary table extension with the data
# array, which is a numpy array with fields, or "recarray".
data = np.zeros(10, dtype=[('id','i8'),('ra','f8'),('dec','f8')])
fitsio.write(filename, data)
# Write an image to the same file. By default a new extension is
# added to the file. use clobber=True to overwrite an existing file
# instead. To append rows to an existing table, see below.
fitsio.write(filename, image)
#
# the FITS class gives the you the ability to explore the data, and gives
# more control
#
# open a FITS file for reading and explore
fits=fitsio.FITS('data.fits')
# see what is in here; the FITS object prints itself
print(fits)
file: data.fits
mode: READONLY
extnum hdutype hduname
0 IMAGE_HDU
1 BINARY_TBL mytable
# at the python or ipython prompt the fits object will
# print itself
>>> fits
file: data.fits
... etc
# explore the extensions, either by extension number or
# extension name if available
>>> fits[0]
file: data.fits
extension: 0
type: IMAGE_HDU
image info:
data type: f8
dims: [4096,2048]
# by name; can also use fits[1]
>>> fits['mytable']
file: data.fits
extension: 1
type: BINARY_TBL
extname: mytable
rows: 4328342
column info:
i1scalar u1
f f4
fvec f4 array[2]
darr f8 array[3,2]
dvarr f8 varray[10]
s S5
svec S6 array[3]
svar S0 vstring[8]
sarr S2 array[4,3]
# See bottom for how to get more information for an extension
# [-1] to refers the last HDU
>>> fits[-1]
...
# if there are multiple HDUs with the same name, and an EXTVER
# is set, you can use it. Here extver=2
# fits['mytable',2]
# read the image from extension zero
img = fits[0].read()
img = fits[0][:,:]
# read a subset of the image without reading the whole image
img = fits[0][25:35, 45:55]
# read all rows and columns from a binary table extension
data = fits[1].read()
data = fits['mytable'].read()
data = fits[1][:]
# read a subset of rows and columns. By default uses a case-insensitive
# match. The result retains the names with original case. If columns is a
# sequence, a numpy array with fields, or recarray is returned
data = fits[1].read(rows=[1,5], columns=['index','x','y'])
# Similar but using slice notation
# row subsets
data = fits[1][10:20]
data = fits[1][10:20:2]
data = fits[1][[1,5,18]]
# Using EXTNAME and EXTVER values
data = fits['SCI',2][10:20]
# Slicing with reverse (flipped) striding
data = fits[1][40:25]
data = fits[1][40:25:-5]
# all rows of column 'x'
data = fits[1]['x'][:]
# Read a few columns at once. This is more efficient than separate read for
# each column
data = fits[1]['x','y'][:]
# General column and row subsets.
columns=['index','x','y']
rows = [1, 5]
data = fits[1][columns][rows]
# data are returned in the order requested by the user
# and duplicates are preserved
rows = [2, 2, 5]
data = fits[1][columns][rows]
# iterate over rows in a table hdu
# faster if we buffer some rows, let's buffer 1000 at a time
fits=fitsio.FITS(filename,iter_row_buffer=1000)
for row in fits[1]:
print(row)
# iterate over HDUs in a FITS object
for hdu in fits:
data=hdu.read()
