What’s new?#

1.0.0 Release Notes#

Despite this version of hnn-core being marked as version 1.0.0, this software package has been stable for several years and is actively being used by many different scientists for ongoing research. We decided to release this as version 1.0.0 primarily to communicate that hnn-core is, and has been, ready for active use. It should not be thought of as “still being built”. That said, we will still continue by adding new features and making it easier to use than ever before!

To be clear, this means the versions go directly from 0.6.1 to 1.0.0, and there is no version 0.7.0, etc. Please enjoy the numerous additions!

New Features#

  • Katharina Duecker has added an exciting, brand-new model, duecker_ET_model, which includes new human-derived pyramidal cell and interneuron models (L2_inhibitory and L5_inhibitory) along with new ion channel and synapse mechanisms (including NMDA_gao2021 and gabab_neymotin2016 synapses, and channels from Hay et al. 2011 and Golomb et al. 2007). A preprint is available here, the final publication is pending, and the source code for the simulations in the paper is available here. Importantly, the model is still subject to change until final publication. You can easily include simulation of its optional ERP drives using the argument add_erp_drives=True. Merged in #1227 with additional updates by Katharina Duecker, Nick Tolley, and Austin E. Soplata in #1345, #1358, and #1363.

  • HNN-Core now supports NEURON 9 (up to 9.0.2, inclusive) and Python up to 3.14 (inclusive). Please note that due to the removal of legacy units in NEURON 9.0, ALL simulations starting with this HNN-Core version will be quantitatively different! However, your simulations should still be qualitatively similar, and the numerical differences will likely be negligible. As an example, this should be equivalent to changing your calcium current reversal potential (E_Ca) from a value of 134.02 to 134.54, or similar magnitude. If your findings are robust to initial conditions and seeds, then you should have nothing to worry about. By Austin E. Soplata and Claude in #1230.

    • Note: If you want exact bit-perfect reproduction of your simulations from before this change, please install the earlier stable version using pip install "hnn-core[all]==0.6.1". The prior stable version (v0.6.1) will always force usage of NEURON 8 along with the legacy units. If you run into issues with updating your code, simulations, or reproducibility, then please reach out to us on our Github Discussions page.

  • MPI parallelism in hnn-core can now be installed entirely through pip (or uv)! On MacOS and Linux (including Windows WSL), pip install "hnn-core[parallel]" (or, alternatively, pip install "hnn-core[all]") now installs the mpi4py and openmpi PyPI packages automatically. See our Installation Guide for details. MPIBackend can automatically locate and use the MPI library provided by the openmpi package, without the need to set any environment variables yourself. MPIBackend also has two new arguments for advanced users (such as on HPC systems): autoload_mpi_library to disable this openmpi pip package behavior, and mpi_lib_path if you want to explicitly provide a path to a custom MPI library file that you want NEURON to use. This means you can use venv or uv to install hnn-core with MPI support without needing any Anaconda packages at all. This is backwards-compatible, and older conda-based MPI installs will still work. By Austin E. Soplata and Claude in #1377 (thanks to Mainak Jas for the tip!)

  • For Optimization, users can now specify their own custom objective function in two different ways:

    1. Passing obj_fun='custom' to Optimizer(). If passing the string 'custom', then users can provide a custom function by passing loss_fun=my_function to Optimizer.fit(). The only restrictions on this function are that it must accept both a single Dipole object and obj_fun_kwargs, similar to other objective functions. More details are available in the descriptions for Optimizer and fit().

    2. Passing their own function (or callable) directly via obj_fun=<callable> to Optimizer. If users want to use an objective function without the restrictions of the first way (above), then they can write their own function similar to the existing code in optimization/objective_functions.py and pass that function to Optimizer(... obj_fun=<callable>). Note that this function must accept all the same arguments as the other objective functions in optimization/objective_functions.py and run a simulation.

    By Vaishnavi Baghel, Nick Tolley, and Austin E. Soplata in #1271.

  • (Beta feature) plot_dipole() can now show overlay arrows marking the onset times of evoked drives on dipole plots, similar to “Original HNN”, via the new arguments show_drive_arrows, arrow_width, arrow_height, and net. This feature is in beta, may still have bugs, is currently disabled in the GUI, and only works for evoked drives. By Sourav Rajvi in #1354, #1387, and #1395.

