Developer HOWTO

This page is a collection of notes for Gammapy contributors and maintainers, in the form of short “How to” or “Q & A” entries.

Where should I import from?

You should import from the “end-user namespaces”, not the “implementation module”.

from import EventList  # good
from import EventList # bad

from gammapy.stats import cash  # good
from gammapy.stats.fit_statistics import cash  # bad

The end-user namespace is the location that is shown in the API docs, i.e. you can use the Sphinx full-text search to quickly find it.

To make code maintenance easier, the implementation of the functions and classes is spread across multiple modules (.py files), but the user shouldn’t care about their names, that would be too much to remember.

The only reason to import from a module directly is if you need to access a private function, class or variable (something that is not listed in __all__ and thus not imported into the end-user namespace.

Note that this means that in the definition of an “end-user namespace”, e.g. in the gammapy/data/ file, the imports have to be sorted in a way such that modules in gammapy/data are loaded when imported from other modules in that sub-package.

Functions returning several values

It is up to the developer to decide how to return multiple things from functions and methods. For up to three things, if callers usually will want access to several things, using a tuple or collections.namedtuple is OK. For three or more things, using a Python dict instead should be preferred.

Python 2 and 3 support

In Gammapy we currently support Python 2.7, as well as Python 3.5 - 3.7. Code should be written in the common subset, using the bundled gammapy.extern.six to help smooth over differences where needed.

The following import should be at the top of every file in Gammapy, to allow us to write code that’s closer to Python 3:

from __future__ import absolute_import, division, print_function, unicode_literals

We do plan do drop legacy Python (Python 2.7) support when the Fermi ST and Fermi support Python 3.

For further information, see PIG 3 in GH 1278

Skip unit tests for some Astropy versions

import astropy
import pytest

ASTROPY_VERSION = (astropy.version.major, astropy.version.minor)
@pytest.mark.xfail(ASTROPY_VERSION < (0, 4), reason="Astropy API change")
def test_something():

Fix non-Unix line endings

In the past we had non-Unix (i.e. Mac or Windows) line endings in some files. This can be painful, e.g. git diff and autopep8 behave strangely. Here’s to commands to check for and fix this (see here):

$ git clean -fdx
$ find . -type f -print0 | xargs -0 -n 1 -P 4 dos2unix -c mac
$ find . -type f -print0 | xargs -0 -n 1 -P 4 dos2unix -c ascii
$ git status
$ cd astropy_helpers && git checkout -- . && cd ..

What checks and conversions should I do for inputs?

In Gammapy we assume that “we’re all consenting adults”, which means that when you write a function you should write it like this:

def do_something(data, option):
    """Do something.

    data : `numpy.ndarray`
    option : {'this', 'that'}
    if option == 'this':
        out = 3 * data
    elif option == 'that':
        out = data ** 5
        ValueError('Invalid option: {}'.format(option))

    return out
  • Don’t always add `isinstance` checks for everything … assume the caller passes valid inputs, … in the example above this is not needed:

    assert isinstance(option, str)
  • Don’t always add `numpy.asanyarray` calls for all array-like inputs … the caller can do this if it’s really needed … in the example above document data as type ndarray instead of array-like and don’t put this line:

    data = np.asanyarray(data)
  • Do always add an `else` clause to your `if`-`elif` clauses … this is boilerplate code, but not adding it would mean users get this error if they pass an invalid option:

    UnboundLocalError: local variable 'out' referenced before assignment

Now if you really want, you can add the numpy.asanyarray and isinstance checks for functions that end-users might often use for interactive work to provide them with better exception messages, but doing it everywhere would mean 1000s of lines of boilerplate code and take the fun out of Python programming.

Float data type: 32 bit or 64 bit?

Most of the time what we want is to use 32 bit to store data on disk and 64 bit to do computations in memory.

Using 64 bit to store data and results (e.g. large images or cubes) on disk would mean a factor ~2 increase in file sizes and slower I/O, but I’m not aware of any case where we need that precision.

On the other hand, doing computations with millions and billions of pixels very frequently results in inaccurate results … e.g. the likelihood is the sum over per-pixel likelihoods and using 32-bit will usually result in erratic and hard-to-debug optimizer behaviour and even if the fit works incorrect results.