# Note dvarr shows type varray[10] and svar shows type vstring[8]. These
# are variable length columns and the number specified is the maximum size.
# By default they are read into fixed-length fields in the output array.
# You can over-ride this by constructing the FITS object with the vstorage
# keyword or specifying vstorage when reading. Sending vstorage='object'
# will store the data in variable size object fields to save memory; the
# default is vstorage='fixed'. Object fields can also be written out to a
# new FITS file as variable length to save disk space.
fits = fitsio.FITS(filename,vstorage='object')
# OR
data = fits[1].read(vstorage='object')
print(data['dvarr'].dtype)
dtype('object')
# you can grab a FITS HDU object to simplify notation
hdu1 = fits[1]
data = hdu1['x','y'][35:50]
# get rows that satisfy the input expression. See "Row Filtering
# Specification" in the cfitsio manual (note no temporary table is
# created in this case, contrary to the cfitsio docs)
w=fits[1].where("x > 0.25 && y < 35.0")
data = fits[1][w]
# read the header
h = fits[0].read_header()
print(h['BITPIX'])
-64
fits.close()
# now write some data
fits = FITS('test.fits','rw')
# create a rec array. Note vstr
# is a variable length string
nrows=35
data = np.zeros(nrows, dtype=[('index','i4'),('vstr','O'),('x','f8'),
('arr','f4',(3,4))])
data['index'] = np.arange(nrows,dtype='i4')
data['x'] = np.random.random(nrows)
data['vstr'] = [str(i) for i in xrange(nrows)]
data['arr'] = np.arange(nrows*3*4,dtype='f4').reshape(nrows,3,4)
# create a new table extension and write the data
fits.write(data)
# can also be a list of ordinary arrays if you send the names
array_list=[xarray,yarray,namearray]
names=['x','y','name']
fits.write(array_list, names=names)
# similarly a dict of arrays
fits.write(dict_of_arrays)
fits.write(dict_of_arrays, names=names) # control name order
# append more rows to the table. The fields in data2 should match columns
# in the table. missing columns will be filled with zeros
fits[-1].append(data2)
# insert a new column into a table
fits[-1].insert_column('newcol', data)
# insert with a specific colnum
fits[-1].insert_column('newcol', data, colnum=2)
# overwrite rows
fits[-1].write(data)
# overwrite starting at a particular row. The table will grow if needed
fits[-1].write(data, firstrow=350)
# create an image
img=np.arange(2*3,dtype='i4').reshape(2,3)
# write an image in a new HDU (if this is a new file, the primary HDU)
fits.write(img)
# write an image with rice compression
fits.write(img, compress='rice')
# control the compression
fimg=np.random.normal(size=2*3).reshape(2, 3)
fits.write(img, compress='rice', qlevel=16, qmethod='SUBTRACTIVE_DITHER_2')
# lossless gzip compression for integers or floating point
fits.write(img, compress='gzip', qlevel=None)
fits.write(fimg, compress='gzip', qlevel=None)
# overwrite the image
fits[ext].write(img2)
# write into an existing image, starting at the location [300,400]
# the image will be expanded if needed
fits[ext].write(img3, start=[300,400])
# change the shape of the image on disk
fits[ext].reshape([250,100])
# add checksums for the data
fits[-1].write_checksum()
# can later verify data integridy
fits[-1].verify_checksum()
# you can also write a header at the same time. The header can be
# - a simple dict (no comments)
# - a list of dicts with 'name','value','comment' fields
# - a FITSHDR object
hdict = {'somekey': 35, 'location': 'kitt peak'}
fits.write(data, header=hdict)
hlist = [{'name':'observer', 'value':'ES', 'comment':'who'},
{'name':'location','value':'CTIO'},
{'name':'photometric','value':True}]
fits.write(data, header=hlist)
hdr=FITSHDR(hlist)
fits.write(data, header=hdr)
# you can add individual keys to an existing HDU
fits[1].write_key(name, value, comment="my comment")
# Write multiple header keys to an existing HDU. Here records
# is the same as sent with header= above
fits[1].write_keys(records)
# write special COMMENT fields
fits[1].write_comment("observer JS")
fits[1].write_comment("we had good weather")
# write special history fields
fits[1].write_history("processed with software X")
fits[1].write_history("re-processed with software Y")
fits.close()
# using a context, the file is closed automatically after leaving the block
with FITS('path/to/file') as fits:
data = fits[ext].read()
# you can check if a header exists using "in":
if 'blah' in fits:
data=fits['blah'].read()
if 2 in f:
data=fits[2].read()
# methods to get more information about extension. For extension 1:
f[1].get_info() # lots of info about the extension
f[1].has_data() # returns True if data is present in extension
f[1].get_extname()
f[1].get_extver()