  • Cells can now use synaptic mechanisms other than NEURON’s Exp2Syn, which can be specified using new, optional parameter keys of the form <cell_type>_<receptor>_mechname (e.g. L2Pyr_nmda_mechname). This allows custom synaptic mechanisms (such as those used in the new Duecker model) to be used, by Austin E. Soplata in #1229. You can also set parameters specific to your custom mechanisms using HNN-Core, as long as those parameters are properly recognized by NEURON. This still relies on the old “legacy params” interface, so if you are interested in using this, please reach out to us on our Github Discussions page.

  • GUI buttons for saving simulated Networks or simulated dipole output data have been overhauled. In the GUI Simulation tab, each of the “Save Network and Drives” and “Save Simulation Output” buttons now have their own respective dropdowns for which simulation you want to save. This also adds the functionality to save PRIOR simulated networks from the GUI, which was not possible before! By Dylan Daniels in #1399.

Deprecations#

  • (Breaking change) BatchSimulate now expects the set_params callback to accept arguments in the order (net, param_values) instead of (param_values, net), aligning it with the order used by the Optimizer class, by Rahul Tripathi in #1208, fixing #1203. This was their first PR, thanks Rahul!

  • Please note that all simulations produced from this hnn_core version and going forward will be slightly numerically different than those of version v0.6.1 and prior. See note above about NEURON 9.0 support.

  • The hnn-core-all Anaconda (or conda) package has been pulled from distribution, and will no longer be developed. (Note: this is different from the pip install "hnn-core[all]" package which will always be supported.) Starting with this version, pip and uv will be the recommended ways to install hnn-core. Users who have previously installed the conda package using

    conda install hnn-core-all -c jonescompneurolab -c conda-forge
    

    are strongly recommended to remove the conda package and instead update to the latest hnn-core package using pip. If you installed using the instructions in our Installation Guide, you can do this upgrade with the following commands in your Terminal / Anaconda Prompt:

    conda activate hnn-core-env
    conda remove hnn-core-all
    pip install "hnn-core[all]"
    

    If you have any questions or concerns about updating from the discontinued conda package, reach out to us on our Github Discussions page.

  • Example scripts are no longer included in the hnn-core repository or on our API website (the website you are reading this on). All of the former examples/ scripts have been migrated to polished and updated Jupyter notebooks on our Textbook website, such as here. All references to them in these release notes now link to their corresponding Textbook notebooks. This also removes our use of sphinx-gallery, Binder, etc. By Austin E. Soplata in #1347.

Upcoming Deprecations#

  • There is now a new, preferred way for you to add the “canonical” ERP drives to your Neymotin 2020 network: Users are strongly advised to begin using the new argument:

    # use this new argument `add_erp_drives`
    neymotin_2020_model(add_erp_drives=True)
    

    INSTEAD OF either

    # deprecated
    neymotin_2020_model(add_drives_from_params=True)
    

    or

    # deprecated
    net = neymotin_2020_model()
    add_erp_drives_to_jones_model(net)
    

    Both the argument neymotin_2020_model(add_drives_from_params=True) and the function add_erp_drives_to_jones_model() will eventually be deprecated and have been marked as such.

    The reasoning for this was that the two deprecated ways to add these drives were slightly different (seed values were different), and that add_drives_from_params’s final values required reading a lot of code. You can see all parameters for the canonical drives easily here.

    The only exception to using the new neymotin_2020_model(add_erp_drives=True) argument is if you are still using legacy_mode=True. If you still need legacy_mode, then you must continue using add_drives_from_params, though legacy_mode is planned to be deprecated as well in the future.

    This same argument has been added to jones_2009_model(add_erp_drives=True) for backwards compatibility, though we recommend users switch from jones_2009_model() to neymotin_2020_model().

    Relatedly, the new duecker_ET_model also has an argument of the same name (duecker_ET_model(add_erp_drives=True)) for adding its canonical ERP drives, however be advised that its ERP drives are very different than that of neymotin_2020_model, and you can read about them at the Duecker preprint here. By Austin E. Soplata and Sourav Rajvi in #1340.

  • Network.set_cell_positions has been renamed to Network.update_cell_positions. Now, if you want to change where the cell positions are to new values, use Network.update_cell_positions, and if you instead want to reset the cell positions back to where they were before any Network.update_cell_positions calls, you can use Network._reset_to_original_cell_positions. By Katharina Duecker, Nick Tolley, and Austin E. Soplata in #1227.

  • The postproc argument of simulate_dipole() is now documented as deprecated. Please explicitly use the smooth() and scale() methods after your simulation instead, by Katharina Duecker in #1343.

