Implement NaimaModel wrapper class #2124

luca-giunti · 2019-04-18T15:20:21Z

This PR introduces a preliminary implementation for a interface with Naima models.

A few things are still not in place (see the TODO list), but I would like to receive some feedback before carrying on!

I have added a new SpectralModel, called NaimaModel. The purpose of this class is wrapping the models defined in naima.models/naima.radiative, i.e. translating the parameters in a format that allows to perform a spectral fit directly within gammapy.

In this implementation, the user is supposed to have his own Naima installation. He needs to define a particle distribution and a radiative model, and pass the latter as an argument to a new NaimaModel instance:

 >>> particle_distribution = Naima.models.ExponentialCutoffPowerLaw(
 ...       amplitude =1e34 / u.eV,
 ...           e_0 =3 * u.TeV,
 ...           alpha = 2.7,
 ...           e_cutoff =10 * u.TeV,
... )
>>> radiative_model = Naima.radiative.InverseCompton(
...            particle_distribution,
...            seed_photon_fields=["CMB"],
...            Eemin=100 * u.GeV,
... )
>>> model = NaimaModel(radiative_model)

The model parameters and are left free by default. They can be freezed/unfreezed through the .freeze()method:

>>> model.parameters.parameters
[Parameter(name='amplitude', value=1e+34, factor=1e+34, scale=1.0, unit='eV-1', min=nan, max=nan, frozen=False),
 Parameter(name='e_0', value=3.0, factor=3.0, scale=1.0, unit='TeV', min=nan, max=nan, frozen=False),
 Parameter(name='alpha', value=2.7, factor=2.7, scale=1.0, unit='', min=nan, max=nan, frozen=False),
 Parameter(name='e_cutoff', value=10.0, factor=10.0, scale=1.0, unit='TeV', min=nan, max=nan, frozen=False),
 Parameter(name='beta', value=1.0, factor=1.0, scale=1.0, unit='', min=nan, max=nan, frozen=False)]
>>> model.freeze('e_0')
>>> model.freeze('beta')

The distance to the source is 1 kpc by default, but can be changed. In the case of an inverse Compton radiative model, instantiated with a list of seed_photon_fields, the model can still be evaluated using only a subset of that list. This might be useful, for example, for plotting purpose: indeed, one may want to see a different curve for each separate photon field.

As a test, I have reproduced the joint-crab paper fit (see the notebook).

TODO:

Implement a way to set min/max values for the parameters;
add the possibility to fit the magnetic field strenght;
add a way to access the total energy in electrons/protons (useful to check if the fit result is realistic)
give the possibility to normalize the model using the (observed) gamma-ray flux, instead of the amplitude of the particle distribution. (Not easy probably)
add documentation (possibly including an example)

Any idea or comment is welcome! Cheers

registerrier · 2019-04-23T07:16:52Z

gammapy/spectrum/tests/test_models.py

@@ -1,6 +1,7 @@
 # Licensed under a 3-clause BSD style license - see LICENSE.rst
 import pytest
 import numpy as np
+from naima import models, radiative


shouldn't there be a @require_dependency("naima") to avoid some build fails?

also you should call naima specific code in a way that is clearly different from spectrum.models.
e.g. import naima and later naima.models and naima.radiative
or from naima import models as naima_models

Thank you! I apply all the changes you suggest

The @require_dependency("naima") decorator should be just used for the test_models() function?

Yes, it's OK if the actual code fails with an ImportError, but for the test we want to skip if the import fails.

registerrier · 2019-04-23T07:18:11Z

gammapy/spectrum/models.py

+
+        if distance == None:
+            distance = self.distance.quantity
+        eval = self.evaluate(


you should avoid using eval as a variable name. It is a python built-in.

registerrier · 2019-04-23T07:21:40Z

gammapy/spectrum/tests/test_models.py

@@ -1,6 +1,7 @@
 # Licensed under a 3-clause BSD style license - see LICENSE.rst
 import pytest
 import numpy as np
+from naima import models, radiative


also you should call naima specific code in a way that is clearly different from spectrum.models.
e.g. import naima and later naima.models and naima.radiative
or from naima import models as naima_models

adonath

Thanks @luca-giunti! I've left a few comments to address...