f[1].get_extnum() # return zero-offset extension number
f[1].get_exttype() # 'BINARY_TBL' or 'ASCII_TBL' or 'IMAGE_HDU'
f[1].get_offsets() # byte offsets (header_start, data_start, data_end)
f[1].is_compressed() # for images. True if tile-compressed
f[1].get_colnames() # for tables
f[1].get_colname(colnum) # for tables find the name from column number
f[1].get_nrows() # for tables
f[1].get_rec_dtype() # for tables
f[1].get_rec_column_descr() # for tables
f[1].get_vstorage() # for tables, storage mechanism for variable
# length columns
# public attributes you can feel free to change as needed
f[1].lower # If True, lower case colnames on output
f[1].upper # If True, upper case colnames on output
f[1].case_sensitive # if True, names are matched case sensitiveThe easiest way is using pip or conda. To get the latest release
pip install fitsio
# update fitsio (and everything else)
pip install fitsio --upgrade
# if pip refuses to update to a newer version
pip install fitsio --upgrade --ignore-installed
# if you only want to upgrade fitsio
pip install fitsio --no-deps --upgrade --ignore-installed
# for conda, use conda-forge
conda install -c conda-forge fitsioYou can also get the latest source tarball release from
https://pypi.python.org/pypi/fitsioor the bleeding edge source from GitHub or use git. To check out the code for the first time
git clone https://github.com/esheldon/fitsio.gitOr at a later time to update to the latest
cd fitsio
git updateUse tar xvfz to unpack the file, enter the fitsio directory and type
pip install .- python >=3.10
- a C compiler and build tools like
make,patch, etc. - numpy (See the note below. Generally, numpy 1.11 or later is better.)
There is a serious performance regression in numpy 1.10 that results
in fitsio running tens to hundreds of times slower. A fix may be
forthcoming in a later release. Please comment on GitHub issue
numpy/issues/6467
here if this has already impacted your work
The unit tests should all pass for full support.
pytest fitsioSome tests may fail if certain libraries are not available, such as bzip2. This failure only implies that bzipped files cannot be read, without affecting other functionality.
We use the pre-commit framework for linting and code formatting. To
run the linting and code formatting, use the following command
pre-commit run -aWe bundle cfitsio partly because many deployed versions of cfitsio in the wild do not have support for interesting features like tiled image compression. Bundling a version that meets our needs is a safe alternative.
Since numpy uses C order, FITS uses fortran order, we have to write the TDIM and image dimensions in reverse order, but write the data as is. Then we need to also reverse the dims as read from the header when creating the numpy dtype, but read as is.
As of version 1.0.0, fitsio has been transitioned to setuptools for packaging
and installation. There are many reasons to do this (and to not do this). However,
at a practical level, what this means for you is that you may have trouble uninstalling
older versions with pip via pip uninstall fitsio. If you do, the best thing to do is
to manually remove the files manually. See this stackoverflow question
for example.
As of version 1.0.0, fitsio now supports Python 3 strings natively. This support
means that for Python 3, native strings are read from and written correctly to
FITS files. All byte string columns are treated as ASCII-encoded unicode strings
as well. For FITS files written with a previous version of fitsio, the data
in Python 3 will now come back as a string and not a byte string. Note that this
support is not the same as full unicode support. Internally, fitsio only supports
the ASCII character set.
fitsio is a Python wrapper for the cfitsio library and so inherits the constraints
on multithreaded programs from cfitsio. Specifically this means that
- Concurrent reading from FITS files is thread-safe, but every thread must open the FITS file
on its own, getting a unique
fitsio.FITSobject. - Concurrent writing to FITS files is NOT thread-safe.
fitsio.FITSfile objects can be shared between threads for reading, but only one thread can use the file object at a time. On Python 3.13 or newer,fitsioemploys a lock on the underlyingcfitsiodata structure to enforce this condition and help prevent race conditions. Even with this lock, you will likely need to employ your own locks from thethreadingmodule in order to prevent race conditions arising from how thefitsiolibrary is being used. See the example below.
fitsio is compatible with Python free threading, and will not reenable the GIL
when imported. However, the constraints above must be respected even when using Python
free threading.
