Skip to content

Latest commit

 

History

History
349 lines (251 loc) · 21.7 KB

CHANGELOG.md

File metadata and controls

349 lines (251 loc) · 21.7 KB
Raw

Changelog

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

Added

  • Add dataframe_to_live_points function to nessai.livepoint for converting from a pandas.DataFrame to live points.
  • Add fallback_reparameterisation to FlowProposal. This allows the user to specify which reparameterisation to use for parameters that are not included in the reparameterisations dictionary. Default behaviour remains unchanged (defaults to no reparameterisation).
  • Add rolling_mean to nessai.utils.stats.
  • Add nessai.flows.utils.create_linear_transform as a common function for creating linear transforms in the flows.
  • Add nessai.flows.transforms.LULinear to address a bug in nflows that has not been patched and prevents the use of CUDA with LULinear.
  • Add calibration_example.py to the gravitational wave examples.
  • Add defaults keyword argument to nessai.reparameterisations.get_reparameterisation for overriding the dictionary of default reparameterisations.
  • Add explicit tests for nessai.flowsampler
  • Add more tests for nessai.reparameterisations
  • Add more tests for nessai.gw
  • Add support for vectorised likelihoods and automatically detect if the likelihood is vectorised.
  • Add support for passing a user-defined pool instead of using n_pool.

Changed

  • NestedSampler.plot_state now includes the log-prior volume in one of the subplots and the rolling mean of the gradient (|dlogL/dLogX|) is plotted instead of the gradient directly.
  • The figure produced by NestedSampler.plot_state now includes a legend for the different vertical lines that can appear in the subplots.
  • RealNVP and NeuralSplineFlow now use nessai.flows.utils.create_linear_transform.
  • The figure produced by NestedSampler.plot_state now includes a legend for the different vertical lines that can appear in the subplots.
  • Updated all of the examples to reflect the new defaults.
  • Rework nessai.gw.reparameterisations.get_gw_reparameterisation to use get_reparameterisation with the defaults keyword argument.
  • Switch to os.path.join for joining paths in nessai.flowsampler.
  • Context is now passed to the transform in nessai.flows.base.NFlow enabling the use of flows with conditional transforms.
  • Add context_features to RealNVP and NeuralSplineFlows
  • Rework MaskedAutoregressiveFlow to add context_features
  • Rework how likelihood parallelisation is handled. The model now contains the pool instead of the sampler and proposals.
  • Update parallelisation_example.py to show use of n_pool and pool for parallelisation.
  • Simplify how the normalising flow is reset in FlowModel and NestedSampler.

Fixed

  • Fixed a bug in RescaleToBounds when using pre_rescaling without boundary inversion.

Removed

  • Removed legacy_gw_example.py
  • Removed keep_samples from FlowProposal.

0.4.0 - 2021-11-23

Added

  • Add a constant volume mode to FlowProposal. In this mode the radius of the latent contour is fixed to the q'th quantile, which by default is 0.95. (#125)
  • Add a check for resume_file when resume=True. (#126)
  • Change default logging level to WARNING. (#126)
  • Add angle-cosine reparameterisation. (#126)
  • Added an explicit check for one-dimensional models that raises a custom exception OneDimensionalModelError. (#121)
  • RealNVP and NeuralSplineFlow now raise an error if features<=1. (#121)
  • Add option in nessai.reparameterisations.Angle to set scale=None, the scale is then set as 2 * pi / angle_prior_range. (#127)
  • Add 'periodic' reparameterisation that uses scale=None in nessai.reparameterisations.Angle. (#127)
  • Add the use_default_reparameterisations option to FlowProposal to allow the use of the default reparameterisations in GWFlowProposal without specifying any reparameterisations. (#129)
  • Add chi_1, chi_2 and time_jitter to known parameters in GWFlowProposal with corresponding defaults. (#130)