Public API Changes#

  • simulate_dipole() now accepts a baseline_correction boolean argument (default True), which controls whether the baseline correction appropriate for the Network’s model is applied to the dipole after simulation, by Katharina Duecker and Austin E. Soplata in #1343. If you have always used the default behavior, then you do not need to take any action. The new flag is also supported by users of simulate_dipole such as BatchSimulate.

  • duecker_ET_model is now available via from hnn_core import duecker_ET_model (see “New Features” above).

  • pip install "hnn-core[parallel]" or pip install "hnn-core[all]" now install MPI support on Mac, Linux, and Windows-WSL automatically (see “New Features” above).

  • The dask[distributed] package can now be installed using pip install "hnn-core[dask]" for use with JoblibBackend’s dask backend, by M Yaswanth Reddy in #1248. It will also be installed with pip install "hnn-core[all]".

  • plot_spikes_raster() now supports a new argument gid_ranges, allowing users to specify that they want the spike trains of specific GIDs to be plotted, by Katharina Duecker in #1356.

  • Many API changes introduced in #1227:

    • Spike output saved to *.txt files should now always be sorted in time, even if MPI was used in the simulation.

    • For developers, there is a new Makefile recipe regenerate-duecker-test-data. This needs to be run by developers whenever there are any changes to either the network or drive parameters of our “canonical” Duecker model, in order to update the “ground truth” data that is used in testing.

    • Network objects now have a new model_variant attribute that is handled automatically throughout JSON files, etc. This is currently only used for discriminating Duecker-models from non-Duecker models (i.e. Neymotin 2020, etc.), but will be a key part of supporting other models in our future repertoire.

    • Cell.create_tonic_bias is now demoted to a private function Cell._create_tonic_bias, since it is what is used to actually create the biases in NEURON at simulation. Users who want to introduce a new bias to the network should continue to use Network.add_tonic_bias.

    • CellResponse now accepts a new argument cell_type_metadata, in case you want to provide your own custom metadata for your own custom celltypes. This is not recommended for general use, and is instead mostly used automatically when CellResponse objects are created after simulations. If you do decide to pass your own cell_type_metadata, see the top of network_models.py for an example that uses the default (Neymotin) network.

    • CellResponse now has new methods rate_over_time() and the related plot_firing_rate_time(). rate_over_time is similar to mean_rates but with a different API, and allows you to calculate firing rates for specific cells (or cell types) but using only a sliding window of time instead of only measuring across a single time segment. You can visualize its output with the new associated plot_firing_rate_time.

    • The Duecker model has introduced new celltypes in cells_default.py, including pyramidal_humanL5ET, pyramidal_humanL23, and human_gen_interneuron.

    • Optimizer now has a new objective function named dipole_rmse_corr, which calculates the “correlation-weighted RMSE” between dipoles, penalizing solutions with low correlation. See dipole.py’s _rmse_corr() for details on the calculation.

    • There are now several new MOD files available for use, all of which are used in the new Duecker model, such as mod/NMDA_gao2021.mod, etc. You can thus view the exact source code of the model mechanisms used in the new Duecker model, many of which are adapted from (Hay et al., 2011) (see the MOD files for publication references).

    • If you are creating raw Network objects from scratch (instead of using the recommended neymotin_2020_model or duecker_ET_model functions), and if you pass either a custom pos_dict or cell_types argument, you MUST pass your own version of the other argument. This is a small part of our plan to gradually reduce the amount of hardcoding of specific celltypes. Relatedly, if you do pass a custom pos_dict, then related variables like inplane_distance will be recalculated automatically based on your positions, instead of being hardcoded to constant values.

    • (Upcoming deprecation) Network.set_cell_positions has been renamed to Network.update_cell_positions. Now, if you want to change where the cell positions are to new values, use Network.update_cell_positions, and if you instead want to reset the cell positions back to where they were before any Network.update_cell_positions calls, you can use Network._reset_to_original_cell_positions.

    • Network.add_tonic_bias now accepts a new argument gid, which allows several different ways for you to specify which exact GIDs (or cells) you want a particular bias to connect to. See documentation here add_tonic_bias().

    • Network.plot_cells now accepts new arguments colors and markers, where you can map colors/markers to specific cell types.

    • CMA-ES optimization will now save (or overwrite) a cma_checkpoint.pkl file every 10th iteration of your optimization, in order to not lose your progress if it crashes. We plan to add this functionality to other optimization algorithms soon.

    • plot_spikes_raster now accepts more arguments that correspond to Matplotlib variables including xticks, yticks, xlabel, ylabel, and title. Internally, its color handling, sizing, and scaling have all been improved.