Now you shouldn’t put this line at the top of every function … assume the caller passes 64-bit data:

data = np.asanyarray(data, dtype='float64')

But you should add explicit type conversions to 64 bit when reading float data from files and explicit type conversions to 32 bit before writing to file.

Clobber or overwrite?

In Gammapy we consistently use an overwrite bool option for gammapy.scripts and functions that write to files. This is in line with Astropy, which had a mix of clobber and overwrite in the past, and has switched to uniform overwrite everywhere.

The default value should be overwrite=False, although we note that this decision was very controversial, several core developers would prefer to use overwrite=True. For discussion on this, see GH 1396.

Pixel coordinate convention

All code in Gammapy should follow the Astropy pixel coordinate convention that the center of the first pixel has pixel coordinates (0, 0) (and not (1, 1) as shown e.g. in ds9).

You should use origin=0 when calling any of the pixel to world or world to pixel coordinate transformations in astropy.wcs.

When to use C or Cython or Numba for speed

Most of Gammapy is written using Python and Numpy array expressions calling functions (e.g. from Scipy) that operate on Numpy arrays. This is often nice because it means that algorithms can be implemented with few lines of high-level code,

There is a very small fraction of code though (one or a few percent) where this results in code that is either cumbersome or too slow. E.g. to compute TS or upper limit images, one needs to do a root finding method with different number of iterations for each pixel … that’s impossible (or at least very cumbersome / hard to read) to implement with array expressions and Python loops over pixels are slow.

In these cases we encourage the use of Cython or Numba, or writing the core code in C and exposing it to Python via Cython. These are popular and simple ways to get C speed from Python.

To use several CPU cores consider using the Python standard library multiprocessing module.

Note that especially the use of Numba should be considered an experiment. It is a very nice, but new technology and no-one uses it in production. Before the Gammapy 1.0 release we will re-evaluate the status of Numba and decide whether it’s an optional dependency we use for speed, or whether we use the much more established Cython (Scipy, scikit-image, Astropy, … all use Cython).

At the time of writing (April 2015), the TS map computation code uses Cython and multiprocessing and Numba is not used yet.

Assert convention

When performing tests, the preferred numerical assert method is numpy.testing.assert_allclose. Use

from numpy.testing import assert_allclose

at the top of the file and then just use assert_allclose for the tests. This makes the lines shorter, i.e. there is more space for the arguments.

assert_allclose covers all use cases for numerical asserts, so it should be used consistently everywhere instead of using the dozens of other available asserts from pytest or numpy in various places.

For assertions on Quantity objects, you can do this to assert on the unit and value separately:

from numpy.testing import assert_allclose
import astropy.units as u

actual = 1 / 3 * u.deg
assert actual.unit == 'deg'
assert_allclose(actual.value, 0.33333333)

Note that Quantity can be compared to unit strings directly. Also note that the default for assert_allclose is atol=0 and rtol=1e-7, so when using it, you have to give the reference value with a precision of rtol ~ 1e-8, i.e. 8 digits to be on the safe side (or pass a lower rtol or set an atol).

The use of assert_quantity_allclose is discouraged, because it only requires that the values match after unit conversions. This is not so bad, but units in test cases should not change randomly, so asserting on unit and value separately establishes more behaviour.

If you don’t like the two separate lines, you can use gammapy.utils.testing.assert_quantity_allclose, which does assert that units are equal, and calls numpy.testing.assert_equal for the values.

Random numbers

All functions that need to call a random number generator should take a random_state input parameter and call the get_random_state utility function like this (you can copy & paste the three docstring lines and the first code line to the function you’re writing):

from gammapy.utils.random import get_random_state

def make_random_stuff(X, random_state='random-seed'):

    random_state : {int, 'random-seed', 'global-rng', `~numpy.random.RandomState`}
        Defines random number generator initialisation.
        Passed to `~gammapy.utils.random.get_random_state`.
    random_state = get_random_state(random_state)
    data = random_state.uniform(low=0, high=3, size=10)
    return data

This allows callers flexible control over which random number generator (i.e. which numpy.random.RandomState instance) is used and how it’s initialised. The default random_state='random-seed' means “create a new RNG, seed it in a random way”, i.e. different random numbers will be generated on every call.

