Below, you can find a linked version of the changelog available at the PyRates github repository.




  • updated readthedocs configuration file

  • added keyword argument adaptive to the CircuitTemplate.get_run_func method, which allows to indicate whether the generated equation file is expected to be called with an adaptive step-size solver (adaptive=True) or not

  • reduced computational overhead for the creation and simulation of delayed differential equation systems

  • removed a bug where edge attribute dictionaries were changed by mistake during the CircuitIR instantiation

  • improved working directory management in the backend


  • simplified automated generation of unique variable names (recursive calls etc. were replaced with look-up tables)

  • improved variable passing between different operators within a node. Less additional variables are now created, thus reducing the memory load


  • fixed bug in fortran backend where the NPAR parameter for Auto-07p files was not properly set

  • improved code readability in fortran backend

  • moved selection of output variables from the results of a numerical simulation from the backend to the computegraph, thus reducing the amount of variables that had to be passed between the different classes

  • after each simulation, the value of all state variables in the compute graph is updated to the value at the final simulation step

  • added functionalities to the CircuitTemplate that allow to remember the state of all network variables from a previous simulation, even if a new backend is chosen for function generation or more simulations


  • added a background input parameter to the izhikevich population template

  • updated the documentation example for parameter sweeps to account for recent changes in the keyword arguments to the grid_search function

  • changed keyword argument vectorization of the function grid_search to vectorize, to be consistent with the naming of the same argument in

  • updated the CircuitTemplate.add_edges_from_matrix method to allow for edges that connect separate network nodes


This official release is a combination of all the bug fixes and improvements in the pre 1.0 versions.

Minor improvements since 0.17.4:

  • removed typos in documentation

  • improved layout of the online documentation

  • updated documentation to account for latest changes

  • removed bug where a delayed differential equation model would lead to a KeyError when trying to generate the run function



  • added sign function to the backend that returns the sign of its input (1 or -1)

  • improved readthedocs documentation (removed bug with display of math, added new use example for edge templates)

  • added a safe guard for defining edge templates: An error is raised now when edge template input variables have the same name as their source variable.


  • minor debugging of the model introduction use examples

  • adjustments of the template cheat sheet template_specification.rst

  • debugged issue in base backend, where file names specified by the users that contained a file ending wre not handled properly

  • debugged issue with fortran backend where file names that contained a directory path were not handled properly for module imports

  • debugged issue with fortran backend where adjustments of the default auto meta parameters were not applied correctly


  • the state variable indices and parameter names returned as the fourth and third return values of CircuitTemplate.get_run_func, respectively, now use the frontend variable names instead of the backend variable names

  • implemented a method CircuitIR.get_frontend_varname that returns the frontend variable name given a backend variable name


  • changed the theme of the readthedocs documentation website

  • added documentation for all supported backend functions

  • added documentation for dependencies and requirements

  • added documentation for YAML template structure to the documentation website

  • added documentation for mathematical syntax

  • added the changelog to the documentation website


  • added __getitem__ methods on all frontend template classes that allow for a less convoluted examination of the major properties of the template classes

  • added pytests that test these new features

  • users can now quickly access each node on CircuitTemplate, each operator on NodeTemplate and EdgeTemplate, and each variable on OperatorTemplate



  • added class for interactive grid search results visualization to utility

  • changed organization of the pandas DataFrames that grid-search returns: Each different parameterization of the model appears only once in the param_grid.index and the results DataFrame uses a full hierarchical column organization.

  • The pandas DataFrame returned by uses a fully hierarchical column organization now: Every node hierarchy level is a separate level in the column index hierarchy.

  • minor docstring improvements

  • fixed bug in edge equation setup where a wrong index was provided to the target variable sometimes

  • fixed bug in variable updating that occurred for numpy.ndarray variables where the shape attribute was an empty tuple

  • applied all changed to the gallery examples in the documentation



  • added generic method for state variable indexing to that is used for all edge-related indexing operations now (replacing multiple, slightly different implementations at various places in

  • added an alternative compute graph class that can be used to generate function files that do not perform in-place manipulations of the vectorfield dy but instead just create a new variable. This is relevant for gradient-based optimization.