Bug Fixes#

  • The notorious MPITimeout errors should now be a thing of the past and no longer happen! Document and remove problematic MPI Timeouts (while keeping code intact), and implement a new tempfile-based mechanism for transmitting MPI data from child processes back to rank 0 child process then parent process, by Camilo Diaz in #1312 and #1315.

  • User-facing deprecation notices now emit FutureWarning instead of DeprecationWarning, making them visible under Python’s default warning filters, by William Kang in #1339. This was their first PR, thanks William!

  • Loading external drives in the GUI now extends the simulation duration when needed, preserving configured drive and tonic-bias stop times, by William Kang in #1348.

  • The RMSE and Corr comparison text on GUI dipole comparison plots is no longer drawn over the evoked drive arrows, and now stays at the bottom right of the plot, by Sourav Rajvi in #1378, fixing #1374.

  • Camilo Diaz did considerable work in fixing our long-standing MPI Timeout issues and implementing a permanent solution that uses tempfiles instead of relying on standard input/output/error streams in #1312. Thanks Camilo!

  • Reading in older Network JSON files (created before the v0 initial voltage value was introduced into the Network objects in #1185) no longer incorrectly initializes all of their membrane voltages to -65 mV, and instead uses each section’s correct initial voltage. Thanks to Anna Cattani for finding this bug, which was fixed by Austin E. Soplata in #1327. This means that most simulations that are run using hnn-core==0.6.1 end up being slightly different than the plots shown in our Textbook website, but this is now fixed.

  • Network model functions such as neymotin_2020_model() once again accept their params argument as a str or Path filename (which was accidentally removed in #1227), and all ways of passing params are now tested, by Austin E. Soplata in #1352.

  • plot_spikes_raster() no longer raises an error when one of the cell types does not spike, by Katharina Duecker in #1356.

  • Optimizer now ensures that the keys of initial_params and constraints are in the same order, which previously could cause the CMA-ES solver to raise an error, by Katharina Duecker in #1359.

  • Interactive plotting buttons in the GUI have been re-enabled by correctly detecting interactive notebook backends, by Tushar Jamdade in #1273, fixing #1125.

  • Errors that occur when running a simulation in the GUI are now printed at the top of the log window (with their tracebacks) instead of the bottom, by Dev Agarwal in #1264. This was their first PR, thanks Dev!

  • When a user loads a Network file in the GUI that has custom global synaptic gains, the GUI now displays those custom global gains (“smartly”) instead of resetting them to their default values, by Austin E. Soplata in #1183.

  • Fixed a failure to remove spectrogram colorbars in the GUI due to changes in Matplotlib 3.11, in by Austin E. Soplata and Claude in #1320.

  • Fixed caret coloring and pixel shifting when opening or closing the GUI log window, by Dylan Daniels in #1318.

  • Fixed stray scrollbars appearing on empty GUI dropdowns in Safari, by Sourav Rajvi in #1360, fixing #1321. This was their first PR, thanks Sourav!

  • Fixed broken behavior with “simulation data” / “experimental data” visualization dropdown bug, and refactor some related code, by Sourav Rajvi in #1394, fixing #1389.

  • Fixed crash from when add_spike_train_drive uses spikes that are not sorted, by Austin E. Soplata in #1400, identified by Sourav Rajvi.

Miscellaneous Changelog#

  • For the published stable 1.0.0 version, all user-facing dependencies now have maximum version caps, in order to prevent future dependency updates from breaking this release. Additionally, notebook is now a base dependency, so that anyone installing hnn-core can easily run our Textbook’s Jupyter notebooks using the standard jupyter notebook <file.ipynb> command, in #1385, fixing #1384.

  • Cell metadata is now used around the codebase much more cleanly. Additionally, each cell type’s cell_metadata now includes a default color and marker, which are used by our visualization functions (e.g. spike raster and histogram plots) instead of hardcoded values, by Chetan Kandpal in #1269.

  • Improvements to some visualization documentation, by Arnesh Banerjee in #1332. This was their first PR, thanks Arnesh!

  • Easier optimization custom function usage by Vaishnavi Baghel in #1271.

  • Improve y-axis labels (depth units) and ticks of LFP and CSD plotting, by Shivansh Bhageria in #1191. This was their first PR, thanks Shivansh!

  • The GUI’s data management has been completely overhauled. Simulation data and loaded experimental data are now kept separate and managed through a single, centralized data store object instead of being passed around between many functions, by Camilo Diaz in #1328.

  • os.path usage throughout the codebase has been migrated to pathlib, by Satvik Saluja in #1311.

