Loading a configuration

Various aspects of PyPhi’s behavior can be configured.

When PyPhi is imported, it checks for a YAML file named pyphi_config.yml in the current directory and automatically loads it if it exists; otherwise the default configuration is used.

The various settings are listed here with their defaults.

>>> import pyphi
>>> defaults = pyphi.config.defaults()

Print the config object to see the current settings:

>>> print(pyphi.config)  
  'CACHE_SIAS': False,

Setting can be changed on the fly by assigning them a new value:

>>> pyphi.config.PROGRESS_BARS = False

It is also possible to manually load a configuration file:

>>> pyphi.config.load_file('pyphi_config.yml')

Or load a dictionary of configuration values:

>>> pyphi.config.load_dict({'PRECISION': 1})

Approximations and theoretical options

These settings control the algorithms PyPhi uses.

Parallelization and system resources

These settings control how much processing power and memory is available for PyPhi to use. The default values may not be appropriate for your use-case or machine, so please check these settings before running anything. Otherwise, there is a risk that simulations might crash (potentially after running for a long time!), resulting in data loss.

Memoization and caching

PyPhi provides a number of ways to cache intermediate results.


These settings control how PyPhi handles messages. Logs can be written to standard output, a file, both, or none. If these simple default controls are not flexible enough for you, you can override the entire logging configuration. See the documentation on Python’s logger for more information.

Numerical precision

The config API

class pyphi.conf.Option(default, values=None, on_change=None, doc=None)

A descriptor implementing PyPhi configuration options.


default – The default value of this Option.

Keyword Arguments
  • values (list) – Allowed values for this option. A ValueError will be raised if values is not None and the option is set to be a value not in the list.

  • on_change (function) – Optional callback that is called when the value of the option is changed. The Config instance is passed as the only argument to the callback.

  • doc (str) – Optional docstring for the option.

class pyphi.conf.ConfigMeta(cls_name, bases, namespace)

Metaclass for Config.

Responsible for setting the name of each Option when a subclass of Config is created; because Option objects are defined on the class, not the instance, their name should only be set once.

Python 3.6 handles this exact need with the special descriptor method __set_name__ (see PEP 487). We should use that once we drop support for 3.4 & 3.5.

class pyphi.conf.Config

Base configuration object.

See PyphiConfig for usage.

classmethod options()

Return a dictionary of the Option objects for this config.


Return the default values of this configuration.


Load a dictionary of configuration values.


Load config from a YAML file.


Return a snapshot of the current values of this configuration.


Decorator and context manager to override configuration values.

The initial configuration values are reset after the decorated function returns or the context manager completes it block, even if the function or block raises an exception. This is intended to be used by tests which require specific configuration values.


>>> from pyphi import config
>>> @config.override(PRECISION=20000)
... def test_something():
...     assert config.PRECISION == 20000
>>> test_something()
>>> with config.override(PRECISION=100):
...     assert config.PRECISION == 100

Reconfigure PyPhi logging based on the current configuration.

class pyphi.conf.PyphiConfig

pyphi.config is an instance of this class.



In certain cases, making a cut can actually cause a previously reducible concept to become a proper, irreducible concept. Assuming this can never happen can increase performance significantly, however the obtained results are not strictly accurate.



When determining the MIP for \(\Phi\), this restricts the set of system cuts that are considered to only those that cut the inputs or outputs of a single node. This restricted set of cuts scales linearly with the size of the system; the full set of all possible bipartitions scales exponentially. This approximation is more likely to give theoretically accurate results with modular, sparsely-connected, or homogeneous networks.



The measure to use when computing distances between repertoires and concepts. A full list of currently installed measures is available by calling print(pyphi.distance.measures.all()). Note that some measures cannot be used for calculating \(\Phi\) because they are asymmetric.

Custom measures can be added using the pyphi.distance.measures.register decorator. For example:

from pyphi.distance import measures

def always_zero(a, b):
    return 0

This measure can then be used by setting config.MEASURE = 'ALWAYS_ZERO'.

