v0.7.0

Added

• Added the function sampler.asymptotic_sampling_efficiency which returns an estimate of the sampling efficiency in the sampling phase.

Changed

• One can now change whether to discard points in the exploration phase after calling run by changing the discard_exploration argument of the sampler. To achieve this, information about points in the exploration phase is never dropped. Consequently, the sampler does not create a backup of the end of the exploration stage under "filename_exp.hdf5", anymore.
• The computational overhead was slightly reduced when new bounds are rejected. Also, the output now specifically mentions when adding a new bound failed because the volume increased.

Fixed

• Likelihoods with large plateaus shouldn't crash, anymore.
• Information updates about bounds are now written during the sampling phase. Previously, if computations were interrupted and restarted during the sampling phase, the volume estimates were noisier than necessary, and some points may have been proposed twice.

Removed

• The n_jobs keyword argument for the sampler has been removed. The pool used for likelihood calls is now also used for sampler parallelization, by default. To use independent pools for likelihood calls and sampler calculations, pass a tuple to pool.
• The random_state keyword argument for the sampler has been removed. Use seed, instead.
• The enlarge keyword argument has been removed. Use enlarge_per_dim, instead.

v0.6

Added

• Added the keyword arguments n_points_min and split_threshold to the sampler. Previously, they were not accessible.
• Sampling new points can now be parallelized using the n_jobs keyword argument.

Changed

• The code now uses the more modern numpy.random.Generator framework instead of numpy.random.RandomState.
• The default value for n_points_min is now the number of dimensions plus 50. Previously, it was hard-coded to be the number of dimensions plus 1.
• The multi-ellipsoidal decomposition has been tweaked with the goal of reducing computational overhead for high-dimensional problems.
• The default number of parallel processes has been changed to one. By default, the sampler will not use parallelization.
• Multi-ellipsoidal decomposition now uses Gaussian mixture modeling instead of K-Means. The former typically results in better performance, i.e., smaller boundaries with fewer ellipsoids.

Fixed

• The sampler now correctly writes the random number generator in the sampling phase.
• The keyword argument n_jobs is now being correctly passed when training networks. Previously, all cores were used regardless of n_jobs.
• The sampler doesn't crash when setting verbose=Trueand n_networks=0.

Deprecated

• The random_state keyword argument for the sampler has been deprecated in favor of the new keyword argument seed.

v0.5

Changed

• Introduced neural network ensembles. Instead of relying on a single network, the sampler now uses 4 networks by default. This should lead to better sampling performance. The training of the networks is done in parallel, by default. Which means that on multi-core systems, time spent on neural network training shouldn't increase dramatically. The number of networks and training parallelization can be adjusted by the user.
• Introduced new bounds that lead to lower overhead for high-dimensional problems. The number of likelihood calls should be close to unaffected by this change.
• Increased the default number of live points from 1500 to 2000.
• Limited number of threads to 1 when calculating minimum volume enclosing ellipsoid and doing K-means clustering. More often than not, thread parallelization would slow things down while increasing CPU usage.
• When setting the filepath to "filename.hdf5" and discarding points in the exploration stage, the sampler now creates a backup of the end of the exploration stage under "filename_exp.hdf5".

Fixed

• Previously, the sampler would sometimes reject new bounds because their volume was estimated to be larger than the previous bound. However, this was based on very noisy volume estimates in certain situations. The sampler now uses more precise volume estimates, which can increase sampling performance for high-dimensional problems since bounds are updated more often.
• The sampler now doesn't perform unnecessary prior transformations when calculating the fraction of the evidence in the live set. This helps lower the computational overhead when using the sampler in CosmoSIS.
• When discarding points in the exploration phase, blobs are now also correctly removed.
• Fixed a crash when blobs_dtype is a single elementary dtype and multiple blobs are returned.

Deprecated

• The enlarge keyword argument for the sampler has been deprecated in favor of the new keyword enlarge_per_dim. Specifying enlarge will be ignored but not raise an error.

Removed

• The use_neural_network keyword argument for the sampler has been deprecated. To not use neural networks, set the new keyword argument n_networks to 0.

v0.4

Changed

• Renamed "tessellation phase" to "exploration_phase". Accordingly, changed the keyword argument discard_tessellation to discard_exploration for run function.
• Increased performance for sampling new points from shells.
• Likelihood functions will now never be passed structured numpy arrays. In case they previously received structured numpy arrays, they will now receive dictionaries. Similarly, nautilus.Prior.physical_to_structure has been renamed to nautilus.Prior.physical_to_dictionary and the keyword argument of nautilus.Sampler from pass_struct to pass_dict.
• The n_like_update keyword argument for nautilus.Sampler has been renamed to n_like_new_bound.
• The number of points for which the likelihood is evaluated is now always the batch size.
• Renamed return_dict keyword argument in Sampler.posterior to return_as_dict.

Added

• Checkpointing and writing the sampler to disk.
• Support for so-called blobs, following the implementation in emcee.

v0.3

Added

• Made several parameters such as those of the neural networks available to the user.

Changed

• Changed the default number of live points and neural network structure.
• Now using threadpoolctl to limit the number of cores used by the neural network.
• The keyword argument threads passed to the Sampler class has been removed and absorbed into the pool keyword argument.

Fixed

• Fixed crash in Prior class.

v0.2

Changed

• Improved sampling strategy after tessellation. Sampling each shell now depends on the evidence, the number of likelihood calls and the effective sample size in each shell.
• Changed the way points are assigned to different neural network emulators. Effectively, the number of neural network emulators for each bound has been reduced, often to only one.
• Changed the way neural network emulators are trained.

Removed

• Removed TensorFlow backend in favor of only using scikit-learn.

Fixed

• nautilus now raises an error if sampling with less than 2 parameters. (#2)
• nautilus now correctly passes a dictionary as input to the likelihood function if pass_struct is True and vectorized is False.

v0.1

Added

• Initial beta release.