Skip to content

Latest commit

 

History

History
249 lines (165 loc) · 20.4 KB

DEVEL.rst

File metadata and controls

249 lines (165 loc) · 20.4 KB
Raw

Development notes

This document gathers some notes about the development flow, release checklist, and general design decisions and their motivation. If you intend to heavily modify Cobaya, we recommend you to read it first.

Note

This document may not always be up to date

git development model

  • Non-breaking travis-passing latest changes and fixes are in master
  • Development that may break tests is done in temp branches or forks and merged to master once OK
  • Breaking changes developed in separate branches, and merged when releases updated
  • Releases are branched out, and only critical bug fixes are pushed onto them.

Development flow for contributors

  1. Fork and clone the repo from github.
  2. From its folder, install in editable mode: pip install -e .
  3. Modify stuff.
  4. Test with pytest (first "pip install pytest pytest-forked pytest-cov flaky dill")
  5. Make a pull requests and check (about about 15 minutes) if the tests have passed.
  6. Iterate until tests pass, then wait for or request feedback/merge

Contributors must agree to the license (see LICENCE.txt in the root folder).

Before committing check code formatting and spacing are consistent, and run a check (in PyCharm this is automatic, or can be run on all files in the main cobaya source using Code -> Inspect; this finds static errors, generally anything at Warning and above should be fixed, but strict PEP8 is not required).

To set Cobaya's line-length convention in PyCharm set Pycharm: File -> Settings -> Editor -> Code Style -> Hard wrap to 90. Ctrl+Alt+L will then auto-format code consistently, including wrap and spacing etc, though some line wrapping may need manual tweaking for nicer formatting.

Using PyCharm is the easiest way to keep things consistent. If you use another editor you can check most formatting and static errors on the command line using:

flake8 cobaya --select=E713,E704,E703,E714,E741,E10,E11,E20,E22,E23,E25,E27,E301,E302,E304,E9,F405,F406,F5,F6,F7,F8,W1,W2,W3,W6 --show-source

Release checklist

  • Make sure all tests pass in Travis (or the package won't be pushed to PyPI).
  • Make sure everything relevant has been added to the Changelog.
  • Delete old deprecation notices (>=2 versions before)
  • Bump version number in __init__.py and CHANGELOG.md (also date)
  • If archived version: - make __obsolete__ = True in __init__.py - Fix CAMB's version to latest release (right now, it installs master by default)
  • Update year of copyright in __init__.py.
  • Update year of copyright in LICENCE.txt.
  • Commit + tag with new version + git push + git push --tags
  • If needed, re-build the documentation.
  • If applicable, delete branches merged for this version.
  • Notify via the e-mail list.

Notes on some design choices

Generality and modularity

This code is being developed as a general-purpose sampler, with a Bayesian focus. The different aspects of the sampling, namely prior, likelihood and sampler, are kept as isolated as possible: the prior and sampler know which parameters are sampled and fixed, but not the likelihood, which does not need to know; the sampler does now know which likelihood understands which parameters, since it does not care (just cares about their respective priors, speeds, etc). This designs choices take some compromises, but are a fair price for making the code more easily extendable and maintainable: e.g. adding a new likelihood on the fly, or modify their parameters, needs no modification of the main code.

The cosmology only enters through particular likelihood and theory modules, and the main source does not contain significant cosmological code or experimental data, just wrappers, so that the general user does not need to download gigabytes of data.

Ideally, in the near future, the source will be split in two: a general sampler package on one side, and the cosmological modules on the other.

Dealing with parameters

Parameter roles

Parameters have different roles with respect to different parts of the code:

The :class:`~.parameterization.Parameterization` class (see diagram) takes care of interfacing between these two sets of roles, which, as it can be seen below, is sometimes not as simple as sampled + fixed = input, and derived = output.

Warning

Despite generating some ambiguity, we call output parameters sometimes also derived, when it is clear that we are in the likelihood context, not the sampler context.

How likelihoods and theory decide which input/output parameters go where

Once the :class:`.~parameterization.Parameterization` has decided which are the input and output parameters, the :class:`.~model.Model` needs to decide how to distribute them between the likelihood and theory components.

The simplest way to do that would be tagging each parameter with its corresponding likelihood(s) or theory, but this would make the input much more verbose and does not add much. Alternatively we could hard-code parameter routes for known parameters (e.g. for cosmological models), but hard-coding parameter names impose having to edit Cobaya's source if we want to modify a theory code or likelihood to add a new parameter, and we definitely want to avoid people having to edit Cobaya's source (maintainability, easier support, etc).

So, in order not to have tag parameters or hard-code their routes, the only option left is that each likelihood and theory can tell us which parameters it understands. There are a number of possible ways a likelihood or theory could do that:

  • If it is defined as a Python function (an external likelihood, in our terminology), we can use introspection to get the possible arguments. Introspection for output parameters is a little more complicated (see note below).
  • For internal likelihoods and theories (i.e. more complex classes that allow more flexibility and that have no function to inspect), we need either:
    • to keep a list of possible input/output parameters
    • to define a rule (e.g. a prefix) that allows us to pick the right ones from a larger set
  • Finally, if there is a likelihood or theory that cannot be asked and does not keep a list of parameters, that would not necessarily be a problem, but we would have to choose between passing it either all of the parameters, or just those that have not been claimed by anyone else (in this last case, there could obviously be only one likelihood or theory in the collection with this property).

