Releases | Open Source Science

Fixes

Fixed a bug introduced into pyphi.utils.load_data() by a breaking change in recent versions of NumPy that caused an error on import.
Fixed a bug where changing config.PRECISION dynamically did not change constants.EPSILON, causing some comparisons that relied on constants.EPSILON to not reflect the new precision.
Changing config.FS_CACHE_DIRECTORY and config.FS_CACHE_VERBOSITY now causes a new joblib.Memory cache to be created. Previously, changing these options dynamically had no effect.
Made test suite compatible with stricter usage of pytest fixtures required by recent versions of pytest.

API additions

Added pyphi.tpm.reconstitute_tpm().

API changes

Renamed pyphi.partition.partition_registry to pyphi.partition.partition_types.
Renamed pyphi.distance.bld() to pyphi.distance.klm().
Fixed the connectivity matrix of the disjunction_conjunction_network().
Removed 'expanded_*_reperotire' attributes of JSON-serialized Concepts.

Config

Added the WELCOME_OFF option to turn off the new welcome message.

Documentation

Added documentation for the partition_types registry.
Added documentation for the filesystem and database caches.

- Python
Published by wmayner over 6 years ago

Fixes

Fixed a memory leaked when concepts returned by parallel CES computations were returned with distinct subsystem objects. Now all objects in a CES share the same subsystem reference.
Fixed a race condition caused by newly introduced tqdm synchronization. Removed the existing ProgressBar implementation and pinned tqdm to version >= 4.20.0.
Made model hashes deterministic (6b59061). This fixes an issue with the Redis MICE cache in which cached values were not shared between processes and program invokations.

API additions

Added a NodeLabels object for managing the labels of network elements. Most models now carry a NodeLabels instance that is used for string formatting.
Added the cut_node_labels property to Subsystem and MacroSubsystem.
Added utils.time_annotated decorator to measure execution speed.

API changes

Specifying the nodes of a Subsystem is now optional. If not provided, the subsystem will cover the entire network.
Removed the labels2indices, indices2labels and parse_node_indices methods from Network, and the indices2labels method from Subsystem.
Renamed config.load_config_file to config.load_file, and config.load_config_dict to config.load_dict
Removed backwards-compatible Direction import from constants module.
Renamed macro.coarse_grain to coarse_graining.
Exposed coarse_grain, blackbox, time_scale, network_state and micro_node_indices as attributes of MacroSubsystem.

Config

Removed the LOG_CONFIG_ON_IMPORT configuration option.

- Python
Published by wmayner over 7 years ago

API changes

Modules

Renamed:
- compute.big_phi to compute.network
- compute.concept to compute.subsystem
- models.big_phi to models.subsystem
- models.concept to models.mechanism

Functions

Renamed:
- compute.main_complex() to compute.major_complex()
- compute.big_mip() to compute.sia()
- compute.big_phi() to compute.phi()
- compute.constellation() to compute.ces()
- compute.conceptual_information() to compute.conceptual_info()
- subsystem.core_cause() to subsystem.mic()
- subsystem.core_effect() to subsystem.mie()
- subsystem.mip_past() to subsystem.cause_mip()
- subsystem.phi_mip_past() to subsystem.phi_cause_mip()
- subsystem.phi_mip_future() to subsystem.phi_effect_mip()
- distance.small_phi_measure() to distance.repertoire_distance()
- distance.big_phi_measure() to distance.system_repertoire_distance()
- For all functions in convert:
- loli to le (little-endian)
- holi to be (big-endian)
Removed compute.concept(); use Subsystem.concept() instead.

Arguments

Renamed connectivity_matrix keyword argument of Network() to cm

Objects

Renamed BigMip to SystemIrreducibilityAnalysis
- Renamed the unpartitioned_constellation attribute to ces
- sia is used throughout for attributes, variables, and function names instead of big_mip
Renamed Mip to RepertoireIrreducibilityAnalysis
- Renamed the unpartitioned_repertoire attribute to repertoire
- ria is used throughout for attributes, variables, and function names instead of mip
Renamed Constellation to CauseEffectStructure
- ces is used throughout for attributes, variables, and function names instead of constellation
Renamed Mice to MaximallyIrreducibleCauseOrEffect
- mic or mie are used throughout for attributes, variables, and function names instead of mip
Similar changes were made to the actual and models.actual_causation modules.

Configuration settings

Changed configuration settings as necessary to use the new object names.

Constants

Renamed Direction.PAST to Direction.CAUSE
Renamed Direction.CAUSE to Direction.EFFECT

API additions

Configuration settings

Added CACHE_REPERTOIRES to control whether cause/effect repertoires are cached. Single-node cause/effect repertoires are always cached.
Added CLEAR_SUBSYSTEM_CACHES_AFTER_COMPUTING_SIA to control whether subsystem caches are cleared after calling compute.sia().