Here is an example of using a lock to share a fitsio.FITS file pointer across threads:
import concurrent.futures
import threading
import fitsio
lock = threading.RLock()
def _read_file(fp):
with lock:
# do something with fp here
pass
with fitsio.FITS(fname) as fp:
with ThreadPoolExecutor(max_workers=10) as exc:
futs = [
exc.submit(_read_file, fp) for _ in range(10)
]
for fut in futs:
res = fut.result()On Python 3.13 and above, we release the GIL (for GIL-enabled Python builds) or detach the thread state (for free-threading Python builds). Some background information is helpful in understanding how this works.
- In the Python C API, the GIL and the thread state (i.e., attached or detached) are two separate concepts. A thread that is attached to the Python C runtime can make calls into it, use data from it, etc. In GIL-enabled builds of Python, only one thread can be attached at a time, and the GIL is the lock that enforces this constraint. In free-threading builds of Python, the interpreter must sometimes "stop the world" in order to do key tasks (e.g., garbage collection). Thus threads still must either attach to the runtime or not (but there is no constraint on how many threads can be attached, and thus no GIL). For performance reasons, if a thread is doing I/O or some other long-running computation where it does not need the Python runtime, it is good to detach it so that any "stop the world" tasks are not blocked.
- The Python C API uses the same functions for these handling both the GIL
and the thread state (wrapped in the
*_NOGILmacros in the C code.) - The
cfitsiolibrary can be compiled in such a way that it is "reentrant." Versions of the library that are reentrant allow library functions to be called concurrently by different threads, but only on different FITS file handles. Even in reentrant builds of thecfitsiolibrary, it is not safe to call library functions concurrently on the same FITS file handle. The typical way to manage access to reentrant libraries is via a lock on the data structure returned by the library (i.e., the FITS file handle). - When dealing with both the GIL and a lock for reentrant libraries, it is very easy to create deadlocks (i.e., two threads that are each waiting on one another). To help with this, on Python 3.13 or newer, the Python C API provides a lock that is hooked into the Python runtime in such a way that it will not deadlock with the GIL.
The Python Free-Threading Guide is a very useful resource for learning more about the concepts above.
To enforce the threading constraints, we use the following macros in the C layer:
LOCK_FITS(x)&UNLOCK_FITS(x): These macros take a pointer to thePyFITSObjectobject, and lock/unlock the underlying FITS file pointer for use by a single thread. This lock is not reentrant (i.e., everyLOCK_FITScall must be paired with anUNLOCK_FITScall). The implementation of this lock uses the one from the Python C API so it will not deadlock with the GIL-related macros below.ALLOW_NOGIL: This macro defines variables needed for handling the GIL/thread state, and it must be used in any C function where the other GIL-related macros below are used.RELEASE_GIL&CAPTURE_GIL: These macros are used to actually release/capture the GIL and/or attach/detach the thread state. Like the locks, these macros are not reentrant and so everyRELEASE_GILcall must be paired with aCAPTURE_GILcall.NOGIL(x): This macro wraps a single function call with the release and capture operations, returning the value of the function call. It can be used to make code more concise. You cannot use this macro in between calls toRELEASE_GILandCAPTURE_GIL.
All of these macros (except NOGIL) must be followed by a semicolon when used in C code
(e.g., ALLOW_NOGIL;). You must also take care to properly unlock the FITS file pointer
and/or release the GIL for all possible execution paths through your function (including
branches for error handling). C goto statements can be very helpful for this task.
In the C wrapper of cfitsio on Python 3.13 and above, we always lock the underlying FITS
file pointer, and we do our best to release the GIL during I/O and/or long-running operations.
- HDU groups: does anyone use these? If so open an issue!