Changed

  • Reparameterisation angle-sine is now an alias for RescaledToBounds instead of Angle with specific keyword arguments. (#126)
  • maximum_uninformed=None now defaults to 2 times nlive instead of np.inf. (#126)
  • nlive=2000 by default. (#126)
  • Default batch_size is now 1000. (#126)
  • Default n_neurons is now 2 times the dimensions of the normalising flow. (#126)
  • Default mode for FlowProposal is constant_volume_mode=True. (#126)
  • Proposal plots are now disabled by default. (#126)
  • cooldown now defaults to 200 to reflect the change in nlive. (#126)
  • Default optimiser is now adamw. (#126)
  • Rework AugmentedFlowProposal to work with the new defaults. (#126)
  • Model.names and Model.bounds are now properties by default and their setters include checks to verify the values provided are valid and raise errors if not. (#121)
  • Logger now has propagation enabled by default. (#128)
  • FlowProposal.configure_reparameterisations can now handle an input of None. In this case only the default reparameterisations will be added. (#129)
  • Changed default reparameterisation for gravitational-wave parameters a_1 and a_2 to 'default'. (#130)

Fixed

  • Fixed a bug where the parameters list passed to Reparameterisation (or its child classes) wasn't being copied and changes made within the reparameterisation would change the original list. (#127)

Deprecated

  • keep_samples in FlowProposal will be removed in the next minor release.

0.3.3 - 2021-11-04

Fixed

  • Fixed a bug in nessai.livepoint.dict_to_live_points when passing a dictionary where the entries contained floats instead of objects with a length raised an error. (#119)

0.3.2 - 2021-10-12

Added

  • Added more checks to the init method for nessai.reparameterisations.AnglePair to catch invalid combinations of priors and/or angle conventions. Now supports RA or azimuth defined on [-pi, pi] in addition to [0, 2pi]. (#114)
  • Add a check in nessai.flowmodel.update_config for 'noise_scale', a ValueError is now raised if noise_scale is not a float or 'adaptive'. (#115)
  • Add codespell to the pre-commit checks. (#116)

Changed

  • The dtype for tensors passed to the flow is now set using torch.get_default_dtype() rather than always using float32. (#108)
  • Incorrect values for mask in nessai.flows.realnvp.RealNVP now raise ValueError and improved the error messages returned by all the exceptions in the class. (#109)
  • Change scale of y-axis of the log-prior volume vs. log-likelihood plot from symlog to the default linear axis. (#110)
  • nessai.plot.plot_trace now includes additional parameters such as logL and logP default, previously the last two parameters (assumed to be logL and logP were always excluded). (#111)

Fixed

  • Fixed an issue where nessai.reparameterisations.AnglePair would silently break when the prior range for RA or azimuth was set to a range that wasn't [0, 2pi]. It now correctly handles both [0, 2pi] and [-pi, pi] and raises an error for any other ranges. (#114)
  • Fixed various spelling mistakes throughtout the source code and documentation. (#116)

0.3.1 Minor improvements and bug fixes - 2021-08-23

This release has a few minor improvements and bug fixes. It also explicitly adds support for python 3.9, which worked previously but was not tested.

Added

  • Add in_bounds, parameter_in_bounds and sample_parameter methods to nessai.model.Model. (#90)
  • Implemented the option to specify the cosmology in nessai.gw.utils.ComovingDistanceConverter using astropy. Previously changing the value had no effect of the transformation. (#91)
  • Improve test coverage for nessai.proposal.base.Proposal (#92)
  • Add 'logit' to the default reparameterisations (#98)
  • Add example using the Rosenbrock likelihood in two dimensions (#99)
  • Add a colours argument to nessai.plot.plot_1d_comparison (#102)
  • Explicitly support Python 3.9 (Added Python 3.9 to unit tests) (#103)

Changed

  • nessai.gw.utils.DistanceConverter now inherits from abc.ABC and to_uniform_parameter and from_uniform_parameter are both abstract methods. (#91)
  • nessai.proposal.base.Proposal now inherits from abc.ABC and draw is an abstract method. (#92)
  • nessai.proposal.rejection.RejectionProposal now inherits from nessai.proposal.analytic.AnalyticProposal. Functionality is the same but the code will be easier to maintain since this removes several methods that were identical. (#93)
  • noise_scale='adaptive' option in FlowModel now correctly uses a standard deviation of 0.2 times the mean nearest neighbour separation as described in Moss 2019. Note that this feature is disabled by default, so this does not change the default behaviour. (#95)
  • Refactor nessai.utils into a submodule. (#96)
  • Change behaviour of determine_rescaled_bounds so that rescale_bounds is ignored when inversion=True. This matches the behaviour in RescaledToBounds where when boundary inversion is enabled, values are rescaled to [0, 1] and then if no inversion if applied, changed to [-1, 1]. (#96)
  • Tweaked detect_edges so that both is returned in cases where the lower and upper regions contain zero probability. (#96)
  • NestedSampler no longer checks capitalisation of flow_class when determining which proposal class to use. E.g. 'FlowProposal' and 'flowproposal' are now both valid values. (#100)
  • NestedSampler.configure_flow_proposal now raises ValueError instead of RuntimeError if flow_class is an invalid string. (#100)
  • Raise a ValueError if nessai.plot.plot_1d_comparison is called with a labels list and the length does not match the number of sets of live points being compared. (#102)
  • nessai.flow.base.BaseFlow now also inherits from abc.ABC and methods that should be defined by the user are abstract methods. (#104)
  • Changed default to fuzz=1e-12 in nessai.utils.rescaling.logit and nessai.utils.rescaling.sigmoid and improved stability. (#105)

