v1.0.5

Added

• The equal-weighted posterior can now be made larger and more precise using the equal_weight_boost keyword argument of sampler.posterior.

v1.0.4

Added

• Nautilus can now use Dask clusters for parallelization (#49).

v1.0.3

Added

• The user can now specify a timeout interval for the sampler. If that time is exceeded, the sampler will not start new calculations. (#46)
• The sampler now returns whether it finished normally or stopped because the timeout or maximum number of likelihood computations was reached. (#46)

v1.0.2

Changed

• Further improved the way the sampler handles plateaus. For example, if the lowest-likelihood point in the live set is part of a plateau, the sampler will jump past the plateau if enough live points with higher likelihood exist. The sampler also behaves reasonably if most or all points are part of a plateau. This leads to less freezes and better performance.

v1.0.1

Fixed

• Fixed a crash when multiple blobs per likelihood call are returned as a single array.

v1.0.0

Added

• The user can now specify a maximum number of likelihood calls. If that number is exceeded, the sampler will automatically stop. (#23)

Changed

• Updated the terminal output to be more compact and more friendly for log files. This also removes the dependency on tqdm. (#36)
• By default, the batch size is now dynamically determined at the start based on the pool size. This should prevent issues for new users parallelizing over a large number of CPUs.

Fixed

• Fixed a crash when returning the posterior as a dictionary and with equal weight.
• Fixed a potential crash when n_update is extremely low.

Depcrecated

• The evidence, asymptotic_sampling_efficiency, and effective_sample_size sampler functions are deprecated and replaced by the log_z, eta, and n_eff properties, respectively.

Removed

• The n_jobs parameter when initializing the sampler has been removed. Use pool, instead.

v0.7.4

Fixed

• Increased numerical stability when finding the minimum-volume enclosing ellipsoid. Instances of np.linalg.LinAlgError should be reduced or eliminated. (#34)

v0.7.3

Changed

• When passing a number to the pool keyword argument, the sampler automatically caches the likelihood function in the subprocesses of the multiprocessing pool. This reduces communication between processes and can substantially improve performance.
• Changed the way ellipsoids are split, preventing rare instances where ellipsoid splitting stops prematurely. (#28)
• Improved the performance of the algorithm for finding the minimum-volume enclosing ellipsoid.

Fixed

• Fixed a rare freeze when the neural networks predict the same score for all input values. (#27)
• Prevented a rare freeze if the likelihood is multi-modal and one mode has much less volume than the other. (#29)
• Fixed an occasional crash when resuming a calculation without blobs from a checkpoint file. (#31)

v0.7.2

Changed

• The sampler should now deal better with likelihood plateaus.

Fixed

• The prior function now always receives a copy of the points, preventing buggy behavior if the user's prior function does the prior transform on the original array.

v0.7.1

Fixed

• Parallelization with MPIPoolExecutor should now work correctly.