Note

For callable (external) likelihood functions, output parameters cannot be simple keyword arguments, since in Python parameter values (float's) are immutable: they are passed by value, not by reference, so their value cannot be modified back. Thus, we interface them via a dictionary passed through a _derived keyword argument. Since dictionaries are mutable objects, when their contents are modified the modifications are permanent, which makes a natural way of dealing with derived parameters on the same ground as sampled parameters. At function definition, we assign this keyword argument a list of possible keys, which we can get, via introspection, as the list of output parameters understood by that likelihood.

We should also take into account the following:

  • Different likelihoods may share part of the same model, so they may have input parameters in common (but not output parameters; or if they do, we still only need to compute them once).
  • Some likelihoods may not take any input parameter at all, but simply get an observable through their interface with a theory code.
  • Some parameters may be both input and output, e.g. when only a subset of them can determine the value of the rest of them; e.g. a likelihood may depend on a and b, but we may want to expose a+b too, so that the user can choose any two of the three as input, and the other one as output.
  • External functions may have a variable number of input parameters, since some may be represented by keyword arguments with a default value, and would thus be optional.

To implement these behaviours, we have taken the following design choices:

  • Two parameters with the same name are considered by default to be the same parameter. Thus, when defining custom likelihoods or creating new interfaces for external likelihoods, use preferably non-trivial names, e.g. instead of A, use amplitude, or even better, amplitude_of_something. (The case of two likelihoods naming two different parameter the same is still an open problem: we could defined two parameters prefixed with the name of the likelihood, and have the :class:`model.Model` deal with those cases; or we could define some dynamical renaming.)
  • If a likelihood or theory (with method get_allow_agnostic() returning True) does not specify a parameter set/criterion and it is not the only element in the collection, we pass it only the parameters which have not been claimed by any other element.
  • Cosmology theory codes may understand a very large number of input/output parameters. These can be obtained by from the code by internal introspection or they will often be the "no knowledge" (agnostic) kind. On the other hand, they should not usually share parameters with the likelihoods: if the likelihoods do depend on any theoretical model parameter, they should request it via the same interface the theory-computed observables are, so that the parameterization of the theoretical model can be changed without changing the parameterization of the likelihoods (e.g. an H_0 likelihood may require the Hubble constant today, but if it where an input parameter of the likelihood, it would be more complicated to choose an alternative parameterization for the theoretical model e.g. some standard ruler plus some matter content).
  • Given the ambiguity between input and output roles for particular parameters, likelihood and theory classes that keep a list known parameters can do so in two ways:
    • The preferred one: a common list of all possible parameters in a params block in the defaults file. There, parameters would appear with their default role. This has the advantage that priors, labels, etc can be inherited at initialisation from these definitions (though the definitions in the user-provided input file would take precedence). If there is a conflict between the priors (or fixed value, or derived state) for the same parameter defined in different defaults files of likelihoods that share it, an error will be produced (unless the user settles the conflict by specifying the desired behaviour for said parameter in the input file).
    • Alternatively (and preferred when there is a conflict), they could keep two lists: one of input and one of output parameters.
    • If the parameters used depend on input options, or have to be obtained from internal introspection, the supported parameters must be returned programmatically from the get_can_support_params class method.
  • It may be that the likelihood does not depend on (i.e. has constraining power over) a particular parameter(s). In that case, we still throw an error if some input parameter has not been recognised by any likelihood, since parameter names may have been misspelled somewhere, and it is easier to define a mock likelihood to absorb the unused ones than maybe finding a warning about unused parameters (or use the unit likelihood described below).
  • Some times we are not interested in the likelihood, because we want to explore just the prior, or the distribution the prior induces on a derived parameter. In those cases, we would need a mock unit likelihood. This unit likelihood would automatically recognise all input parameters (except those absorbed by the theory, if a theory is needed to compute derived parameters).
  • For external likelihood functions, where we can get input and output parameters via introspection, we may not want to use all of the input ones, as stated above, since they may have a fixed default value as keyword arguments. This would be treated as a special case of having a list of input parameters.

Given these principles, we implement the following algorithm to resolve input/output parameter dependencies: (in the following, components include theory and likelihood codes)

  1. Start with a dictionary of input parameters as keys, and another one for output parameters. The values will be a list of the component that depend on each parameter.
  2. Iterate over components that have knowledge of their own parameters, either because they are callable, or because they have input/output parameters lists, a prefix, a mixed params list, or get_can_provide_params() or get_requirements(), in that order of priority. Add them to the lists in the initial parameters dictionaries if applicable.
  3. Deal with the case (check that it is only one) of a component with get_allow_agnostic() returning true, and assign it all unclaimed parameters.
  4. If the unit likelihood is present, assign it all input parameters (if not already used by component with get_allow_agnostic() ).
  5. Check that there are no unclaimed input/output parameters, and no output parameters with more than one claim.