I noticed that the parameter values for the radiative_model and the NaimaModel are not synchronised:

from naima import models, radiative
from gammapy.spectrum.models import NaimaModel
from astropy import units as u

pwl = models.PowerLaw(amplitude=2e33 / u.eV, e_0=10 * u.TeV, alpha=2.5)
ic = radiative.InverseCompton(pwl, seed_photon_fields=["CMB"])

naima_model = NaimaModel(ic)

naima_model.parameters["alpha"].value = 10

print(naima_model.radiative_model.particle_distribution.alpha)

This might lead to confusion, but I'm not sure if there is an easy way around this...

adonath · 2019-04-23T11:30:27Z

gammapy/spectrum/models.py

+
+        super().__init__(parameters)
+
+    def __call__(self, energy, seed=None, distance=None):


If possible let's try to avoid implementing the __call__ method for sub-classes. The .evaluate() method does not have to be static-method. So once you make this change, I think there is not need to implement the __call__ method anymore.

adonath · 2019-04-23T11:32:02Z

gammapy/spectrum/models.py

+            energy.flatten(), self.radiative_model, seed, distance, kwargs
+        )
+
+        return eval.reshape(energy.shape)


Please move the .flatten() and .reshape() call to .evaluate() and make a short inline comment, why this is necessary. I guess the naima models don't support Numpy broadcasting for some reason?

Yes, exactly. Flattening the input energy list and later reshaping the flux list ensures the
sane behaviour of all the radiative models. Namely, it prevents the InverseCompton and Synchrotron models from having broadcasting problems

adonath · 2019-04-23T11:32:19Z

gammapy/spectrum/models.py

+
+
+class NaimaModel(SpectralModel):
+    r""""""


Please implement the docstring with the parameters and also add an example, how this class is used.

adonath · 2019-04-23T11:35:52Z

gammapy/spectrum/models.py

+
+        self.parameters.parameters[par_idx].frozen = freeze
+
+    @staticmethod


.evaluate() does not have to be a static method. Once you make this change you can directly use self.radiative_model and you don't have to pass it to .evaluate() anymore. Maybe we should also take seed as an argument on __init___ (NaimaModel(seed="")).

adonath · 2019-04-23T11:39:15Z

gammapy/spectrum/models.py

+        else:
+            dnde = radiative_model.flux(energy, seed=seed, distance=distance)
+
+        return dnde.to("cm-2 s-1 TeV-1")


In principle there is no need to convert to specific units, because users can always do it themselves and for likelihood fitting or model plotting we do it in the Gammapy code at the corresponding position in the code. I have a slight preference to remove this...but I see that the default units returned here are probably very unusual.

I think that it would be nice if the energy part of the flux was returned in the same units as the input energy array. So for example, if I compute the model using an array like [1.00000000e-02, 3.72759372e-02, 1.38949549e-01] MeV I get a flux in units of cm-2 s-1 MeV-1. Do you think this might make sense? If not, I will just drop any specification and leave it to the default units (which are cm-2 s-1 eV-1)

Yes, I agree this would make sense from a user-perspective. You can achieve this by creating a composite unit like so unit = 1 / (u.s * u.sr * energy.unit).

adonath · 2019-04-23T11:41:26Z

gammapy/spectrum/tests/test_models.py

@@ -1,6 +1,7 @@
 # Licensed under a 3-clause BSD style license - see LICENSE.rst
 import pytest
 import numpy as np
+from naima import models, radiative


Yes, it's OK if the actual code fails with an ImportError, but for the test we want to skip if the import fails.

adonath · 2019-04-23T11:42:45Z

gammapy/spectrum/models.py

+
+    def __init__(self, radiative_model, distance=1.0 * u.kpc):
+        self.radiative_model = radiative_model
+        self.distance = Parameter("distance", distance)


I guess it's probably a good choice to freeze the distance parameter by default.

adonath · 2019-04-23T11:46:23Z

gammapy/spectrum/models.py

+        self.distance = Parameter("distance", distance)
+
+        parameters = []
+        parameters_dict = self.radiative_model.particle_distribution.__dict__


Note that there is a particle_distribution.param_names attribute you can use. You don't have to relay on the private __dict__.

adonath · 2019-04-23T11:54:06Z

gammapy/spectrum/models.py

+        for (name, quantity) in parameters_dict.items():
+            if name[0] == "_" or name == "unit":
+                continue
+            parameters.append(Parameter(name, quantity))


Please also set the parameters as an attribute on the NaimaModel object. So that users can access it like parameters on other spectral models in Gammapy.

