History

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog and this project adheres to Semantic Versioning.

[Unreleased]

Summary

Fixed

Added

Changed

Deprecated

Removed

Attribution

[v0.7.5] - 2021-02-10

Fixed

  • Corner-case stability improvements for default background marginalisation.
  • If likelihood function is below llzero after evaluation, the parameter vector is included in the prior support as intended.
  • Typo in _precision function in xpsi/PostProcessing/__init__.py. (S.V.)
  • Math typo on the HotRegion page. (S.V.)
  • Explanatory text in the multiple-imaging tutorial. (T.S.)

Changed

  • A few image components appearing on the HotRegion page. (S.V.)
  • Bounds exception now prints the name of the offending parameter in Parameter. (S.V.)

Added

  • An extension module for calculating hot region local variables from global variables for hot region configurations under the umbrella of the PST-U model introduced in Riley et al. 2019 (ApJL, 887, L21).

Attribution

  • With thanks to Serena Vinciguerra (S.V.) and Tuomo Salmi (T.S.).

[v0.7.4] - 2021-01-26

Fixed

  • Missing packages in setup.py causing errors when importing xpsi.
  • A few typos in the documentation.

Added

  • A few images in the documentation.

Attribution

  • Serena Vinciguerra, Yves Kini, Devarshi Choudhury.

[v0.7.3] - 2020-11-12

Fixed

  • Phase-averaging issue that can sometimes occur due to numerical effects when comparing two numbers that should be the same but can differ by tiny degrees at machine precision level.
  • Some documentation typographic errors.

[v0.7.2] - 2020-11-04

Fixed

  • Error raised while running setup.py for linking rayXpanda with clang compiler.

Attribution

  • Serena Vinciguerra.

[v0.7.1] - 2020-10-01

Fixed

  • An AttributeError raised during runtime linking to the fallback rayXpanda implementation.

Attribution

  • With thanks to Devarshi Choudhury for bug testing.

[v0.7.0] - 2020-09-30

Summary

  • New plotting functionality.
  • Should be backwards compatible, but some small internal tweaks or default behaviour changes could result in small differences in plots that might not even be discernable.

Added

  • Option to specify only the number of phases per cycle when calling image(), instead of having to supply the phase set.
  • New plot type for animated photon specific intensity skymaps with their associated photon specific flux pulse-profiles and the photon specific flux spectrum that connects the signals at those energies. See the documentation of the image() method for options, details, and an example.
  • Example plots to the Photosphere documentation.
  • New helper methods write_image_data() and load_image_data() to write ray map data, photon specific intensity image data, and photon specific flux signal data to disk, and then read the data back into memory as attributes so that the data can be reused to accelerate calls to calculate images and generate static and animated plots.
  • Option to _plot_sky_maps(), add_zero_intensity_level, that applies a colormap such that zero intensity corresponds to the lowest colour. In this case a non-radiating part of the stellar surface, and the background sky, have well-defined colour. If lowest colour in the colormap is instead associated with the lowest finite intensity in the skymap panel, then the background sky for instance is assigned the same colour so that the least bright part of the image merges with the background sky colour. The latter choice resolves the variation in the intensity as a function of phase and sky direction better with colour, but the former might give more of an indication of the magnitude of the variation in intensity as a function of phase and sky direction relative to the background sky.

Changed

  • A phase set supplied to image() can have units of cycles, not radians as was previously the requirement, by setting the phase_in_cycles keyword argument to True if the supplied phase array as units of cycles.
  • The photon specific flux can be calculated with image() at far more energies than photon specific intensities are cached at, by using the cache_energy_indices keyword to supply and array of integers to index the energy array. This saves memory and means that imaging with an extension module can be executed once to generate both skymaps (which require cached intensities but only typically at a few representative energies) and the photon specific flux (which does not require cached intensities, but typically is computed for a much finer energy array).

Attribution

  • With thanks to Anna Bilous and Serena Vinciguerra for helpful suggestions about the new animated plot type.

[v0.6.3] - 2020-10-01

Fixed

  • An AttributeError raised during runtime linking to the fallback rayXpanda implementation.