There’s a few ways to get deterministic results from a script that call functions that generate random numbers.

One option is to create a single RandomState object seeded with an integer and then pass that random_state object to every function that generates random numbers:

from numpy.random import RandomState
random_state = RandomState(seed=0)

stuff1 = make_some_random_stuff(random_state=random_state)
stuff2 = make_more_random_stuff(random_state=random_state)

Another option is to pass an integer seed to every function that generates random numbers:

seed = 0
stuff1 = make_some_random_stuff(random_state=seed)
stuff2 = make_more_random_stuff(random_state=seed)

This pattern was inspired by the way scikit-learn handles random numbers. We have changed the None option of sklearn.utils.check_random_state to 'global-rng', because we felt that this meaning for None was confusing given that numpy.random.RandomState uses a different meaning (for which we use the option 'global-rng').


Gammapy is a library. This means that it should never contain print statements, because with print statements the library users have no easy way to configure where the print output goes (e.g. to stdout or stderr or a log file) and what the log level (warning, info, debug) and format is (e.g. include timestamp and log level?).

So logging is much better than printing. But also logging is only rarely needed. Many developers use print or log statements to debug some piece of code while they write it. Once it’s written and works, it’s rare that callers want it to be chatty and log messages all the time. Print and log statements should mostly be contained in end-user scripts that use Gammapy, not in Gammapy itself.

That said, there are cases where emitting log messages can be useful. E.g. a long-running algorithm with many steps can log info or debug statements. In a function that reads and writes several files it can make sense to include info log messages for normal operation, and warning or error log messages when something goes wrong. Also, command line tools that are included in Gammapy should contain log messages, informing the user about what they are doing.

Gammapy uses the Python standard library logging module. This module is extremely flexible, but also quite complex. But our logging needs are very modest, so it’s actually quite simple …

Generating log messages

To generate log messages from any file in Gammapy, include these two lines at the top:

import logging
log = logging.getLogger(__name__)

This creates a module-level logging.Logger object called log, and you can then create log messages like this from any function or method:

def process_lots_of_data(infile, outfile):'Starting processing data ...')

    # do lots of work'Writing {}'.format(outfile))

You should never log messages from the module level (i.e. on import) or configure the log level or format in Gammapy, that should be left to callers … except from command line tools …

There is also the rare case of functions or classes with the main job to check and log things. For these you can optionally let the caller pass a logger when constructing the class to make it easier to configure the logging. See the EventListDatasetChecker as an example.

Configuring logging from command line tools

Every Gammapy command line tool should have a --loglevel option:

parser.add_argument("-l", "--loglevel", default='info',
                    choices=['debug', 'info', 'warning', 'error', 'critical'],
                    help="Set the logging level")

This option is then processed at the end of main using this helper function:


This sets up the root logger with the log level and format (the format isn’t configurable for the command line scripts at the moment).

See gammapy/scripts/ as an example.

Command line tools using click

Command line tools that use the click module should disable the unicode literals warnings to clean up the output of the tool:

import click
click.disable_unicode_literals_warning = True

See here for further information.

BSD or GPL license?

Gammapy is BSD licensed (same license as Numpy, Scipy, Matplotlib, scikit-image, Astropy, photutils, yt, …).

We prefer this over the GPL3 or LGPL license because it means that the packages we are most likely to share code with have the same license, e.g. we can take a function or class and “upstream” it, i.e. contribute it e.g. to Astropy or Scipy if it’s generally useful.

Some optional dependencies of Gammapy (i.e. other packages like Sherpa or Gammalib or ROOT that we import in some places) are GPL3 or LGPL licensed.

Now the GPL3 and LGPL license contains clauses that other package that copy or modify it must be released under the same license. We take the standpoint that Gammapy is independent from these libraries, because we don’t copy or modify them. This is a common standpoint, e.g. astropy.wcs is BSD licensed, but uses the LGPL-licensed WCSLib.

Note that if you distribute Gammapy together with one of the GPL dependencies, the whole distribution then falls under the GPL license.


In Gammapy we keep a Changelog with a list of pull requests. We sort by release and within the release by PR number (largest first).

As explained in the Updating and Maintaining the Changelog section in the Astropy docs, there are (at least) two approaches for adding to the changelog, each with pros and cons.

