This page gives an overview of the main concepts in Gammapy. It is a theoretical introduction to Gammapy, explaining what data, packages, classes and methods are involved in a data analysis with Gammapy, proviging many links to other documentation pages and tutorials, but not giving code examples here. For a hands-on introduction how to use Gammapy, see Installation, Getting Started and the many Tutorials.
Gammapy is an open-source Python package for gamma-ray astronomy built on Numpy and Astropy. It is a prototype for the Cherenkov Telescope Array (CTA) science tools, and can also be used to analyse data from existing gamma-ray telescopes, if their data is available in the standard FITS format (gadf). There is good support for CTA and existing imaging atmospheric Cherenkov telescopes (IACTs, e.g. H.E.S.S., MAGIC, VERITAS), and some analysis capabilities for Fermi-LAT and HAWC and multi-mission joint likelihood analysis.
To use Gammapy you need a basic knowledge of Python, Numpy, Astropy, as well as matplotlib for plotting. Many standard gamma-ray analyses can be done with few lines of configuration and code, so you can get pretty far by copy and pasting and adapting the working examples from the Gammapy documentation. But eventually, if you want to script more complex analyses, or inspect analysis results or intermediate analysis products, you need to acquire a basic to intermediate Python skill level.
To analyse data from CTA or existing IACTs, the usual workflow is to use the
high-level interface in gammapy.analysis as shown in the
example First analysis tutorial notebook, i.e.
to write a YAML config file, and then to use
Analysis to perform the data reduction from event lists
and instrument response functions (IRFs) to a reduced data format called
datasets, using either 3D cube analysis or 1D region-based spectral analysis.
The IACT data distributed by instruments is called “data level 3” (DL3) and is
given as FITS files, as shown in the CTA with Gammapy
and H.E.S.S. with Gammapy notebooks and explained in
more detail in Data reduction below. Then
class is then used to compute intermediate reduced analysis files like counts
and exposure maps or spectra, and reduced point spread function (PSF) or energy
dispersion (EDISP) information, combined in container objects called datasets
The second step is then typically to model and fit the datasets, either
individually, or in a joint likelihood analysis, using the
Fit and model classes (see Modeling and Fitting below).
You can specify your model using a YAML model specification, or write Python
code to specify which spectral and spatial models to use and what their rough
parameters are to start the fit (such as sky position and source extension, or
approximate flux level). Methods to run global model fits are available, as well
as methods to compute flux points or light curves, or run simple source
The analysis config file and
Analysis class currently mostly
scripts the data reduction up to the datasets level for the most common analysis
cases. It might be extended in the future to become the “manager” or “driver”
class for modeling or fitting as well, or that might remain the responsibility
of the datasets, models and fit classes. Advanced users that need to run
specialises analyses such as e.g. complex background modeling, or grouping of
observations, have a second-level API available via dataset makers, that offer
more flexibility. An example of this is shown in the Second analysis tutorial
Gammapy ships with a
gammapy command line tool, that can be used to check
your installation and show version information via
gammapy info, to download
example datasets and tutorials via
gammapy download or to bootstrap an
analysis by creating a default config file via
gammapy analysis. To learn
about the Gammapy command line tool, see gammapy.scripts.
As already mentioned in the Workflow section above, IACT analysis starts with data level 3 (DL3) FITS files consisting of event lists, instrument response information (effective area, point spread function, energy dispersion, background) and extra information concerning the observation (pointing direction, time), as well as two index tables that list the observations and declare which response should be used with which event data.
The main classes in Gammapy to access that DL3 data are the
EventList and various other
classes. To learn how to work with this data, and how to reduce it to the
datasets level, see gammapy.analysis, gammapy.data, gammapy.irf, gammapy.cube,
gammapy.spectrum. Note that the current IACT DL3 data model
and format is a prototype (documented at gadf), in the coming years CTA will
develop and define it’s release data format, and Gammapy and other IACTs will
adapt to that.
There are many data reduction options, but the main ones are whether to do a 3D cube analysis or a 1D spectral analysis, and whether to keep individual observations as separate datasets for a joint likelihood fit or whether to group and stack them. Partly background modeling choices are also already made at this data reduction stage. If you have a deep IACT observation, e.g. 100 observation runs, the data reduction can take a while. So typically you write the output datasets to file after data reduction, allowing you to read them back at any time later for modeling and fitting.