Attribution

  • With thanks to Devarshi Choudhury for bug testing.

[v0.6.2] - 2020-09-28

Fixed

  • Bug in nested() when initialisation of nested sampler class tries to call set_default dictionary method instead of the correct setdefault method.
  • Import errors associated with the PostProcessing module.

Changed

  • The cached property of a Parameter instance can be set to None.
  • The ParameterSubspace initialiser is decorated to avoid verbose output by every MPI process.
  • The Prior uses the class attribute __draws_from_support__ to set the number of Monte Carlo draws from the joint prior support to require to set the MultiNest hypervolume expansion factor appropriately. The default value is 5, which means \(10^5\) draws from the joint prior support.
  • Checks if an instance of six.string_types in Metadata, e.g., to allow unicode strings in posterior ID labels.

[v0.6.1] - 2020-09-14

Fixed

  • Bug wherein multiple Signal instances passed to a Likelihood instance do not have references stored.
  • The tools synthesis functions adhering to the global phase interpolant switch, and updated tutorial accordingly.

Changed

Removed

  • An unused prototype extension module.

[v0.6.0] - 2020-09-05

Summary

  • Backwards compatible for most use cases, but possible corner cases.
  • Includes a non-critical, but important patch for animating intensity skymaps, and updates to the environment file for cloning.
  • The new feature is support for higher-order images when invoking an integrator that discretises the surface (with a regular mesh). Secondary images can be very important, whilst tertiary images less so. Quaternary, quinary, and possibly senary images can sometimes be detected and included too, with accuracy that decreases with order. Fortunately, the contribution to the photon specific flux generally decays rapidly with image order beyond the secondary or tertiary images. The computational cost scales almost linearly with order if an appreciable fraction of every iso-latitudinal ring on the surface is multiply-imaged at each order. Note that multiple-imaging manifests entirely naturally when an image-plane is discretised in such away that the regular mesh resolves the stellar limb sufficiently well, where higher-order images get insanely squeezed.

Fixed

  • The memory consumption problem of the animator method in Photosphere. Now animation should generally require an entirely tracable amount of memory.

Added

  • Multiple-imaging support including an option to specify the maximum image order to iterate up to, with automatic truncation when no image at a given order is detected. If no limit is specified (the default), then images are included as far as they can be detected given the numerical resolution settings, which is typically between quaternary and senary images.
  • A multiple-imaging tutorial.
  • A global switch for changing phase and energy interpolants without recompilation of extensions. To change interpolants, you can use top-level functions xpsi.set_phase_interpolant() and xpsi.set_energy_interpolant(). Generally computations are more sensitive to the phase interpolants, of which the options from GSL are: Steffen spline (pre-v0.6 choice), Akima periodic spline, and cubic periodic spline. The default choice is now an Akima periodic spline in an attempt to improve interpolation accuracy of the interpolant at function maxima, where the accuracy is generally most important in the context of likelihood evaluations. Note that in some corner cases, the signal from a hot region is negative in specific flux because there is a correction computed to yield the intended signal from Elsewhere when it is partially masked by hot regions. In this case, when using phase interpolant tools from the tools and likelihood modules it is necessary to use a allow_negative option when calling the tools to specify that a negative interpolant is permitted.
  • Automatic linking of the package rayXpanda for calculation of the inverse of the deflection integral, and it’s derivative via a high-order symbolic expansion, for a subset of primary images. The purpose is to mainly as an orthogonal validation of a subset of integrals executed via numerical quadrature and inversion via spline interpolation. The other reason is because to support multiple-imaging with the surface-discretisation integrators this aforementioned interpolation had to change due to non-injectivity of functions when interpolating with respect to the cosine of the deflection angle. However, to calculate the convergence derivative sufficiently accurately, interpolating with respect to the cosine of the deflection seems necessary. Therefore rayXpanda can be linked in, if it is available, for low deflection angles instead of avoid having to allocate additional memory and construct splines specifically for low-deflection primary images. Simple testing suggests there are no valuable speed gains, however, possibly because the high-order expansion and simultaneous evaluation of the polynomial and it’s derivate with a nested Horner scheme itself requires a substantial number of floating point operations.
  • A helper method merge() that merges a set of parameters, or a parameter subspace, or a set of subspaces, into a subspace that has already been instantiated.