Fixed

  • Fixed a typo in nessai.gw.utils.NullDistanceConverter.from_uniform_parameter that broke the method. (#91)
  • Fixed a bug in nessai.reparameterisations.RescaleToBounds when using offset=True and pre_rescaling where the prime prior bounds were incorrectly set. (#97)
  • Fixed a bug that prevented disabling periodic checkpointing. (#101)
  • Fixed a bug when calling nessai.plot.plot_1d_comparison with live points that contain a field with only infinite values. (#102)
  • Fixed the log Jacobian determinant for nessai.utils.rescaling.logit and nessai.utils.rescaling.sigmoid which previously did not include the Jacobian for the fuzz when it was used. (#105)

0.3.0 Testing, testing and more testing - 2021-07-05

This release contains a large number of changes related to bugs and issues that were discovered when writing more tests for nessai.

It also adds a number of feature and examples.

Note: Runs produced with previous releases are incompatible with this release and cannot be resumed with out manual intervention.

Added

  • Added code to catch errors when calling plot_live_points when gwpy is installed.
  • Added tests for _NSIntegralState.
  • Add code coverage upload
  • Added an example of using unbounded priors, bilby_unbounded_priors.py
  • Added Rescale reparameterisation that just rescales by a constant and does not require prior bounds. Also add tests for this reparameterisation.
  • Added more GW examples.
  • Added tests for AugmentedFlowProposal.
  • Added an example using AugmentedFlowProposal.
  • Added eggbox example.
  • Added an error if calling FlowProposal.rejection_sampling with FlowProposal.truncate=True but worst_q=None.
  • Add option to train using dataloaders or directly with tensors. This is faster when using CUDA.
  • Add options to train with different optimisers: Adam, AdamW, SGD
  • Add tests for NestedSampler
  • Explicitly check prior bounds when using reparameterisations. This catches cases where infinite bounds are used and break some reparameterisations. (#82)
  • Add error when calling FlowProposal.populate without initialising the proposal.
  • Add NestedSampler.plot_insertion_indices to allow for easier plotting of insertion indices.
  • Add filename keyword argument to NestedSampler.plot_trace.
  • Added batch_norm_within_layers to NeuralSplineFlow

Changed

  • Plotting logX vs logL now returns the figure is filename=None
  • NestedSampler.plot_state now has the keyword argument filename and the figure is only saved if it is specified.
  • Changed name from _NSintegralState to _NSIntegralState.
  • nessai.model.Model now inherits from abc.ABC and log_prior and log_likelihood are now abstractmethods. This prevents the class from being used without redefining those methods.
  • Updated AumgentedFlowProposal to work with current version of FlowProposal
  • Fix random seed unit tests.
  • Improved FlowProposal.reset so that all attributes that are changed by calling draw are reset.
  • Move _NSIntegralState and some functions from posterior.py to evidence.py
  • NestedSampler.check_flow_model_reset will now NOT reset the flow it has never been trained (i.e proposal.training_count==0)
  • Moved all legacy gw functions to nessai/gw/legacy.py and removed them from the coverage report.
  • Minor improvements to NestedSampler
  • Better handling on NaNs in NestedSampler.populate_live_points
  • Minor improvements to plotting in FlowProposal and moved plotting to separate methods in FlowProposal.
  • Switch to using os.path.join when joins paths.
  • Improved FlowProposal.reset
  • Renamed FlexibleRealNVP to RealNVP, shouldn't affect most uses since the default way to specify a flow is via strings in configure_model.
  • Renamed nessai.flows.utils.setup_model to configure_model.
  • Renamed nessai.flows.utils.CustomMLP to MLP
  • Changed default value for tail_bound in NeuralSplineFlow to 5.