Objects

Added two new objects, MaximallyIrreducibleCause and MaximallyIrreducibleEffect, that are subclasses of MaximallyIrreducibleCauseOrEffect with a fixed direction.

Refactor

Moved network-level functions in compute.big_phi to pyphi.compute.network
Moved subsystem-level functions in compute.big_phi and compute.concept to compute.subsystem

Documentation

Added a description of TPM representations.
Improved the explanation of conditional independence and updated the example to reflect that PyPhi now raises an error if a conditionally-dependent TPM is provided.
Added detailed installation instructions.
Little-endian and big-endian replace LOLI and HOLI terminology
Added documentation for the following modules:
- distribution
- cache
- compute.parallel
- compute top-level module
- module top-level module

- Python
Published by wmayner about 8 years ago

Fixes

Refactored parallel processing support to fix an intermittent deadlock.

- Python
Published by wmayner about 8 years ago

API changes

Many functions have been refactored to different modules; see the "Refactor" section for details.
compute.possible_complexes no longer includes the empty subsystem.
Made is_cut a property.
Renamed macro.list_all_partitions and macro.list_all_groupings to all_partitions and all_groupings. Both are now generators and return nested tuples instead of lists.
Moved macro.make_mapping to CoarseGrain.make_mapping.
Moved macro.make_macro_tpm to CoarseGrain.macro_tpm.
Added blackbox functionality to macro.emergence. Blackboxing and coarse- graining are now parametrized with the blackbox and coarse_grain arguments.
Removed utils.submatrix.
Made Network.tpm and Network.cm immutable properties.
Removed the purview argument from Subsystem.expand_repertoire.
Moved validate.StateUnreachableError and macro.ConditionallyDependentError to the exceptions module.
Removed perturbation vector support.
Changed tpm.marginalize_out to take a list of indices.
Fixed macro.effective_info to use the algorithm from the macro-micro paper.
Replace constants.DIRECTIONS, constants.PAST, and constants.FUTURE with a proper Enum class: constants.Direction. Past and future are now represented by constants.Direction.PAST and constants.Direction.FUTURE.
Simplifed logging config to use config.LOG_STDOUT_LEVEL, config.LOG_FILE_LEVEL and config.LOG_FILE.
Removed the location property of Concept.

API Additions

Added subsystem.evaluate_partition. This returns the φ for a particular partition.
Added config.MEASURE to choose between EMD, KLD, or L1 for distance computations.
Added macro.MacroSubsystem. This subclass of Subsystem is used to performs macro computations.
Added macro.CoarseGrain to represent coarse-grainings of a system.
Added macro.Blackbox to represent system blackboxes.
Added validate.blackbox and validate.coarse_grain.
Added macro.all_coarse_grains and macro.all_blackboxes generators.
Added Subsystem.cut_indices property.
Added Subsystem.cm connectivity matrix alias.
Added utils.all_states, a generator over all states of an n-element system.
Added tpm.is_state_by_state for testing whether a TPM is in state-by-state format.
Network now takes an optional node_labels argument, allowing nodes to be referenced by a canonical name other than their indices. The nodes of a Subsystem can now be specified by either their index or their label.
Added models.normalize_constellation for deterministically ordering a constellation.
Added a Makefile.
Added an exceptions module.
Added distribution.purview for computing the purview of a repertoire.
Added distribution.repertoire_shape.
Added config.PARTITION_TYPE to control the ways in which φ-partitions are generated.
Added more functions to the convert module:
- holi2loli and loli2holi convert decimal indices between HOLI and LOLI formats.
- holi2loli_state_by_state and loli2holi_state_by_state convert between HOLI and LOLI formats for state-by-state TPMs.
- Added short aliases for some functions:
- h2l is holi2loli
- l2h is loli2holi
- l2s is loli_index2state
- h2s is holi_index2state
- s2h is state2loli_index
- s2l is state2holi_index
- h2l_sbs is holi2loli_state_by_state
- l2h_sbs is loli2holi_state_by_state
- sbn2sbs is state_by_node2state_by_state
- sbs2sbn is state_by_state2state_by_node
Added the Constellation.mechanisms, Constellation.labeled_mechanisms, and Constellation.phis properties.
Add BigMip.print method with optional constellations argument that allows omitting the constellations.

Refactor

Refactored the utils module into the connectivity, distance, distribution, partition, timescale, and tpm modules.
Existing macro coarse-grain logic to use MacroSubsystem and CoarseGrain.
Improved string representations of PyPhi objects.
Refactored JSON support. The jsonify module now dumps PyPhi models to a a format which can be loaded to reproduce the full object graph of PyPhi objects. This causes backwards incompatible changes to the JSON format of some model representations.
Refactored pyphi.config to be an object. Added validation and callbacks for config options.