Changed

  • Updated the Conda environment.yml file for replication of the development environment. The basic_environment.yml file was also updated in an earlier release in an additional necessary package, wrapt.

Deprecated

  • The repeat, repeat_delay, and ffmpeg_path keyword arguments for the animator method in Photosphere. These were ultimately not effective. To repeat the animation intrinsically, set the number of cycles, and extrinsically, this can be looped when embedded in another environment.

[v0.5.4] - 2020-09-01

Fixed

  • Bug due to local variable NameError when setting instrument channel energy edges.
  • Bug that prevented a hot region phase parameter from being a fixed or derived variable.

Attribution

  • With thanks to Devarshi Choudhury.

[v0.5.3] - 2020-08-14

Summary

  • Improvement patches. Deliberately backwards incompatible for safety in memory allocation.

Fixed

  • Add try-except block to global_to_local_file property so that explicit setting of None by user is not required if file I/O is not needed in the extension module. Actually, None could not be set for the property anyway due to type checking.
  • Bug when declaring that sky maps should be animated and memory freed beforehand.

Added

  • The surface to image-plane ray map is cached in Python process memory so it can be efficiently reused for same spacetime configuration and ray map resolution settings. Explicit support for writing the ray map to disk and loading it is not included, but this should be entirely possible to achieve manually. Backwards compatible except for corner cases, such as not using keyword arguments when calling image(), or if resolution settings changed between calls to the imager but a ray map otherwise exists in Python process memory and the spacetime configuration has not been changed.
  • A secret keyword argument to image(), _OVERRIDE_MEM_LIM, which can be used to change an internal hard limit on the intensity cache size. This setting is for safety and designed so that higher memory consumption is deliberate or if something goes awry, it is deemed the responsibilty of the user to have read method docstring carefully. The tutorials will not use this secret keyword, so if the user tries to run them and encounters an exception, they will need to investigate the docstring and either adapt the resolution to their system or take the responsibility of setting the cache size limit for their system to accomodate the resolution settings in the tutorial.
  • Optional argument to image(), single_precision_intensities, which flags whether or not to cache the intensities in single precision do halve intensity cache memory requirements. The default is to cache in single precision.
  • Verbosity to image() because execution can take many minutes depending on settings chosen. The verbosity can be deactivated via a keyword argument (see the method docstring).

Changed

  • The usage of the image() argument cache_intensities. Instead of simply activating intensity caching with boolean, the user must specify a cache size limit that is adhered to. If the required cache size given the resolution settings is larger than the limit, imaging does not proceed. If the cache size limit is zero or equivalent, then imaging safely proceeds without caching the intensities.
  • Intensities are by default cached in single precision to reduce cache memory requirements.

[v0.5.2] - 2020-08-12

Summary

  • Python API: small backwards compatible patches to add useful features.
  • C API: small backwards incompatible patch to support Python API patch.

Added

  • Support for hyperparameters (i.e., parameters of the prior distribution), by making Prior inherit from ParameterSubspace. Custom hyperparameters can then be defined in a subclass initiliser, or otherwise. The hyperparameters are merged into the Likelihood parameter subspace as mostly normal parameters (with small caveat in the form of property is_hyperparameter) and can have their own prior (the hyperprior) implemented in a Prior subclass along with the other free parameters in the model. A tutorial will be delivered in due course. These modifications are backwards compatible.
  • Simple support for transforming from global to local variables (for image- plane calculations) with the help of a file on disk, whose path can be specified dynamically in Python and relayed to the relevant extension where a custom model implemention can do I/O with the file. This is useful if one has a set of files containing precomputed data, but understandably does not want to do filesystem acrobatics or recompile an extension every time the file path changes. Setting the file path dynamically in this way is akin to changing the value of some discrete variable in the mapping between global and local variables. With thanks to Anna Bilous for the suggestion. A tutorial will be delivered when possible.
  • Added channel_edges property, and updated tutorials to reflect this new concrete implementation.