The problem here is that you don't know in principle what are the parameters of the model.. If we want to have this general NaimaModel class supporting all Naima models I am not sure how to implement the change you suggest. In other words, I don't see how to make something like this work:

naima_model = NaimaModel(some_radiative_model) naima_model.apha

Maybe the NaimaModel class could have a long __slots__ list containing all possible model parameters.. but this doesn't seem an ideal solution

Thanks for the explanation, indeed this is a problem. Technically you could use something along the lines of:

for par in radiative_model.particle_distribution.param_names: setter(self, par, Parameter())

Declaring the parameters in __slots__ is not strictly required by the design, but is basically just what we use to avoid that users can set new attributes after init. It might be OK, to leave this part out for the NaimaModel for now and maybe change to a different solution for the new attribute handling later.

The other solution I see would be to implement a dedicated wrapper for all radiative models, very similar to what is done in https://github.com/zblz/naima/blob/master/naima/sherpa_models.py. Not sure if this is good a option...

adonath · 2019-04-23T11:58:24Z

gammapy/spectrum/models.py

+
+        return eval.reshape(energy.shape)
+
+    def freeze(self, name, freeze=True):


Why is this needed? I think once the parameters are available as attributes, users can just do model.distance.frozen = False...

luca-giunti · 2019-04-30T10:40:45Z

Hi! @adonath, I did my best to apply all requested changes.

Thanks @luca-giunti! I've left a few comments to address...

I noticed that the parameter values for the radiative_model and the NaimaModel are not synchronised:
from naima import models, radiative
from gammapy.spectrum.models import NaimaModel
from astropy import units as u

pwl = models.PowerLaw(amplitude=2e33 / u.eV, e_0=10 * u.TeV, alpha=2.5)
ic = radiative.InverseCompton(pwl, seed_photon_fields=["CMB"])

naima_model = NaimaModel(ic)

naima_model.parameters["alpha"].value = 10

print(naima_model.radiative_model.particle_distribution.alpha)
This might lead to confusion, but I'm not sure if there is an easy way around this...

Indeed, I did not find an easy way to fix this. However, I guess that now that parameters are exposed as attributes the temptation of doing naima_model.radiative_model.particle_distribution.alpha instead of naima_model.alpha will be rather low..

The example in the Docstring produces the following plot:

Here is a link to the notebook where I use the NaimaModel to reproduce the joint-crab paper fit (Note that I stil used the SpectrumFit class..)

cdeil

a small comment on docstring / Sphinx inline.

cdeil · 2019-05-03T08:10:26Z

gammapy/spectrum/models.py

+
+
+class NaimaModel(SpectralModel):
+    r"""A wrapper for `Naima <https://naima.readthedocs.io/en/latest/>`_ models


Remove the URL here in the summary line.

cdeil · 2019-05-03T08:13:49Z

gammapy/spectrum/models.py

+
+    This class provides an interface with the models defined in the `~naima.models` module.
+    The model accepts as a positional argument a Naima
+    `radiative model <https://naima.readthedocs.io/en/latest/radiative.html>`_ instance, used


If you just put this, does it create a link to the Naima docs?

`naima`_

I think it should because of

gammapy/docs/conf.py

Line 74 in 7b186a7

intersphinx_mapping["naima"] = ("https://naima.readthedocs.io/en/latest/", None)

Please run python setup.py build_docs locally and check the HTML output.

cdeil · 2019-05-03T08:16:23Z

gammapy/spectrum/models.py

+        plt.legend(loc='best')
+        plt.show()
+    """
+    #TODO: prevent users from setting new attributes after init


Remove TODO? Or is this really something we want / need to do here?

All spectral models prevent the intialization of new attributes after __init__ because they have parameters declared in __slots__. For Naima models this turns out to be tricky (see previous answer: #2124 (comment)), but for consistency I think we will probably need to implement this TODO here

cdeil · 2019-05-03T08:33:02Z

If you add a tutorial notebook or example, please make it fast. 10 seconds is good, 1 min is max.

Tutorials that run for many minutes are bad: users have to sit and wait for the example to run, but to learn you need to play with parameters. And if the tutorial takes long to execute, then it's difficult to test in our continuous integration setup.