  • improved the modularity of the ComputeGraph

  • added a method add_import to the backend that allows adding import statements to the top of a function file

  • added a backend function concatenate that can be used in equation strings now in order to combine vectorized variables

  • removed a bug where calling clear_frontend_caches did not clear all IR caches properly


  • added support for models with vectorized state-variables

  • improved performance of edge operations

  • more detailed output about returned function arguments when calling CircuitTemplate.get_run_func

  • improved memory consumption during model initialization

  • complex-valued models use complex variable types for all variables and parameters now, to prevent type conversions

  • added a new method CircuitTemplate.get_var that allows users to access backend variables after calling CircuitTemplate.get_run_func

  • added automated reduction of vectorized constants, if all constants are identical

  • added possibility to pass iterables to CircuitTemplate.update_var, thus allowing to update vectorized variables in one go

  • updated CircuitTemplate.add_edges_from_matrix such that only edges with non-zero weights are added to the CircuitTemplate instance



  • run-function generating method of ComputeGraph now returns the keys of the function arguments together with the arguments

  • implemented a method in CircuitTemplate that allows to get the indices of state variables within the system state vector


  • updated changelog


  • added different versions of the Izhikevich mean-field model (the dimensionless model, the biophysical model with distributed background currents, and the biophysical model with distributed spike thresholds)

  • improved documentation gallery examples (debugged equations, added images, added Izhikevich model references)


  • added Heun’s method as a new differential equation solver method

  • Heun’s method was integrated with all backends

  • a test was added that ensures correct functionality of Heun’s method

  • the usage of the method is demonstrated in the simulations gallery example

  • added hyperlinks to websites explaining the different numerical solvers in the gallery example

  • improved the backend implementation of choosing between different solvers (less code overlap between backends now)



  • added support for delayed differential equation (DDE) systems

  • a function past(y, tau) is now available for any backend that allows to evaluate a state variable y at time t-tau

  • edges with discrete delays that are to be used in combination with an adaptive step-size solver are translated into past calls

  • a gallery example was added that demonstrates how to interface the Python package ddeint via a DDE system generated by PyRates

  • the Julia backend received support for performing DDE simulations from within PyRates via its interface to DifferentialEquations.jl



  • debugged latex equation error in Izhikevich model gallery example

  • bugfix in julia backend where a wrong file ending was provided

  • added new pytests for the izhikevich model, the python model definition interface and the CircuitIR translation

  • updated the readme

  • added a new QIF model template that includes conductance-based synapses


  • added gallery example for the izhikevich mean-field model

  • updated readme

  • updated changelog

  • updated default parameterization of the izhikevich model


  • added a matlab backend (mainly for code generation, since simulations are very slow due to array conversion between numpy and matlab)

  • added a mean-field model of the Izhikevich neuron

  • small bug fixes

    • removed an issue of the fortran interface to Auto-07p that led to wrong function argument indices being generated

    • removed an issue with synaptic weights of -1 being converted to 1

    • removed a compatibility issue between old and new versions of the ‘to_yaml’ methods

  • added the natural logarithm ‘log’ as backend function



  • removed bug where vectorized circuits with multiple edges to the same target wre not resolved correctly

  • removed bug where creating deepcopies of a CircuitTemplate raised an error for scalar-valued models

  • added a new gallery example demonstrating different ways of adding delays to models

  • added a new gallery example demonstrating the different options to optimize run times of numerical simulations


  • added support for complex-valued systems

  • added model templates for the kuramoto order parameter and the theta neuron model

  • added model templates for the van der pol oscillator and the stuart-landau oscillator

  • added support for Python 3.9

  • added new example galleries

  • extended pytest library

  • added the CircuitTemplate.to_yaml method that allows to save a given CircuiTemplate instance to a YAML definition file

  • added the CircuitTemplate.add_edges_from_matrix method that allows to connect nodes in a CircuiTemplate instance via connectivity matrices

  • deleted old, deprecated code fragments

  • removed the dependecy on pyparsing



  • updates to changelog and


  • reworked features:

    • Restructured backend

      • new backends (torch, Julia)

      • sympy-based equation parsing

      • improved compute graph

      • improved generation of run functions from compute graphs

    • Improved frontend

      • easier imports

      • additional convenience functions for simulations

      • less steps from model definition to simulation

      • reduced syntax for model definitions

    • Removed utility package

      • utility packages for parameter optimization, signal analysis and visualization have been removed from the pyrates main package

      • most utility functionalities have been moved to separate repositories of the pyrates-neuroscience organization

      • less package requirements

    • new model templates

      • improved structure of the model templates

      • New model templates and documentation examples

      • new example galleries and jupyter notebooks with hands-on use examples



  • Reworked features:

    • CircuitIR._add_edge_buffer() was re-worked, such that the algorithm that translates gamma-kernel convolutions for edges into ODE systems is more transparent and computationally less expensive

    • additionally improved the source code documentation of CircuitIR._add_edge_buffer()

    • removed unnecessary copying/indexing operations of original edge source variable


  • Bug fixes:

    • fixed a bug in CircuitIR._add_edge_buffer() that caused a mix-up between edges when data was transferred from the originial output into the buffer variables.

  • Performance improvements:

    • zero-weight edges are now removed much earlier in the compilation process, thus reducing compilation time.