We’ve had some pain due to merge conflicts in the changelog and having to wait until the contributor rebases (and having to explain git rebase to new contributors).

So our recommendation is that changelog entries are not added in pull requests, but that the core developer adds a changelog entry after right after having merged a pull request (you can add [skip ci] on this commit).

File and directory path handling

In Gammapy use Path objects to handle file and directory paths.

from gammapy.extern.pathlib import Path

dir = Path('folder/subfolder')
filename = dir / 'filename.fits'

Note how the / operator makes it easy to construct paths (as opposed to repeated calls to the string-handling function os.path.join) and how methods on Path objects provide a nicer interface to most of the functionality from os.path (mkdir in this example).

One gotcha is that many functions (such as table.write in this example) expect str objects and refuse to work with Path objects, so you have to explicitly convert to str(path).

Note that pathlib was added to the Python standard library in 3.4 (see here), but since we support Python 2.7 and the Python devs keep improving the version in the standard library (by adding new methods and new options for existing methods), we decided to bundle the latest version (from here) in gammapy/extern/ and that should always be used.

Bundled gammapy.extern code

We bundle some code in gammapy.extern. This is external code that we don’t maintain or modify in Gammapy. We only bundle small pure-Python files (currently all single-file modules) purely for convenience, because having to explain about these modules as Gammapy dependencies to end-users would be annoying. And in some cases the file was extracted from some other project, i.e. can’t be installed separately as a dependency.

For gammapy.extern we don’t generate Sphinx API docs. To see what is there, check out the gammapy/extern directory locally or on Github. Notes on the bundled files are kept in the docstring of gammapy/extern/

Interpolation and extrapolation

In Gammapy, we use interpolation a lot, e.g. to evaluate instrument response functions (IRFs) on data grids, or to reproject diffuse models on data grids.

Note: For some use cases that require interpolation the NDDataArray base class might be useful.

The default interpolator we use is scipy.interpolate.RegularGridInterpolator because it’s fast and robust (more fancy interpolation schemes can lead to unstable response in some cases, so more careful checking across all of parameter space would be needed).

You should use this pattern to implement a function of method that does interpolation:

def do_something(..., interp_kwargs=None):
    """Do something.

    interp_kwargs : dict or None
        Interpolation parameter dict passed to `scipy.interpolate.RegularGridInterpolator`.
        If you pass ``None``, the default ``interp_params=dict(bounds_error=False)`` is used.
    if not interp_kwargs:
        interp_kwargs = dict(bounds_error=False)

    interpolator = RegularGridInterpolator(..., **interp_kwargs)

Since the other defaults are method='linear' and fill_value=nan, this implies that linear interpolation is used and NaN values are returned for points outside of the interpolation domain. This is a compromise between the alternatives:

  • bounds_error=True – Very “safe”, refuse to return results for any points if one of the points is outside the valid domain. Can be annoying for the caller to not get any result.
  • bounds_error=False, fill_value=nan – Medium “safe”. Always return a result, but put NaN values to make it easy for analysers to spot that there’s an issue in their results (if pixels with NaN are used, that will usually lead to NaN values in high-level analysis results.
  • bounds_error=False, fill_value=0 or bounds_error=False, fill_value=None – Least “safe”. Extrapolate with zero or edge values (this is what None means). Can be very convenient for the caller, but can also lead to errors where e.g. stacked high-level analysis results aren’t quite correct because IRFs or background models or … were used outside their valid range.

Methods that use interpolation should provide an option to the caller to pass interpolation options on to RegularGridInterpolator in case the default behaviour doesn’t suit the application.

TODO: we have some classes (aeff2d and edisp2d) that pre-compute an interpolator, currently in the constructor. In those cases the interp_kwargs would have to be exposed e.g. also on the read and other constructors. Do we want / need that?

Locate origin of warnings

By default, warnings appear on the console, but often it’s not clear where a given warning originates (e.g. when building the docs or running scripts or tests) or how to fix it.