Fixed

  • Fixed a bug when plotting the state plot from a saved instance of the sampler where the sampling time was changed based on the current time.
  • Fixed a bug when using plot_trace, plot_1d_comparison or plot_live_points with a single parameter
  • Total sampling time is now correctly displayed when producing a state plot from a saved sampler.
  • Fixed a bug when using unbounded priors related to Model.verify_model
  • Fix inversion-split with RescaleToBounds
  • Fixed AugmentedGWFlowProposal.
  • Fixed a bug with plot_live_points when the hue parameter (c) was constant.
  • Fixed a bug with the reparameterisation Rescale when scale was set to a negative number.
  • Fixed a bug where scale could not be changed in ToCartesian.
  • Fixed a error when specifying NullReparameterisation (!82)
  • Fix typo in FlowProposal.set_poolsize_scale when acceptance=0
  • Fixed unintended behaviour when rescale_parameters is a list and boundary_inversion=True, where the code would try apply inversion to all parameters in Model.names.
  • Fixed bug where z returned by FlowProposal.rejection_sampling was incorrect when using truncation (which is not recommended).
  • Fix prior_sampling
  • Fixed minor typos in nessai.proposal.flowproposal.py

Removed

  • Remove "clip" option in FlowProposal, this was unused and untested.

0.2.4 - 2021-03-08

This release includes a number of bug fixes, changes to make the GWFlowProposal consistent with LegacyGWFlowProposal and a number of new unit tests to improve test coverage.

Added

  • Add poolsize to AnalyticProposal
  • Add a test for sampling with multiprocessing.
  • Add a test for sampling with AnalyticProposal and RejectionProposal.
  • Add a test for using the proposal methods with n_pool
  • Add tests for reparameterisations.
  • Add a test for comparing GWFlowProposal and LegacyGWFlowProposal.

Changed

  • Changed prime priors in LegacyGWFlowProposal to not update. This improves efficiency.
  • Changes to the reparameterisations to the the proposal consistent with LegacayGWFlowProposal:
    • Use [-1, 1] when inversion is enabled but not applied
  • Improved errors when reparameterisations are configured incorrectly.

Fixed

  • Fixed a bug with saving results when multiprocessing is enabled.
  • Fixed a bug with AnalyticProposal introduced in the last release.
  • Fixed a bug with resuming when using certain reparameterisations.

0.2.3 - 2021-02-24

Add support for Python >= 3.6 and other minor changes and bug fixes

Added

  • Badges for DOI and PyPI versions.
  • Add support for Python >= 3.6.
  • Improve doc-strings and tweak settings for doc-strings in the documentation.
  • Add tests for plotting functions.
  • Added sections to README and docs on citing nessai.

Changed

  • Remove := operator to enable support for Python >= 3.6.
  • Plotting functions are now more consistent and all return the figure if filename=None.

Fixed

  • Fixed bug when plotting non-structured arrays with plot_1d_comparison and specifying parameters.
  • Fixed bug where plot_indices failed if using an empty array but worked with an empty list.

Removed

  • Remove plot_posterior because functionality is include in plot_live_points.
  • Remove plot_likelihood_evaluations because information is already contained in the state plot.
  • Remove plot_acceptance as it is only by augmented proposal which is subject to change.
  • Remove plot_flow.

0.2.2 - 2021-02-19

This release was added to trigger Zenodo for producing a DOI.

Added

  • Docs badge

0.2.1 - 2021-02-18

Minor repository related fixes. Core code remains unchanged.

Added

  • PyPI workflow to automatically release package to PyPI

Fixed

  • Fixed issue with README not rendering of PyPi

0.2.0 - 2021-02-18

First public release.

Added

  • Complete documentation
  • Use setup.cfg and pyproject.toml for installing package
  • reparemeterisations submodule for more specific reparameterisations
  • half_gaussian.py example

Changed

  • Change to use main instead of master
  • Default GWFlowProposal changed to used reparameterisations
  • Split proposal.py into various submodules
  • Minor updates to examples
  • max_threads default changed to 1.

Fixed

  • Fix a bug where maximum_uninformed did not have the expected behaviour.

Deprecated

  • Original GWFlowProposal method renamed to LegacyGWFlowProposal. Will be removed in the next release.