This algorithm runs after initialize of the components is called, but before initialize_with_params.

After parameters have been assigned, we save the assignments in the updated (full) info using the unambiguous "input/output lists" option, for future use by e.g. post-processing: during post-processing, unused likelihoods are not initialised, in case they do not exist any more (e.g. an external function), but we still need to know on which parameters it depended.

Dynamical reparameterization layer (a bit old!)

As stated above, parameters are specified according to their roles for the sampler: as fixed, sampled and derived. On the other hand, the likelihood (and the theory code, if present) cares only about input and output arguments. In a trivial case, those would correspond respectively to fixed+sampled and derived parameters.

Actually, this needs not be the case in general, e.g. one may want to fix one or more likelihood arguments to a function of the value of a sampled parameter, or sample from some function or scaling of a likelihood argument, instead of from the likelihood argument directly. The reparameterization layer allow us to specify this non-trivial behaviour at run-time (i.e. in the input), instead of having to change the likelihood code to make it understand different parameterizations or impose certain conditions as fixed input arguments.

In general, we would distinguish between two different reparameterization blocks:

  • The in block: f(\text{fixed and sampled params})\,\Longrightarrow \text{input args}.
  • The out block: f(\text{output [and maybe input] args})\,\Longrightarrow \text{derived params}.

Note

In the out block, we can specify the derived parameters as a function of the output parameters and either the fixed+sampled parameters (pre-in block) or the input arguments (post-in block). We choose the post case, because it looks more consistent, since it does not mix likelihood arguments with sampler parameters.

Let us look first at the in case, in particular at its specification in the input. As an example, let us assume that we want to sample the log of a likelihood argument x.

In principle, we would have to specify in one block our statistical parameters, and, in a completely separate block, the input arguments as a series of functions of the fixed and sampled parameters. In our example:

params:
  logx:
    prior: ...  # whatever prior, over logx, not x!
    ref: ...    # whatever reference pdf, over logx, not x!

arguments:
  x: lambda logx: numpy.exp(logx)

This is a little redundant, specially if we want to store x also as a derived parameter: it would appear once in the params block, and again in the arguments block. Let us assume that in almost all cases we communicate trivially with the likelihood using parameter names that it understands, such that the default functions are identities and we only have to specify the non-trivial ones. In that case, it makes sense to specify those functions as substitutions, which in out example would look like:

params:
  logx:
    prior: ...  # whatever prior, over logx, not x!
    ref: ...    # whatever reference pdf, over logx, not x!
    subs:
      x: lambda logx: numpy.exp(logx)

If the correspondences are not one-to-one, because some number of statistical parameters specify a larger number of input arguments, we can create additional fixed parameters to account for the extra input arguments. E.g. if a statistical parameter y (not understood by the likelihood) defines two arguments (understood by the likelihood), u=2y and v=3y, we could do:

params:
  y:
    prior: ...  # whatever prior, over y
    subs:
      u: lambda y: 2*y
  v: lambda y: 3*y

or even better (clearer input), change the prior so that only arguments known by the likelihood are explicit:

params:
  u:
    prior: ...  # on u, *transformed* from prior of y
  v: lambda u: 3/2*u

Note

The arguments of the functions defining the understood arguments should be statistical parameters for now. At the point of writing this notes, we have not implemented multi-level dependencies.

Now, for the out reparameterization.

First, notice that if derived parameters which are given by a function were just specified by assigning them that function, they would look exactly like the fixed, function-valued parameters above, e.g. v in the last example. We need to distinguish them from input parameters. Notice that an assignment looks more like how a fixed parameter would be specified, so we will reserve that notation for those (also, derived parameters may contain other sub-fields, such as a range, which are incompatible with a pure assignment). Thus, we will specify function-valued derived parameters with the key derived, to which said function is assigned. E.g. if we want to sampling x and store x^2 along the way, we would input

params:
  x:
    prior: ...  # whatever prior for x
  x2:
    derived: lambda x: x**2
    min: ...  # optional

As in the in case, for now we avoid multilevel dependencies, by making derived parameters functions of input and output arguments only, not of other derived parameters.

Notice that if a non trivial reparameterization layer is present, we need to change the way we check at initialisation that the likelihoods understand the parameters specified in the input: now, the list of parameters to check will include the fixed and sampled parameters, but applying the substitutions given by the subs fields. Also, since derived parameters may depend on output arguments that are not explicitly requested (i.e. only appear as arguments of the function defining the derived parameters), one needs to check that the likelihood understands both the derived parameters which are not specified by a function, and the arguments of the functions specifying derived parameters, whenever those arguments are not input arguments.

Note

In the current implementation, if we want to store as a derived parameter a fixed parameter that is specified through a function, the only way to do it is to defined an additional derived parameter which is trivially equal to the fixed one. In the u,\,v example above, if we would want to store the value of v (fixed) we would create a copy of it, V:

params:
  u:
    prior: ...  # *transformed* from prior of y
  v: lambda u: 3/2*u
  V:
    derived: lambda v: v