Datasets in Gammapy contain reduced data, models, and the likelihood function
fit statistic for a given set of model parameters. All datasets contain a
SkyModels container with one or more
SkyModel objects that represent additive emission
To model and fit data in Gammapy, you have to create a
Datasets container object with one or multiple
Dataset objects. Gammapy has built-in support to create and
analyse the following datasets:
The map datasets represent 3D cubes (
WcsNDMap objects) with two
spatial and one energy axis. For 2D images the same map objects and map datasets
are used, an energy axis is present but only has one energy bin. The
MapDataset contains a counzts map, background is modeled with a
BackgroundModel, and the fit statistic used is
MapDatasetOnOff contains on and off count maps,
background is implicitly modeled via the off counts map, and the
The spectrum datasets represent 1D spectra (
objects) with an energy axis. There are no spatial axes or information, the 1D
spectra are obtained for a given on region. The
SpectrumDataset contains a counts spectrum, background is
modeled with a
CountsSpectrum, and the fit statistic used is
SpectrumDatasetOnOff contains on on and off
count spectra, background is implicitly modeled via the off counts spectrum, and
wstat fit statistic. The
FluxPoints and a spectral model, the fit statistic used is
Modeling and Fitting¶
Assuming you have prepared your gamma-ray data as a set of
Dataset objects (see Datasets above), and
stored one or more datasets in a
Datasets container, you are
all set for modeling and fitting. Either via a YAML config file, or via Python
code, define the
SkyModels to use, which is a list of
SkyModel objects representing additive emission
components, usually sources or diffuse emission, although a single source can
also be modeled by multiple components if you want. The
SkyModel is a factorised model with a
SpectralModel component and a
SpatialModel component. Most commonly used models in
gamma-ray astronomy are built-in, see the Model gallery. It is easy to create user-defined models and
datasets, Gammapy is very flexible.
Fit class provides methods to fit, i.e. optimise
parameters and estimate parameter errors and correlations. It interfaces with a
Datasets object, which in turn is connected to a
SkyModels object, which has a
Parameters object, which contains the model parameters.
iminuit is used as modeling and fitting backend, in the future we
plan to support other optimiser and error estimation methods, e.g. from
scipy. Models can be unique for a given dataset, or contribute to multiple
datasets and thus provide links, allowing e.g. to do a joint fit to multiple
IACT datasets, or to a joint IACT and Fermi-LAT dataset. Many examples are given
in the tutorials.
To learn more about modeling and fitting, see gammapy.modeling.
To learn more about time, see gammapy.time.
Gammapy currently supports binned simulation, Poisson fluctuation of predicted counts maps. The following tutorials illustrate how to use that to predict observability, significance and sensitivity, using CTA examples: 3D map simulation, 1D spectrum simulation, and Point source sensitivity.
Development of event sampling is work in progress, currently Gammapy cannot be used yet to simulate DL3 events data.
Gammapy is organised in sub-packages, containing many classes and functions. In
this overview we only mentioned the most important concepts and parts to get
started. To learn more, see the following sub packages and documentation pages:
gammapy.data, gammapy.irf, gammapy.maps, gammapy.catalog, gammapy.astro,
gammapy command line tool).
Note that in Gammapy, 2D image analyses are partly done with actual 2D images that don’t have an energy axis (e.g. in gammapy.detect), and partly with 3D cubes with a single energy bin, e.g. for modeling and fitting, see the 2D map analysis tutorial.
For 1D spectral modeling and fitting,
used, to provide uniformity within Gammapy, and to allow in future versions of
Gammapy for advanced use cases where a sky region based analysis is used
resulting in 1D counts spetra, but the modeling is done with a spatial model
assumption, allowing for treatment of overlapping emission components, such as
e.g. a gamma-ray binary with underlying emission from a pulsar wind nebula, to
apply proper treatment of containment and contamination corrections. Note that
the spatial model on a
SkyModel is optional, you can
only pass a
SpectralModel, as shown in the First
analysis tutorial notebook and other tutorials.