  • The Network fixtures used in our tests have been expanded and consolidated so that all network models can be tested consistently, including the new Duecker model, by Satvik Saluja and Austin E. Soplata in #1257.

  • test_viz.py has been refactored into smaller, more focused tests, by Camilo Diaz in #1350.

  • Added “progressive” installation tests to our CI, which check that the minimal (no extras), then [opt], then [gui] installations each work before running the full test suite, in #1304.

  • Miscellaneous CI, documentation, testing/development config, and website maintenance in #1305, #1310, #1326, #1333, #1342, #1344, #1351, #1361, #1365, #1368, and #1372.

People who contributed to this release:#

0.6.1 Patch Release Notes#

This is an emergency patch release to fix an install misconfiguration (see #1302).

Recent changes to Optimizer from the new CMA optimization solver require joblib to be installed for any Optimization to be run. Since joblib is a required dependency of scikit-learn, installing with pip install "hnn-core[opt]" is sufficient, even though [opt] installs some, but not all, of HNN-Core’s [parallel] install extras, which are joblib and psutil.

Relatedly, since the GUI can now support optimization, its installation now requires the [opt] packages to be installed as well. This has been fixed.

For the greater release notes that document what has been added since 0.5.0, including important upcoming deprecations, see our Release Notes for 0.6.0 here.

0.6.0 Release Notes#

Important Deprecations#

  • jones_2009_model has been renamed to neymotin_2020_model to better reflect its provenance. From now on, neymotin_2020_model is the recommended “default” model we will be documenting and using going forward. However, for backwards compatibility, jones_2009_model will still be retained as a simple “wrapper function” around neymotin_2020_model, the latter being otherwise identical to the old jones_2009_model behavior. We recommend you update your scripts to use neymotin_2020_model, but jones_2009_model will still continue to work.

  • IMPORTANT: For the default model neymotin_2020_model (renamed from jones_2009_model, see above), in the NEXT release (i.e. 0.7.0), some parameters will be slightly changed, which will cause your simulations to produce slightly different output. Again, this is only expected in the next major release, not this one. This is necessary in order to fix certain parameters (e.g. synaptic delays) which are slightly different than their values in the “Original HNN” code. Using the new parameters, simulation results are expected to only be quantitatively different, but not qualitatively different, and these changes are not expected to change the significance of simulation results.

New Features#

  • When installing hnn-core through pip, there is now an easy way to install all extra dependencies (excluding development packages): pip install "hnn-core[all]". This installs the necessary dependencies for the GUI, Optimization, and parallelism (excluding MPI itself, which you can find how to install at our Installation Guide).

  • Major visual refresh and polishing of the GUI by Dylan Daniels in #1243.

  • Nick Tolley has added Optimization support for both the “Covariance matrix adaptation evolutionary strategy” (CMA-ES) solver and correlation-coefficient as an objective function, in #1221 and #1286! CMA-ES is a very powerful optimization option which can simulate multiple samples in a single epoch, and may possibly become the default optimization solver in the future.

  • The GUI now includes an Optimization tab that can run optimization against any parameter in the Network’s drives, using any of the available built-in solvers or objective functions, including the new CMA-ES solver (see above). By Austin E. Soplata and Nick Tolley in #1190.

Bug Fixes#

  • Accessibility of all HNN websites (excluding auto-generated Jupyter notebook output) has been fixed to prevent basic accessibility errors, by Joyce Gao and Austin E. Soplata in several PRs including #1280, #1293, https://github.com/jonescompneurolab/textbook/pull/153 , and https://github.com/GitHub-at-Brown/hnn-front-website/pull/56 .

  • The COBYLA Optimization solver now returns the Network with the best-performing parameters instead of simply returning the parameters of the final iteration that was run, by Carolina Fernandez Pujol, Nick Tolley, and Austin E. Soplata in #1240.

  • GUI dataset dropdowns have been fixed so that they show what is displayed in the plot, by Tushar Jamdade in #1223.

  • When loading a network using the GUI, sometimes there would arise a KeyError due to a lack of n_trials value. This has been fixed by Tushar Jamdade in #1216.

  • Read-in of Network JSON files with custom cell_types or pos_dict entries has been fixed in #1195.

Public API Changes#

  • jones_2009_model, the main function to spawn the default model, has been replaced with neymotin_2020_model (see note above in ‘Important Deprecations’). jones_2009_model will still continue to work, but users should migrate to the new name for the function. Done in #1290.

  • simulate_dipole now has a verbose flag to control how much output is printed, by Anna Cattani in #1228. This was their first PR, thanks Anna!