  • Bug fixes:

    • fixed a bug in CircuitIR._add_edge_buffer() that caused a mix-up between edges when some outputs of a node had delays while others had not.

  • Usability improvements:

    • changed CircuitIR.vectorize_edges() in such that zero-weight edges are removed during the vectorization, even if they have a delay defined on them (previously, defining a delay on a zero-weight edge kept that edge in the graph).


  • Documentation changes:

    • corrected mistake in the documentation of, where arguments that refer to the source variable of an edge, where erroneously described as target variable information.

  • Bug fixes:

    • fixed bug in where the conversion from discrete delays to gamma-kernel convolutions led to a mix-up between different edges in some special cases.

    • fixed bug in pyrates.utility.pyauto.PyAuto._start_from_solution() where certain special solution branches from Auto-07p could not be properly handled

  • Usability improvements:

    • changed pyrates.utility.grid_search.adapt_circuit() such that node properties are always deep-copied before they are changed. This allows users to change the values of parameters on specific node operators, even though that exact same operator has been used to define multiple nodes in the network. Previously, changing the value of the parameter on one node led to changes on all other nodes as well.

    • improved stability and usability of pyrates.utility.visualization.Interactive2DParamPlot. A title for the 2D plot can now be passed, a colorbar is added, and the location of the axis ticks of the 2D plot was improved


  • Documentation updates:

    • all Jansen-Rit model introductions where changed to track the excitatory and inhibitory post-synaptic potentials of the pyramidal cell population as output variables. Their difference provides the average membrane potential of the pyramidal cells.

    • Changed documentation jupyter notebooks etc. to account for Jansen-Rit model definition change (see below).

    • adjusted to delete all temporary files created by auto-07p

  • model templates updates:

    • added a 3 population model to the qif model templates in simple_montbrio.yaml

    • added qif population template with mono-exponential synaptic depression to simple_montbrio.yaml

    • added a new model template to simple_montbrio.yaml which provides a QIF population with mono-exponential spike-frequency adaptation

    • added bi-exponential short-term adaptation descriptions to QIF models in simple_montbrio.yaml

    • small change to the Jansen-Rit model definition: I removed the observer operator. To investigate the PC membrane potential, please record both PSP variables at the PC population and plot their sum. This has been changed accordingly in all corresponding examples.

  • PyAuto related updates:

    • altered the pyrates.utility.pyauto.PyAuto.to_file method. Additional keyword arguments that are provided by the user are now stored in a dictionary under additional_attributes. Loading a pyauto instance via from_file will thus create an attribute additional_attributes on the instances, which will contain all the keyword arguments as a dictionary.

    • debugged the pyrates.utility.pyauto.get_from_solutions method. Previously, providing more than one attribute key resulted in the method using an erroneous list comprehension style. This was fixed now. Providing multiple keys now results in the method returning a list of lists.

    • changed the way automatic re-runs of starting points computed by auto are detected by pyrates.utility.pyauto.PyAuto

    • fixed problem with extracting a solution from auto via the method pyrates.utility.pyauto.PyAuto.get_solution(). Apparently, sometimes the function call solution_branch(solution_key) does not work and throws an attribute error. I implemented a work around for this inconsistency in the Python interface for auto-07p.

    • changed pyrates.utility.pyauto.continue_period_doubling_bf to return a list that contains the names of all period doubling continuations performed with the pyauto instance that is returned as a second return value

    • now catching an error in the plotting-related method pyrates.utility.pyauto.PyAuto._get_line_collection, if the x argument is a vector of length 1

    • debugged pyrates.utility.pyauto.PyAuto.get_point_idx(). Sometimes, when auto-07p failed to locate the new fixed point of a steady-state solution, it retries the previous step. PyAuto could not recognize the auto-07p diagnostic output for such cases. Now it can.

    • improved period doubling continuation in Only solution branches with new PD bifurcations are saved for plotting etc.

    • adjusted pyrates.utility.pyauto.PyAuto.plot_continuation method such that it can be used to plot continuations of the time parameter “PAR(14)”

    • adjusted pyrates.utility.pyauto.PyAuto.plot_trajectory to be able to plot phase space trajectories of explicit time continuations (continuations in “PAR(14)”)

    • adjusted the return values of the pyrates.utility.pyauto.fractal_dimension method for its extreme cases. If the sum of the lyapunov spectrum is positive, return the number of lyapunov exponents. If the largest lyapunov exponent is smaller or equal to zero, use the normal formula.

    • added a cutoff argument to the pyrates.utility.pyauto.PyAuto.plot_trajectory method that allows to cut off initial transients within the time window from t=0 until t=cutoff.