So far there's always been a way to make a fast example: use small dataset, simple model, use optimisation parameters so that it runs fast - one doesn't learn so much more by running 100s of iterations and have super-precise results.

luca-giunti · 2019-05-03T12:56:17Z

@cdeil I agree, especially because this is a rather minimal interface with Naima and it probably doesn't deserve a huge tutorial. I wonder if a tutorial is needed at all, to be honest. Maybe the current example plus a reference to the joint-crab paper notebook (with a brief explanation on how this class may ease the fit) could be enough?

cdeil · 2019-05-03T13:25:26Z

+1 to focus on the Gammapy package for now, and not add new tutorials here and in most other cases.

The more tutorials we have, the more work it becomes to change / polish anything, and we will do quite a bit of that in the coming months.

adonath · 2019-05-08T09:20:08Z

gammapy/spectrum/tests/test_models.py

@@ -1,6 +1,7 @@
 # Licensed under a 3-clause BSD style license - see LICENSE.rst
 import pytest
 import numpy as np
+import naima


We have to restructure the testing here, because naima is only an optional dependency. That means the import must be "delayed" i.e. within a class or function, but not at the top-level of the module. This also means the NaimaModel must be tested separately from the other models.

My suggestion would be to implement a TestNaimaModel class like so:

@requires_dependency("naima") class TestNaimaModel: def test_pion_decay_(self): import naima model = NaimaModel() assert_quantity_allclose() assert_quantity_allclose() def test_ic(self): import naima model = NaimaModel() assert_quantity_allclose() assert_quantity_allclose() def test_synchrotron(self): import naima model = NaimaModel() assert_quantity_allclose() assert_quantity_allclose()

luca-giunti · 2019-05-08T12:57:50Z

@adonath remaining fail seems unrelated with the PR?

adonath · 2019-05-08T15:28:04Z

Thanks, @luca-giunti! Yes, the remaining fail was unrelated...

luca-giunti requested a review from adonath April 18, 2019 15:20

registerrier reviewed Apr 23, 2019

View reviewed changes

adonath requested changes Apr 23, 2019

View reviewed changes

luca-giunti force-pushed the Add-Naima-spectral-models branch from 5401110 to 76e0624 Compare April 30, 2019 09:13

cdeil reviewed May 3, 2019

View reviewed changes

cdeil assigned registerrier May 3, 2019

cdeil added the feature label May 3, 2019

cdeil added this to the 0.12 milestone May 3, 2019

cdeil added this to To do in gammapy.makers via automation May 3, 2019

luca-giunti added 9 commits May 7, 2019 11:34

Added new spectral model

0f93c5b

Added a test

9593e45

Better freeze method

0930c1e

Better test

c5baf5a

Applied requested changes

ec13e58

Fixed bad try/except usage

ea2f41b

Fixed bad try/except usage

af0aa50

Fixed import issues

f20f116

Suggested changes

619753e

luca-giunti force-pushed the Add-Naima-spectral-models branch from 9ee7c00 to 619753e Compare May 7, 2019 09:34

adonath and others added 3 commits May 7, 2019 11:57

Remove NaimaModel.compute_W method and minor clean up

866e017

Move naima import to NaimaModel.__init__

6fbe601

docstring changes

95f302f

adonath reviewed May 8, 2019

View reviewed changes

Fixed test

300d034

adonath approved these changes May 8, 2019

View reviewed changes

adonath merged commit e5aaede into gammapy:master May 8, 2019

gammapy.makers automation moved this from To do to Done May 8, 2019

adonath changed the title ~~Add naima spectral models~~ Implement NaimaModel wrapper class May 27, 2019


		super().__init__(parameters)

		def __call__(self, energy, seed=None, distance=None):


		self.parameters.parameters[par_idx].frozen = freeze

		@staticmethod


		return eval.reshape(energy.shape)

		def freeze(self, name, freeze=True):



		class NaimaModel(SpectralModel):
		r"""A wrapper for `Naima <https://naima.readthedocs.io/en/latest/>`_ models

Implement NaimaModel wrapper class #2124

Implement NaimaModel wrapper class #2124

Conversation

luca-giunti commented Apr 18, 2019

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

adonath left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

adonath Apr 23, 2019 • edited

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

luca-giunti commented Apr 30, 2019

cdeil left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

cdeil commented May 3, 2019

luca-giunti commented May 3, 2019

cdeil commented May 3, 2019

adonath May 8, 2019 • edited

Choose a reason for hiding this comment

luca-giunti commented May 8, 2019 • edited

adonath commented May 8, 2019

adonath Apr 23, 2019 •

edited

adonath May 8, 2019 •

edited

luca-giunti commented May 8, 2019 •

edited