If the measure is asymmetric you should register it using the asymmetric keyword argument. See distance for examples.



Controls whether concepts are evaluated in parallel when computing cause-effect structures.



Controls whether system cuts are evaluated in parallel, which is faster but requires more memory. If cuts are evaluated sequentially, only two SystemIrreducibilityAnalysis instances need to be in memory at once.



Controls whether systems are evaluated in parallel when computing complexes.



Controls the number of CPU cores used to evaluate unidirectional cuts. Negative numbers count backwards from the total number of available cores, with -1 meaning ‘use all available cores.’



PyPhi employs several in-memory caches to speed up computation. However, these can quickly use a lot of memory for large networks or large numbers of them; to avoid thrashing, this setting limits the percentage of a system’s RAM that the caches can collectively use.



PyPhi is equipped with a transparent caching system for SystemIrreducibilityAnalysis objects which stores them as they are computed to avoid having to recompute them later. This makes it easy to play around interactively with the program, or to accumulate results with minimal effort. For larger projects, however, it is recommended that you manage the results explicitly, rather than relying on the cache. For this reason it is disabled by default.



PyPhi caches cause and effect repertoires. This greatly improves speed, but can consume a significant amount of memory. If you are experiencing memory issues, try disabling this.



Controls whether the potential purviews of mechanisms of a network are cached. Caching speeds up computations by not recomputing expensive reducibility checks, but uses additional memory.



Controls whether a Subsystem’s repertoire and MICE caches are cleared with clear_caches() after computing the SystemIrreducibilityAnalysis. If you don’t need to do any more computations after running sia(), then enabling this may help conserve memory.



Controls whether precomputed results are stored and read from a local filesystem-based cache in the current directory or from a database. Set this to 'fs' for the filesystem, 'db' for the database.


default=0, on_change=configure_joblib

Controls how much caching information is printed if the filesystem cache is used. Takes a value between 0 and 11.


default='__pyphi_cache__', on_change=configure_joblib

If the filesystem is used for caching, the cache will be stored in this directory. This directory can be copied and moved around if you want to reuse results e.g. on a another computer, but it must be in the same directory from which Python is being run.


27017, ‘database_name’: ‘pyphi’, ‘collection_name’: ‘cache’}``

Set the configuration for the MongoDB database backend (only has an effect if CACHING_BACKEND is 'db').




‘localhost’, ‘port’



Specifies whether to use Redis to cache MaximallyIrreducibleCauseOrEffect.


6379, ‘db’: 0, ‘test_db’: 1}``

Configure the Redis database backend. These are the defaults in the provided redis.conf file.




‘localhost’, ‘port’



Specifies whether to suppress the welcome message when PyPhi is imported.

Alternatively, you may suppress the message by setting the environment variable PYPHI_WELCOME_OFF to any value in your shell:

export PYPHI_WELCOME_OFF='yes'

The message will not print if either this option is True or the environment variable is set.


default='pyphi.log', on_change=configure_logging

Controls the name of the log file.


default='INFO', on_change=configure_logging

Controls the level of log messages written to the log file. This setting has the same possible values as LOG_STDOUT_LEVEL.


default='WARNING', on_change=configure_logging

Controls the level of log messages written to standard output. Can be one of 'DEBUG', 'INFO', 'WARNING', 'ERROR', 'CRITICAL', or None. 'DEBUG' is the least restrictive level and will show the most log messages. 'CRITICAL' is the most restrictive level and will only display information about fatal errors. If set to None, logging to standard output will be disabled entirely.



Controls whether to show progress bars on the console.


If you are iterating over many systems rather than doing one long-running calculation, consider disabling this for speed.


default=6, on_change=configure_precision

If MEASURE is EMD, then the Earth Mover’s Distance is calculated with an external C++ library that a numerical optimizer to find a good approximation. Consequently, systems with analytically zero \(\Phi\) will sometimes be numerically found to have a small but non-zero amount. This setting controls the number of decimal places to which PyPhi will consider EMD calculations accurate. Values of \(\Phi\) lower than 10e-PRECISION will be considered insignificant and treated as zero. The default value is about as accurate as the EMD computations get.