Changed

  • The init_local_variables function signature in the header xpsi/surface_radiation_field/local_variables.pxd, and in the corresponding xpsi/surface_radiation_field/archive/local_variables extensions. You would have to modify a custom extension module manually to match the function signature declared in the header.

Fixed

  • Removed remnant manual Sphinx method signatures; the decorator now preserves the method signature so automated Sphinx doc works on those decorated methods.
  • Updated package docstring to reflect name change.
  • Uses of xpsi.Data.channel_range property to adhere to future deprecation.

[v0.5.1] - 2020-08-07

Fixed

  • Bug when plotting intensity sky maps because a line was inadvertently removed.
  • Some mutable defaults in xpsi.Elsewhere and xpsi.Everywhere.
  • Conditional statement in xpsi.Photosphere.embed().

Added

  • Capability to add custom parameters when instantiating xpsi.Photosphere, which is useful for calling image plane extensions whilst passing global variables, without having to instantiate surface-discretisation classes and without having to handle global variable values at compile time or from disk for runtime access.

[v0.5.0] - 2020-08-06

Summary

  • The major change is an update and refactoring of the post-processing module to work again with past API changes. (The module was not being kept up to date with previous releases listed below because it wasn’t to our knowledge being used by anyone yet, and thus we focussed on other features.) The module has been refactored to be more modular, flexible, and extensible. For instance, posterior signal-plot classes can be added by the user and complex plotting routines can thus be developed, as demonstrated in the concrete classes such as xpsi.PostProcessing.PulsePlot. The plot classes have been used to reproduce (with improved functionality and performance) the relevant signal plots from Riley et al. 2019 (ApJL, 887, L21), as demonstrated in the post-processing tutorial notebook and embedded in the class docstrings for reference.
  • Development of online documentation pages, including project organisation pages and a Code of Conduct (please read), and development of docstrings. Note that some snippets of documentation look forward to v1.0 (e.g., release of technical notes in the repo itself).

Fixed

  • The xpsi.Data docstring explanations have been improved for clarity, mainly regarding the instrument channel definitions. The explanation is of how the information contained in a xpsi.Data instance pertains to the loaded instrument response (sub)matrix.
  • The xpsi.Instrument docstrings have also been improved for clarity, explaining the relationship to xpsi.Data in more detail.
  • Update extension module for background marginalisation to take distinct phase sets associated with hot regions.
  • The constructor xpsi.Spacetime.fixed_spin() inclination upper bound is \(\pi/2\) radians to eliminate degeneracy due to equatorial-reflection symmetry in the default prior on source-receiver geometric configuration.
  • Tweak caching (memoization) so that cache and current vectors are equal at the end of likelihood evaluation routine.
  • Generally clean up naming and docstrings for extension modules. Add return types.
  • Bug was fixed for transforming posterior sample sets and prior samples when parameter orders different in sample files and a prior object due to API updates. Whether this solution is to be long-term is to be decided; more generally need to figure out how to elegantly handle derived parameters that are not needed for likelihood evaluation (those derived parameters are instances of xpsi.Parameter) but are of interest for post-processing.
  • Handle param_plot_lims=None correctly in xpsi.PostProcessing.CornerPlotter.
  • Checked for unintended mutable defaults package-wide, and fixed as appropriate.
  • Fix bugs in CustomPrior class (Example script and modules; these example modules were not run at the time of translation between past API versions, so only found bugs when making post-processing tutorial for this release).
  • The formatting of annotated credible intervals in xpsi.PostProcessing.CornerPlotter has been improved by inferring the largest number of decimal places needed for two non-zero decimal digits, and then formatting the median and quantile differences to this shared decimal precision above the on-diagonal panels. If the numbers cannot be well- represented by this scheme, the user could try a unit transformation.
  • Tried to tweak automated margins for intensity sky map multi-panel plots, so as not to sometimes partially cut an axis label.
  • Bug that prevented animation of sky map frames written to disk because the frames were not cached in memory by reimaging.