  • Section can now consume an v0 argument in order to set the initial voltage of the membrane potential for that section. Merged in #1185, adapted from work by Katharina Duecker.

  • BatchSimulate is now properly documented via our public API, thanks to M Yaswanth Reddy in #1218.

People who contributed to this release:#

Changelog#

  • Where possible, all remaining imports have been moved to the top of source files, by Percival Villalva in #1285.

  • In the GUI, when comparing two datasets, both the RMSE and the correlation values are now displayed on the plot instead of just the RMSE, by Nick Tolley and Austin E. Soplata in #1298.

  • Javascript and CSS files in the repository are now included in the Python packaging, in #1292.

  • Enable MacOS Intel-CPU runners in CI, by Vaishnavi Baghel in #1287.

  • Update and re-enable Codecov usage in CI, by Vaishnavi Baghel in #1272. This was their first PR, thanks Vaishnavi!

  • We have removed the Gitter badge on our repository’s homepage (in #1270) in order to push people towards the Github Discussions page.

  • Test coverage of Hamming smoothing via hnn_core.utils.smooth_waveform has been improved, by Satvik Saluja in #1266.

  • The issue-metrics Github Action workflow has been removed (in #1279) entirely in favor of https://github.com/jonescompneurolab/hnn-tracking/pull/1 .

  • Fix testing of mesh_shape to include all available network models, by Satvik Saluja in #1268. This was their first PR, thanks Satvik!

  • Improve UX of GUI’s “Delete all drives” button and small refactors of GUI codebase. By Muhammad Ahmad Amin in #1245. This was their first PR, thanks Muhammad!

  • In the GUI, when simulating multiple trials and plotting the spectrogram, use the averaged Time-Frequency-Representation of the dipoles for spectrogram calculation instead of only the first dipole, by M Yaswanth Reddy in #1233. This was their first PR, thanks Yaswanth!

  • The repository’s AI policy has been clarified in #1224.

  • Issue templates are now used in the hnn_core repository.

  • In the GUI, new simulations are now auto-renamed. This fixes an annoying issue with the GUI where every time a user wanted to run a new simulation, they had to go back to the Simulation tab and manually change the simulation name. Now, whenever they run a new simulation without changing the name manually, the new simulation will use the old name, but with a reasonable number appended. By Tushar Jamdade in #1213.

  • Add support for printing a welcome message that asks the user to fill out the HNN survey, which is printed to the standard output after (and only after) the first time the hnn_core module is imported. Whether or not the message is displayed is determined by the presence of a new file created in the module after first run. by Karthikeya Kodlai in #1158. This was their first PR, thanks Karthikeya!

  • Add progressive minimal install and test to Github Actions, by Tushar Jamdade in #1214. This was their first PR, thanks Tushar!


0.5.0 Release Notes#

New Features#

  • You can now provide custom spike trains as inputs into your simulation; see the description of add_spike_train_drive() below in the Public API Changes section.

Deprecations#

  • Similar to NEURON, we have dropped support for Python 3.8. We have also added support for Python 3.13.

Bug Fixes#

  • Synaptic gain and threshold values are now loaded correctly when reading in a Network JSON file, instead of being overridden with default values, by George Dang and Austin E. Soplata in #918.

Public API Changes#

  • Synaptic gains can now be set and retrieved for different “global” connection types (excitatory-to-excitatory, excitatory-to-inhibitory, inhibitory-to-excitatory, and inhibitory-to-inhibitory) using set_global_synaptic_gains() (formerly called Network.update_weights) and get_global_synaptic_gains(). The GUI now includes widgets for modifying synaptic gains in the Connectivity interface, both on a “global” level and on an individual level. The additive combination of global and individual gains is also always shown in the GUI. Importantly, all of the synaptic gain getters and setters across both the GUI and API do not change or reflect the gains of drives. They only change or reflect non-drive connections.

  • add_connection() now accepts optional threshold and gain arguments for setting connection-specific firing thresholds and synaptic gain values.

  • Optimizer now accepts user-defined initial weights in the form of optional argument initial_params and allows the user to set the number of trials by passing n_trials to fit(), by Carolina Fernandez Pujol in #1057.