Sometimes putting this in gammapy/ can help:

import numpy as np

Following the advice here, putting this in docs/ can also help sometimes:

import traceback
import warnings
import sys

def warn_with_traceback(message, category, filename, lineno, file=None, line=None):
    log = file if hasattr(file,'write') else sys.stderr
    log.write(warnings.formatwarning(message, category, filename, lineno, line))

warnings.showwarning = warn_with_traceback

Object summary info string

If you want to add a method to provide some basic information about a class instance, you should use the Python __str__ method.

class Spam(object):
    def __init__(self, ham):
        self.ham = ham

    def __str__(self):
        ss = 'Summary Info about class Spam\n'
        ss += '{:.2f}'.format(self.ham)
        return ss

If you want to add configurable info output, please provide a method summary, like here. In this case the __str__ method should be a call to summary with default parameters. Do not use an info method, since this would lead to conflicts for some classes in Gammapy (e.g. classes that inherit the info method from astropy.table.Table.

Using the NDDataArray

Gammapy has a class for generic n-dimensional data arrays, NDDataArray. Classes that represent such an array should use this class. The goal is to reuse code for interpolation and have an coherent I/O interface, mainly in irf.

A usage example can be found in :gp-extra-notebooks:nddata_demo.

Also, consult Interpolation and extrapolation if you are not sure how to setup your interpolator.

Sphinx docs build

Generating the HTML docs for Gammapy is straight-forward:

python build_docs
open docs/_build/html/index.html

Generating the PDF docs is more complex. This should work:

python build_docs -b latex
cd docs/_build/latex
makeindex -s gammapy.idx
pdflatex -interaction=nonstopmode gammapy.tex
open gammapy.pdf

You need a bunch or LaTeX stuff, specifically texlive-fonts-extra is needed.

Jupyter notebooks present in the gammapy-extra repository are by default copied to the docs/notebooks and docs/_static/notebooks folders during the process of generating HTML docs. This triggers its conversion to Sphinx formatted HTML files and .py scripts. The Sphinx formatted versions of the notebooks provide access to the raw .ipynb Jupyter files and .py script versions stored in docs/_static/notebooks folder.

Once the documentation built you can optimize the speed of re-building processes, for example in case you are modifying or creating new docs and you would like to check these changes are displayed nicely. For that purpose, if your modified doc file does not contain links to notebooks, you may set the flag build_notebooks to False in the setup.cfg file, so they are not re-written again by Sphinx.

Documentation guidelines

Like almost all Python projects, the Gammapy documentation is written in a format called restructured text (RST) and built using Sphinx. We mostly follow the Astropy documentation guidelines, which are based on the Numpy docstring standard, which is what most scientific Python packages use.

There’s a few details that are not easy to figure out by browsing the Numpy or Astropy documentation guidelines, or that we actually do differently in Gammapy. These are listed here so that Gammapy developers have a reference.

Usually the quickest way to figure out how something should be done is to browse the Astropy or Gammapy code a bit (either locally with your editor or online on Github or via the HTML docs), or search the Numpy or Astropy documentation guidelines mentioned above. If that doesn’t quickly turn up something useful, please ask by putting a comment on the issue or pull request you’re working on on Github, or send an email to the Gammapy mailing list.

Functions or class methods that return a single object

For functions or class methods that return a single object, following the Numpy docstring standard and adding a Returns section usually means that you duplicate the one-line description and repeat the function name as return variable name. See astropy.cosmology.LambdaCDM.w or astropy.time.Time.sidereal_time as examples in the Astropy codebase. Here’s a simple example:

def circle_area(radius):
    """Circle area.

    radius : `~astropy.units.Quantity`
        Circle radius

    area : `~astropy.units.Quantity`
        Circle area
    return 3.14 * (radius ** 2)

In these cases, the following shorter format omitting the Returns section is recommended:

def circle_area(radius):
    """Circle area (`~astropy.units.Quantity`).

    radius : `~astropy.units.Quantity`
        Circle radius
    return 3.14 * (radius ** 2)

Usually the parameter description doesn’t fit on the one line, so it’s recommended to always keep this in the Parameters section.

This is just a recommendation, e.g. for gammapy.cube.SkyCube.spectral_index we decided to use this shorter format, but for gammapy.cube.SkyCube.flux we decided to stick with the more verbose format, because the return type and units didn’t fit on the first line.