Added

  • The xpsi.Data is now concrete in implementation, such that in common usage patterns, it does not need to be subclassed.
  • A constructor to xpsi.Data to load a phase-folded event list and phase-bin the events in a subset of selected channels.
  • A xpsi.Data.channels() property that holds the instrument channels to be checked by a xpsi.Signal instance against those declared for the loaded instrument response (sub)matrix. This property as also required by the post-processing module (namely, xpsi.PostProcessing.ResidualPlot and the other xpsi._signalplot.SignalPlot subclasses).
  • A xpsi.Instrument.channels() property that holds the instrument channels to be checked by a xpsi.Signal instance against those declared for the event data matrix.
  • Support for multiple instruments operating on the same incident signal due to assumed effective time-invariance of the signal generated during one rotational cycle of the surface radiation field.
  • Module xpsi.surface_radiation_field to call atmosphere extensions directly (without the calls being embedded in integration algorithms), for checking implementation of complicated atmospheres such as those requiring interpolation with respect to a numerical lookup table.
  • Support for the extension module for calculating the local surface radiation field variables to read in numerical model data. An example extension module designed to execute nearest-neighbour lookup amonst an general unstructured array of points of the openness of magnetic field lines has been developed.
  • Add simple energy annotation option to photon specific intensity sky-map panels.
  • State the energy units (keV) that the xpsi.Instrument must comply with when energy interval bounds are specified.
  • State the units of variables such as energy and specific intensity in the surface radiation field extension module. These requirements may be found in function body comments.
  • Explain in xpsi.PostProcessing.CornerPlotter docstring the order in which posteriors are plotted given the input order.
  • Post-processing switches to overwrite transformed-sample files and combined-run files on disk.
  • Workaround to handle the case where due to API changes, the relationship between sample parameter vectors on disk and the parameter vector in the current API are related not just by reordering, but transformations. This is demonstrated in the post-processing tutorial instead of transforming the original sample files on disk in place, the transformed files written to disk contain both the transformed vector (same number of elements) to match the parameters defined under the current API (the order of the vector can be different between the xpsi.ParameterSubspace underlying with a xpsi.Likelihood instance and the files on disk containing the transformed samples), and the additional derived parameters.
  • Attempt to free up memory when xpsi.Photosphere.images() is no longer needed, but memory-intensive operations need to be performed.
  • Attempt to free memory properly after animating a sky-map phase sequence.

Changed

  • Change (Earth) inclination parameter \(i\) to \(\cos(i)\) so that the default prior density function is isotropic.
  • The object formerly named xpsi.Pulse has had its name changed to xpsi.Signal, and across the package, names that were pulse are apart from potential corner cases or documentation instances of the word, are now signal, because when support joint likelihood functions over multiple instruments, some data sets are phase averaged. Moreover, signal is arguably clearer in meaning than pulse, once it has been established that the signals the package focuses on are pulsed but depending on the instrument, the data we confront the model with has some degree of phase (timing) resolution that might be insufficient for phase-resolved observations.
  • The xpsi.Data definition of the last channel has changed to be the index of the last row in the loaded instrument response (sub)matrix, instead of being the index of the last row plus one; this means that the value exposed via a property is last+1.
  • For numerical atmospheres of same number of grid dimensions, improved extension surface_radiation_field/archive/{hot,elsewhere}/numerical.pyx module to infer grid size for memory allocation and interpolation searches (implemented automatic inference of grid size, but hard-coded four-dimensional cubic polynomial interpolation persistent). Different those atmospheres can be loaded simply via a Python subclass without the relevant extension module being recompiled.
  • The xpsi.Photosphere class sometimes does no surface discretisation, so allow no hot regions, elsewhere, or everywhere objects; then image-plane discretisation can be accessed without dummy object creation.
  • Tweak xpsi.SpectrumPlot settings to print a warning statement that spectrum plot works best with logarithmic spacing, and the user has to shadow class attribute with logspace_y=False.
  • Do not print xpsi.HotRegion instance parameter properties upon creation if fixed at boundary value so that the region is fully described by fewer parameters.
  • Merged energy integration extension modules into one.
  • Made phase shift parameters (strictly) unbounded; remember however that for a sensible prior, bound the phase shifts on a unit interval, and thus it is required that phase bounds are specified and finite.
  • In extensions, modified phase shifting so that a shift permitted by unbounded phase parameter does not require many iterations to decrement or increment to unit interval (achieved simply with floor operation).