Optimizations

Added an analytic solution for the EMD computation between effect repertoires.
Improved the time complexity of directed_bipartition_of_one from exponential to linear.

Documentation

Updated documentation and examples to reflect changes made to the macro API and usage.
Added documentation pages for new modules.

- Python
Published by wmayner about 8 years ago

Fixes

Fixed a bug in setup.py that prevented installation.

- Python
Published by wmayner about 10 years ago

API changes

Mechanisms and purviews are now passed to all functions and methods in node index form (e.g. (0, 1, 3)). Previously, many functions took these arguments as Node objects. Since nodes belong to a specific Subsystem it was possible to pass nodes from one subsystem to another subsystem's methods, leading to incorrect results.
constellation_distance no longer takes a subsystem argument because concepts in a constellation already reference their subsystems.
Moved utils.cut_mechanism_indices and utils.mechanism_split_by_cut to to Cut.all_cut_mechanisms and Cut.splits_mechanism, respectively; moved utils.cut_mice to Mice.damaged_by_cut.
Concept.__eq__: when comparing concepts for equality, we no longer directly check equality of their subsystems. Concept equality is now defined as follows:
- Same φ
- Same mechanism node indices cause/effect purview node indices
- Same mechanism state
- Same cause/effect repertoires
- Same networks This allows two concepts to be equal when e.g. the only difference between them is that one's subsystem is a superset of the other's subsystem.
Concept.__hash__: the above notion of concept equality is also implemented for concept hashing, so two concepts that differ only in that way will have the same hash value.
Disabled concept caching; removed the config.CACHE_CONCEPTS option.

API Additions

Added config.READABLE_REPRS to control whether __reprs__ of PyPhi models default to using pretty string formatting.
Added a Constellation object.
Added utils.submatrix and utils.relevant_connections functions.
Added the macro.effective_info function.
Added the utils.state_of function.
Added the Subsystem.proper_state attribute. This is the state of the subsystem's nodes, rather than the entire network state.
Added an optional Redis-backed cache for Mice objects. This is enabled with config.REDIS_CACHE and configured with config.REDIS_CONFIG.
Enabled parallel concept evaluation with config.PARALLEL_CONCEPT_EVALUATION.

Fixes

Concept.eq_repertoires no longer fails when the concept has no cause or effect.
Fixed the Subsystem.proper_state attribute.

Refactor

Subsystem Mice and cause/effect repertoire caches; Network purview caches. Cache logic is now handled by decorators and custom cache objects.
Block reducibility tests and Mice connection computations.
Rich object comparisons on phi-objects.

Documentation

Updated documentation and examples to reflect node-to-index conversion.

- Python
Published by rlmv about 10 years ago

pyphi - 0.1.6

Release v0.1.6

- Python
Published by wmayner over 10 years ago

pyphi - 0.1.5

Release v0.1.5

- Python
Published by wmayner over 10 years ago

Fixes

Fixed a bug in compute.constellation_distance where 𝛗 from partitioned concepts was not allowed to move to the null concepts; it was assumed that the unpartitioned constellation would always have greater 𝚺𝛗. In some cases, the partitioned system will actually have greater 𝚺𝛗 than the unpartitioned constellation, in which case some of that 𝛗 needs to move to the null concept during the EMD calculation. This fix changes the distance matrix so that can happen.

- Python
Published by wmayner over 10 years ago

API additions

pyphi.macro provides several functions to analyze networks over different spatial scales.
convert.conditionally_independent(tpm) checks if a TPM is conditionally independent.

API changes

𝚽 and 𝛗 values are now rounded to config.PRECISION when stored on objects.

Fixes

Tests for Subsystem_find_mip_parallel and Subsystem_find_mip_sequential.
Slow tests for compute.big_mip.

Refactor

Subsystem cause and effect repertoire caching.

Documentation

Added XOR and Macro examples.

- Python
Published by wmayner almost 11 years ago

Optimizations

Pre-compute and cache possible purviews.
Compute concept distance over least-common-purview rather than whole system.
Store relevant_connections on MICE objects for MICE cache checking.
Only recheck concepts and cut mechanisms after a system cut.

API additions

The new configuration option CUT_ONE_APPROXIMATION gives an approximation of 𝚽 by only considering cuts that cut off a single node.
Formerly, the configuration was always printed when PyPhi was imported. Now this can be suppressed by setting the LOG_CONFIG_ON_IMPORT option to false in the pyphi_config.yml file.

Fixes

Bipartition function.
MICE caching.

- Python
Published by wmayner almost 11 years ago

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.

- Python
Published by wmayner almost 11 years ago

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.

- Python
Published by wmayner almost 11 years ago

- 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).