    • implemented speed-up of pyrates.utility.pyauto.PyAuto.get_eigenvalues() method and fixed two bugs with the method that (1) led to an empty list being returned, and (2) caused the method to fail when applied to a steady-state solution

    • improved continuation of period doubling cascades via pyrates.utility.pyauto.continue_period_doubling_bf(): It recognizes now which branches it had already switched to at period doubling bifurcations. Reduces the number of overall continuations

    • added the possibility to pass the installation directory of auto-07p to pyrates.utility.pyauto.PyAuto, pyrates.utility.pyauto.PyAuto.from_file and This makes it easier to install auto-07p, since the users do not have to manupilate system path variables themselfes anymore

    • debugged counting of already calculated parameter continuations in pyrates.utility.pyauto.PyAuto

    • adjusted the method together with the pyrates.backend.fortran_backend.FortranBackend.clear() method to remove all temporary files created by us or auto-07p during the model compilation and execution.

  • grid-search updates:

    • added a warning to the pyrates.utility.grid_search.grid_search() function if a certain parameter is not found in the model

    • improved interface between pyrates.utility.grid_search.grid_search() function and pyrates.utility.grid_search.ClusterGridsearch class

    • added a keyword argument clear to grid_search that prevents removal of temporary files if set to False

  • visualization updates:

    • improved the interactive 2D plot in

    • Debugging of pyrates.utility.visualization.Interactive2DParamPlot: retrieving the column index of each column name now handles multi-column Dataframes correctly.

  • backend updates:

    • replaced “is” comparisons with “==” comparisons where appropriate

  • evolutionary optimization updates:

    • changed the way model ids are sampled in pyrates.utility.genetic_algorithm.DifferentialEvolutionAlgorithm. With the old method, multiple workers sometimes generated models with equal IDs, leading to errors.

    • added an argument to that allows to suppress runtime warnings.

  • intermediate representation updates:

    • fixed a bug in that led to a wrong association between node indices and node variables in cases where multiple delayed edges with different delay profiles had to be handled. This mostly affected grid-searches over delay distribution parameters.

    • passed the verbose argument of to the backend run function. Now all printed output of PyRates can be muted.


  • Updated documentation

  • Removed conversion function register, because the functions were not used and made the code unnecessarily complicated

    • might be replaced by a graph-based conversion path-finder in the future, if necessary

  • Extended support for loading circuits from and saving to files

    • supported formats: yaml, pickle

    • supported classes: templates

  • Removed all imports in for increased stability. Previously, importing something from pyrates.utility, would have required a user to install optional packages that might not have been needed. Now all utility functions need to be imported from sub-files in the pyrates.utility module instead of directly from the module.

  • Added optional install collection tests that includes all packages necessary to run the tests. Also restricted the travis CI build to use only the tests installation instead of the full installation.

  • Added feature to pass a dictionary to CircuitTemplate.apply() in order to adapt values of variables on the fly. This behaviour was already supported by all other parts of the hierarchy, only circuits missed out until now.


  • Added experimental support for multiple source variables per edge

    • edges can either have multiple input variable from the same input node, or

    • they can have additional (“modulating”) input from any node in the network

  • Added experimental support for Fortran code creation backend

  • Edge delays can now be transformed into delay distributions via convoluted Gamma-Kernels based on differential equation using a mean and spread parameter for the delay

  • various performance improvements


0.8.2 Included bug fixes from jajcayn:

  • Allow to initialise CircuitTemplate with instances of EdgeTemplate instead of a template path, previous behaviour is unaffected.

  • Fix writing graph to the file by passing _format along until the end

0.8.1 Improved cluster distribution and bug fixes

  • updated tensorflow dependency to >=2.0, fixes some dependency problems

  • Improved cluster distribution system, available under pyrates.utility.grid_search

  • New feature: model optimization with genetic algorithms, available under pyrates.utility.genetic_algorithm

  • Miscellaneous bug fixes


  • removed version ID numbers of operator/node instances in the intermediate representation. I.e. a node label mynode was previously renamed to mynode.0 and will now keep it’s original label.

  • moved all functionality of ComputeGraph into CircuitIR, which is now the main interface for the backend.

    • CircuitIR now has a .compile method that performs all vectorization and transformation into the computable backend form.

  • vectorization will transform all nodes into instances of VectorizedNodeIR that have labels like vector_nodeX with X being a integer index. The map between old nodes and vectorized nodes with respective index is saved in the label_map dictionary attribute of the CircuitIR

  • When adding input or sampling output of a network with multiple stacked levels of circuits, you can now use all to get all nodes within that particular level. For example mysubcircuit1/all/mynode will get all nodes with label mynode that are in one level of sub-circuits below mysubcircuit.

  • Tensorflow support now relies on the current 2.0 release candidate tensorflow-2.0-rc

  • Added optional install requirements via extras_require in