Deprecated

  • The xpsi.Data.channel_range() property has been renamed to xpsi.Data.index_range() so as to avoid confusion between these numbers and the true instrument channels. The old property will be removed for release v1.0.

Removed

  • The ensemble MCMC sample backend for post-processing because we do not expect it to be useful in the immediate future, but requires some non-trivial development work to meld properly with the current post-processing module which is focussed on nested sampling. This functionality will be reintroduced in a future release (refer to Future). The ensemble sampler can still be run, however, and the native backend for accessing sample information on disk is demonstrated in a tutorial notebook. However, the runs cannot be processed for posterior integrals and visualisation using the same tools as available for nested sampling runs.

Attribution

  • With thanks to Sebastien Guillot (testing and feedback), Devarshi Choudhury (testing and feedback), Sam Geen & Bob de Witte (Windows installation advice), and Anna L. Watts (documentation patches and feedback).

[v0.4.1] - 2020-06-03

Fixed

  • Function signatures to match header declarations in atmosphere extensions: xpsi/surface_radiation_field/archive/elsewhere/numerical.pyx to match xpsi/surface_radiation_field/elsewhere_radiation_field.pxd. With thanks to Sebastien Guillot.

[v0.4.0] - 2020-02-14

Summary

  • Mainly new features.
  • Backwards compatible (apart from possible corner cases).

Fixed

  • Removed a spurious geometric factor in the integrator that discretises the surface with a static mesh. This integrator was called by the Elsewhere class. The error when this factor is included is O(1%) at 600 Hz for soft emission from the entire stellar disk, and then scales with spin and energy beyond this. To reproduce the bug, find the commented out / superlum in file xpsi/cellmesh/integrator_for_time_invariance.pyx (line 251) and uncomment it. Then reinstall the package. When this factor is included, the mesh itself is moving in the context of the images subtended by its constituent elements on our sky. We want the mesh to be static so that this integrator can be used for faster calculation of time-invariant signals.
  • Bug in which the prior density factor is incorporated twice if a Likelihood instance held a reference to a Prior object and these are merged into a Posterior object which is fed to the ensemble sampler. If the prior density was flat, this bug will have had no effect on posterior distributions.

Added

  • New features are the simulation of signals from more general surface radiation fields that globally span the stellar surface. This can be done with several types of integrator.
  • The new image-plane discretisation integrator supports imaging of a star, and Python functionality has been added to automate plotting and animation of intensity sky maps.
  • A new tutorial to the documentation to demonstrate these new features and an internal cross-check of distinct integration algorithms.
  • A visual introduction to the documentation pages with some animated sky maps.

[v0.3.6] - 2020-01-24

Fixed

  • Some code snippets in documentation examples of prior implementation with the latest API minor version (v0.3).

Changed

  • Modify the HotRegions class to function with two or more hot region objects.

[v0.3.5] - 2020-01-22

Summary

  • Docstring edits and backwards compatible changes to several class initialisation arguments.

Attribution

  • Based mostly on discussion with and feedback from Devarshi Choudhury.

Fixed

  • Some docs formatting problems.
  • Some corrections to example scripts/modules updated in v0.3.4 to use current API.

Changed

  • The photospheric mode frequency parameter is not converted to an angular frequency until it is used, so the cached value matches the docstring description.

Deprecated

  • The is_secondary argument of the HotRegion class. Use is_antiphased instead to ensure future compatibility.
  • The store argument of the Pulse class. Use cache instead to ensure future compatibility.

[v0.3.4] - 2020-01-20

Summary

  • A few patches including backwards compatible improvements.
  • Various docstring/comment/doc edits.
  • Update docs example model to use v0.3.4 API.