Controls whether PyPhi checks if the subsystems’s state is possible (reachable with nonzero probability from some previous state), given the subsystem’s TPM (which is conditioned on background conditions). If this is turned off, then calculated \(\Phi\) values may not be valid, since they may be associated with a subsystem that could never be in the given state.



Controls whether PyPhi checks if a system’s TPM is conditionally independent.



If set to True, the \(\Phi\) value of single micro-node subsystems is the difference between their unpartitioned CauseEffectStructure (a single concept) and the null concept. If set to False, their \(\Phi\) is defined to be zero. Single macro-node subsystems may always be cut, regardless of circumstances.


default=2, values=[0, 1, 2]

Controls the verbosity of __repr__ methods on PyPhi objects. Can be set to 0, 1, or 2. If set to 1, calling repr on PyPhi objects will return pretty-formatted and legible strings, excluding repertoires. If set to 2, repr calls also include repertoires.

Although this breaks the convention that __repr__ methods should return a representation which can reconstruct the object, readable representations are convenient since the Python REPL calls repr to represent all objects in the shell and PyPhi is often used interactively with the REPL. If set to 0, repr returns more traditional object representations.



Controls whether numbers in a repr are printed as fractions. Numbers are still printed as decimals if the fraction’s denominator would be large. This only has an effect if REPR_VERBOSITY > 0.



Controls the type of partition used for \(\varphi\) computations.

If set to 'BI', partitions will have two parts.

If set to 'TRI', partitions will have three parts. In addition, computations will only consider partitions that strictly partition the mechanism. That is, for the mechanism (A, B) and purview (B, C, D) the partition:

A,B    ∅
─── ✕ ───
 B    C,D

is not considered, but:

 A     B
─── ✕ ───
 B    C,D

is. The following is also valid:

A,B     ∅
─── ✕ ─────
 ∅    B,C,D

In addition, this setting introduces “wedge” tripartitions of the form:

 A     B     ∅
─── ✕ ─── ✕ ───
 B     C     D

where the mechanism in the third part is always empty.

Finally, if set to 'ALL', all possible partitions will be tested.

You can experiment with custom partitioning strategies using the pyphi.partition.partition_types.register decorator. For example:

from pyphi.models import KPartition, Part
from pyphi.partition import partition_types

def single_node_partitions(mechanism, purview, node_labels=None):
   for element in mechanism:
       element = tuple([element])
       others = tuple(sorted(set(mechanism) - set(element)))

       part1 = Part(mechanism=element, purview=())
       part2 = Part(mechanism=others, purview=purview)

       yield KPartition(part1, part2, node_labels=node_labels)

This generates the set of partitions that cut connections between a single mechanism element and the entire purview. The mechanism and purview of each Part remain undivided - only connections between parts are severed.

You can use this new partititioning scheme by setting config.PARTITION_TYPE = 'SINGLE_NODE'.

See partition for more examples.



When computing a MaximallyIrreducibleCause or MaximallyIrreducibleEffect, it is possible for several MIPs to have the same \(\varphi\) value. If this setting is set to True the MIP with the smallest purview is chosen; otherwise, the one with largest purview is chosen.



If set to True, the distance between cause-effect structures (when computing a SystemIrreducibilityAnalysis) is calculated using the difference between the sum of \(\varphi\) in the cause-effect structures instead of the extended EMD.


default='3.0_STYLE', values=['3.0_STYLE', 'CONCEPT_STYLE']

If set to '3.0_STYLE', then traditional IIT 3.0 cuts will be used when computing \(\Phi\). If set to 'CONCEPT_STYLE', then experimental concept-style system cuts will be used instead.


Log current settings.