  • The cell_types attribute and initializer argument of Network has significantly changed. Previously, Network.cell_types was a dictionary where keys were cell-type names and values were Cell instances of that cell-type to be used as templates for building the Network. Now, the structure of dictionary is different: Network.cell_types is a dictionary where keys are cell-type names and values are child-dictionaries. The child-dictionaries of each cell-type have the following keys: cell_object and cell_metadata. The value of Network.cell_types[<cell_type>]["cell_object"] is the Cell instance for that cell-type, similar to before. The value of Network.cell_types[<cell_type>]["cell_metadata"] is a new dictionary containing several key-values pairs that describe different aspects of the cell type, described below:

    • morpho_type : either basket or pyramidal

    • electro_type : either inhibitory or excitatory

    • layer : either 2 or 5

    • measure_dipole : either True or False

    • reference: optional

    You can find a working version of Network.cell_types for the default Jones 2009 model here as an example: jonescompneurolab/hnn-core . The author for this was Chetan Kandpal in #1099. Others who helped with design were Dylan Daniels, Katharina Duecker, Austin E. Soplata, and Nick Tolley.

  • The newer Network object and file input/output module, hnn_io, is now published as part of the Public API. The recommended way to write a Network object is write_network_configuration(), and the recommended way to read the same is read_network_configuration(). By Austin E. Soplata in #1108.

  • Network now has support for calculating a “layer dictionary” layer_dict attribute whose layers can be mapped as desired to an object’s cell_types attribute. This mapping can then be used to make a more flexible position dictionary pos_dict attribute of the object. This also adds support for optional pos_dict and cell_types arguments to the Network constructor. This also begins the process of moving some model-specific celltype-usage from network.py to network_models.py. By Chetan Kandpal in #1095.

  • Network now has a new type of available “external drive”, the “spike train” drive, which you can add using add_spike_train_drive(). This new drive can take in spike data in various formats (see its docstring), including spikes generated by previous HNN simulations, and then “replay” the spike trains as inputs into the current simulation. This offers a large degree of control over both the data of the spike trains and how you want to connect them to your simulation. There is a new example script here which illustrates its use. By Maira Usman in #1064.

People who contributed to this release:#

Changelog#

  • Add methods set_global_synaptic_gains() and get_global_synaptic_gains() to modify and retrieve synaptic gains for different cell type connections in the network. Add GUI support for modifying synaptic gains with interactive widgets in the connectivity interface, for both individual connections and “global” connections (for example, “excitatory-to-excitatory”). Add gain and threshold parameters to add_connection() to allow connection-specific control of synaptic gains and firing thresholds. There is also now a warning displayed when a user loads a Network JSON file that has inconsistent synaptic gains, since both set_global_synaptic_gains() and get_global_synaptic_gains() assume that gains are uniformly equal within each “global” connection type. By George Dang and Austin E. Soplata in #918.

  • Enhance the optimization workflow for more reliable parameter fitting. This includes adding support for user-defined initial weights. For optimal results, we now recommend setting initial_params to hand-tuned values since they often provide a good starting fit. The default objective function (_rmse_evoked()) now averages results over 3 dipoles rather than simulating a single dipole, with the option to control this behavior by setting the n_trials parameter in fit(). To capture the model’s average behavior, it is recommended to set n_trials > 1, as using n_trials=1 may identify parameters that work well for one simulation run but perform poorly on average. The optimization example has also been enhanced with improved markdown and updated contents to better illustrate best practices. By Carolina Fernandez Pujol in #1057.

  • Tiny bugfix for the #1099 to enable backwards-compatibility, by Austin E. Soplata in #1119.

  • Major upgrade to celltype metadata and cell_types in Network, see the description above in Public API Changes for details, By Chetan Kandpal in #1099.

  • Added hnn_io to the Public API by Austin E. Soplata in #1108.

  • Some time allowances for “MPI Timeout” errors were doubled. The infamous “MPI Timeout Error” should occur less frequently now, though this is only a temporary, incomplete solution to #774. By Austin E. Soplata in #1102.

  • Reduction of old “short name vs long name” usage (Chetan Chunk 1.5), By Chetan Kandpal in #1103.

  • Add support for Network to calculate a more flexible “layer dictionary” that can be used to construct your Network’s pos_dict, and add its use in our existing network_models.py. This is the first in a series of code changes meant to allow for more flexible cell types to be used. By Chetan Kandpal in #1095.

  • Introduce new “spike_train” drive to Network by Maira Usman in #1064.

  • Various development-process and documentation changes by Austin E. Soplata in #1081, #1086, #1090, #1083, #1069, #1107, #1100, #1114, and #1126.

  • Drop Python 3.8 support, add Python 3.13 support by Austin E. Soplata in #1070.

0.4.3 Patch Notes#

This is an emergency patch release to fix a dependency issue.