Fixed

  • Ensure consistency between input parameter bounds and values by always requiring dictionaries. Fix applies to Elsewhere and Photosphere. Courtesy Sebastien Guillot.
  • Gravitational mass doc typo fix.

Changed

  • Add input argument checks to Likelihood.check method.
  • Add default hypercube=None to Prior.inverse_sample_and_transform method.
  • If derived parameters found in subspace, assume an update is needed because cache mechanism not in place. (WIP.)

[v0.3.3] - 2020-01-20

Fixed

  • At several places in the Likelihood class, calls were place to self, forgetting that Likelihood overwrites ParameterSubspace.__call__. Now calls are super(Likelihood, self).__call__() to obtain the current parameter vector.

[v0.3.2] - 2020-01-16

Summary

  • Bug fixes. Backwards compatible.
  • When initializing the ensemble-MCMC chains using an nd-ball, the inclusion in the prior support was checked by passing a vector to Prior.__call__ but that code assumed that the parameter vector had already been assigned and can be accessed through the ParameterSubspace. As a result either an exception would be thrown (if parameter objects have no value set) or the support condition would be evaluated for some preset vector that does not change has we iterate through chains.
  • The Likelihood.check method now has a fallback implementation given that the NumPy allclose function in v1.17 does not support Python 2.7.

Attribution

  • Based on testing by Sebastien Guillot.

Fixed

  • The EnsembleSampler so that it does not rely on the CustomPrior.__call__ implementation to handle a vector argument. Chains should now be in prior support from the start and never leave.
  • The Likelihood.check method so that a call to a Likelihood instance updates the parameters with a vector if the physical points are passed for value checking.
  • The Likelihood.check method error error handling and if/else branching has been fixed.
  • Some typographic errors in docs.

Changed

  • The way EnsembleSampler accesses the prior object.

[v0.3.1] - 2019-12-12

Fixed

  • Some docstring and Sphinx-related formatting.

[v0.3.0] - 2019-12-10

Summary

  • Not backwards compatible.
  • The main feature is a more sophisticated backend for handling parameters, parameter subspaces, and the object hierarchy that forms the modelling language. Notably, the parameter objects can be accessed everywhere more readily, with dictionary-like functionality that alleviates the problem of remembering the imposed order of parameters in a vector. Resultantly, there is much more freedom when a user constructs a model and interfaces it with sampling software.
  • Model parameters can either be free, fixed/frozen at some scalar value, or derived deterministically from other model parameters.
  • The docs and tutorials have also been updated to reflect these developments.

Attribution

  • Feedback and ideas for the above development were discussed at an X-PSI workshop in Amsterdam, November 25-29 2019: Sebastien Guillot, Emma van der Wateren, Devarshi Choudhury, Pushpita Das, Anna Bilous, and Anna Watts.

Added

  • A new class xpsi.Parameter of which every model parameter is an instance.

Changed

  • The xpsi.ParameterSubspace class, which has far more sophisticated behaviours as a parameter container. The class, upon initialisation with arguments, also merges parameters and subspaces into a higher-dimensional (sub)space. Most other classes in the modelling language inherit from the xpsi.ParameterSubspace class.
  • The xpsi.TwoHotRegions class is now dedicated to representing antipodally reflection-symmetric configurations only to simplify the choice of which class to use between xpsi.HotRegions and xpsi.TwoHotRegions. However, antipodally reflection-symmetric models can also be constructed using just xpsi.HotRegions because of the new derived parameter support. The may be a minor speed difference: xpsi.TwoHotRegions should be very slightly faster, but it might be imperceptible. Future warning: in the future xpsi.TwoHotRegions might removed altogther for simplication.
  • The xpsi.Photosphere class can be instantiated to encapsulate only a reference to an xpsi.Elsewhere instance, and no xpsi.HotRegion instances. An xpsi.Elsewhere instance can by definition only generate a phase-invariant signal. However, further development is needed to handle this phase-invariant signal efficiently for likelihood functionality, given that operations with respect to phase are not required. Instead likelihood functions would be defined only with respect to energy.

Removed

  • The xpsi.ParameterSpace module. The global model parameter space is also simply an intance of the xpsi.ParameterSubspace class.