0.5.0

Optimizations

• Concepts are only recomputed if they could have been changed by a cut.
• Cuts are evaluated individually, rather than in bidirectional pairs, which
  allows for better parallel performance.

API changes

• Removed the unused validate.nodelist function.

API additions

• The new configuration option ASSUME_CUTS_CANNOT_CREATE_NEW_CONCEPTS gives
  an approximation of 𝚽 by only recomputing concepts that exist in the
  unpartitioned constellation. This is much faster in certain cases.
• The methods used in determining whether a cut could effect a concept are
  exposed via the utils module as:
  • utils.cut_mechanism_indices
  • utils.cut_concepts
  • utils.uncut_concepts
• Added the pyphi.Subsystem.connections_relevant_for_concept method.
• Added the pyphi.Subsystem.cut_matrix property.

Fixes

• pyphi.compute.main_complex now returns an empty BigMip if there are no
  proper complexes.
• No longer using LRU-caches implemented as a circular, doubly-linked list;
  this was causing a huge number of recursive calls when pickling a Subsystem
  object (since caches are stored on subsystems since v0.3.6) as pickle
  traversed the (potentially very large) cache.
• pyphi.json.make_encodable now properly handles NumPy numeric types.

0.4.0

Optimizations

• compute.big_mip is faster for reducible networks when executed in parallel;
  it returns immediately upon finding a reducible cut, rather than evaluating
  all cuts. NOTE: This introduces a race condition in cases where there is
  more than one reducible cut; there is no guarantee as to which cut will be
  found first and returned.
• compute.complexes prunes out subsystems that contain nodes without either
  inputs or outputs (any subsystem containing such a node must necessarily have
  zero 𝚽).

API changes

• compute.complexes: returns only irreducible MIPs; see optimizations.
• compute.big_mip
  • New race condition with cuts; see optimizations.
  • The single-node and null BigMip's constellations are now empty tuples
    instead of empty lists and None, respectively.
• models.Concept.eq_repertoires no longer ensures that the networks of each
  concept are equal; it only checks if the repertoires are the same.

API additions

• compute.all_complexes returns the BigMip of every subsystem in the
  network's powerset (including reducible ones).
• compute.possible_main_complexes returns the subsystems that survived the
  pruning described above.

Fixes

• Network tests.

Refactor

• Network attributes. They're now implemented as properties (with getters and
  setters) to facilitate changing them properly. It should be possible to use
  the same network object with different states.
• Network state validation.
• utils.phi_eq is used wherever possible instead of direct comparisons to
  constants.EPSILON.

0.3.8

• Choosing the minimal MIP in pyphi.compute.big_mip is now more
  deterministic.
• When comparing BigMips with > and <, the 𝚽-values are compared only up
  to config.PRECISION.
• BigMips are now truthy if their 𝚽 is greater than or equal to
  config.EPSILON (previously was strictly greater than).

0.3.7

• BigMips and Concepts now have their calculation time in the time
  attribute. BigMips also have the attribute small_phi_time, which gives
  the total time taken to calculate all concepts (not just the ones that
  exist).
• If PARALLEL_CUT_EVALUATION is set to false, the cut evaluation will short-circuit as soon as a MIP of zero 𝚽 is found.
• Fixed models._numpy_aware_eq, which failed when comparing two iterables where
  one is a prefix of the other.
• Fixed models.Concept.location, which raised an error when the
  concept's cause or effect was None.

0.3.6

• LRU-caches for cause and effect repertoires are now local to the Subsystem, improving memory performance, especially in cases where many different subsystems are being analyzed in the same script.
• Precomputed hamming matrices for up to 9 nodes are now distributed with the PyPhi package, rather than being computed every time. If you're working with more than 9 nodes, be sure to precompute the hamming matrices and add them to your PyPhi installation's data/hamming_matrices directory.

0.3.4

Removed LRU caching on all pyphi.compute methods.

For long computations involving many networks, the compute module's caches
would begin to fill up the available memory, leaving no space for the much more
useful cause and effect repertoire caches.

This will be more comprehensively addressed in future releases.

0.3.3

• Fixed strong connectivity check.
• Added support for expanding repertoires over a particular purview.

0.3.2

• Added configuration option for filesystem-caching verbosity.
• If any options are missing in the configuration file, they will gracefully fall back to the defaults.

0.3.1

Network state validation is configurable (and disabled by default).

0.3.0

Added support for nondeterministic TPMs.