A common case where the short format is appropriate are class properties, because they always return a single object. As an example see, which is reproduced here:

def radec(self):
    """Event RA / DEC sky coordinates (`~astropy.coordinates.SkyCoord`).
    lon, lat = self['RA'], self['DEC']
    return SkyCoord(lon, lat, unit='deg', frame='icrs')

Class attributes

Class attributes (data members) and properties are currently a bit of a mess, see SkyCube as an example. Attributes are listed in an Attributes section because I’ve listed them in a class-level docstring attributes section as recommended here. Properties are listed in separate Attributes summary and Attributes Documentation sections, which is confusing to users (“what’s the difference between attributes and properties?”).

One solution is to always use properties, but that can get very verbose if we have to write so many getters and setters. I don’t have a solution for this yet … for now I’ll go read this and meditate.

TODO: make a decision on this and describe the issue / solution here.

Different versions of notebooks in Binder

Jupyter notebooks may be accessed and executed on-line in the Gammapy Binder space. Each fixed-text sphinx formatted notebook present in the documentation has its own link pointing to its specific space in Gammapy Binder. Since notebooks are evolving with Gammapy functionalities and documentation, it is possible to link the different versions of the notebooks stored in GitHub repository gammapy-extra to the same versions built in Gammapy Binder. For this purpose just edit the variable git_commit in setup.cfg file and provide the branch, tag or commit of GitHub repository gammapy-extra that will be used to access the same version of the notebook in Gammapy Binder.

Include images from gammapy-extra into the docs

Similar to the gp-extra-notebook role, Gammapy has a gp-extra-image directive.

To include an image from gammapy-extra/figures/, use the gp-extra-image directive instead of the usual Sphinx image directive like this:

.. gp-extra-image:: detect/fermi_ts_image.png
    :scale: 100%

More info on the image directive is here

Coordinate and axis names

In Gammapy, the following coordinate and axis names should be used.

This applies to most of the code, ranging from IRFs to maps to sky models, for function parameters and variable names.

  • time - time
  • energy - energy
  • ra, dec - sky coordinates, radec frame (i.e. icrs to be precise)
  • glon, glat - sky coordinates, galactic frame
  • az, alt - sky coordinates, altaz frame
  • lon, lat for spherical coordinates that aren’t in a specific frame.

For angular sky separation angles:

  • psf_theta - offset wrt. PSF center position
  • fov_theta - offset wrt. field of view (FOV) center
  • theta - when no PSF is involved, e.g. to evaluate spatial sky models

For the general case of FOV coordinates that depend on angular orientation of the FOV coordinate frame:

  • fov_{frame}_lon, fov_{frame}_lat - field of view coordinates
  • fov_theta, fov_{frame}_phi - field of view polar coordinates

where {frame} can be one of radec, galactic or altaz, depending on with which frame the FOV coordinate frame is aligned.


  • In cases where it’s unclear if the value is for true or reconstructed event parameters, a postfix _true or _reco should be added. In Gammapy, this mostly occurs for energy_true and energy_reco, e.g. the background IRF has an axis energy_reco, but effective area usually energy_true, and energy dispersion has both axes. We are not pedantic about adding _true and _reco everywhere. Note that this would quickly become annoying (e.g. source models use true parameters, and it’s not clear why one should write ra_true). E.g. the property on the event list energy matches the ENERGY column from the event list table, which is for real data always reco energy.
  • Currently, no sky frames centered on the source, or non-radially symmetric PSFs are in use, and thus the case of “source frames” that have to be with a well-defined alignment, like we have for the “FOV frames” above, doesn’t occur and thus doesn’t need to be defined yet (but it would be natural to use the same naming convention as for FOV if it eventually does occur).
  • These definitions are mostly in agreement with the format spec. We do not achieve 100% consistency everywhere in the spec and Gammapy code. Achieving this seems unrealistic, because legacy formats have to be supported, we are not starting from scratch and have time to make all formats consistent. Our strategy is to do renames on I/O where needed, to and from the internal Gammapy names defined here, to the names used in the formats. Of course, where formats are not set in stone yet, we advocate and encourage the use of the names chosen here.
  • Finally, we realise that eventually probably CTA will define this, and Gammapy is only a prototype. So if CTA chooses something else, probably we will follow suite and do one more backward-incompatible change at some point to align with CTA.