Bug Fixes#

  • NEURON version support is restricted to versions 7.7 through 8.2.7 (including those versions). Going forward, the maximum NEURON version supported will be set explicitly in our install configuration, so that new, incompatible versions of NEURON are not used before testing. by Austin E. Soplata in #1153.

People who contributed to this release:#

0.4.2 Patch Notes#

Version 0.4.2 is a patch release that fixes two important bugs: one affecting LFP and CSD output (#1062) and another preventing users from upgrading to the recently released NEURON 8.2.7 (#1058), which also enables use of HNN-Core with Python 3.13. This also includes other, smaller bug fixes and changes.

If you have been using HNN-Core to simulate LFP or CSD output, we strongly recommend you update to this version.

We have also recently overhauled our HNN Frontpage and our new HNN Textbook website, so check them out!

New Features#

  • We now provide conda packages that offer significantly easier installation of HNN-Core, including an option with all features included (including MPI!). See our Installation Guide for how to install them.

  • We now support Python 3.13, in addition to previously supported Python versions.

Upcoming Deprecations#

  • In the next minor release (0.5.x), we will be dropping Python 3.8 support for two reasons:

    1. It has reached “end-of-life” and is no longer supported, including NOT receiving any future security upgrades (see https://devguide.python.org/versions/).

    2. NEURON 8.2.7 has dropped support for Python 3.8.

If you expect that you will still require hnn_core using Python 3.8, please reach out to us on our Github Discussions page. We will likely be able to continue supporting your use-case.

Bug Fixes#

  • Major bugfix: Cell positions were previously starting at origin (incorrect) instead of origin + their pos_dict locations (correct). This has been fixed. This change should not affect results from dipole or single-cell output, but will affect results from LFP or CSD output. Identified and fixed by Katharina Duecker in #1062.

  • hnn_core/viz.py::plot_spikes_raster() now correctly titles the plot according to if dipoles are overlaid or not, by Dylan Daniels in #1074.

  • Update hnn_core/mod/vecevent.mod to a newer version from NEURON’s source code, enabling use of NEURON’s latest version (8.2.7) and Python’s latest stable version (3.13). The prior version of this file would not correctly compile on Windows using NEURON 8.2.7. By Austin E. Soplata in #1058.

  • Fix test breakage caused by upstream dependency change in scipy.optimize, by Austin E. Soplata in #1076.

Development Workflow Changes#

  • hnn_core code now follows the default ruff code formatting style, and all new contributions are expected to follow this style as well. Fortunately, there is a single command that developers need to run which automatically format their code appropriately (make format-overwrite), see the new Quality Control section of our Contributing Guide for details.

  • Updates to Linkcheck Workflow frequency so it only runs on every merge to master and on the 1st and 15th of every month, by Maira Usman in #1043 (new contributor!!!).

  • Concurrency in Github Actions is now utilized: if you push a new commit to your existing PR branch while prior workflows are still running, the prior workflows are canceled (#1015).

  • You can now build and deploy new package distributions of hnn_core directly from Github to Pypi and TestPypi by pushed any new tag that begins with v (#1073 and #1077). Documentation for how to make a release has not been updated yet, but will be soon.

  • When a developer needs to update parameters of the “default model” used in HNN, there is now both documentation and commands to help guide them. See the new “Making changes to the default network” section of our Contributing Guide (#1068).

API Changes#

  • :class:~hnn_core.CellResponse now requires a cell_type_names argument, whereas before the argument was optional (#970).

  • hnn_core/optimization/optimized_evoked.py::optimized_evoked now enforces a minimum number of iterations, twelve (12). This is a temporary, possibly permanent, fix introduced in #1076.

People who contributed to this release (in alphabetical order of family name):#

Changelog#

  • Fix tiny typo in pip MPI install documentation, by Austin E. Soplata in #1072. (he’s also responsible for the original typo…)

  • Update the Linux path of our Unix unit tests to use conda for its MPI installation, NOT system packages. This makes our unit tests consistent with our recently-changed recommended installation pathways for pip MPI on Linux. By Austin E. Soplata in #1071.

  • Tiny docstring update to jones_2009_model that the “model” is from the Jones 2009 paper, but the “implementation” is from Neymotin 2020. Citations for both papers are included in the docstring. By Austin E. Soplata in #1056.

  • Added support for “regeneration” of both required hierarchical JSON network files alongside documentation, by Austin E. Soplata in #1068.

  • Enable automated package release building and deployment via Github Actions to Pypi and TestPypi, by Austin E. Soplata in #1073 and hotfix in #1077.

  • Version downgraded from 0.5.0dev0 to 0.4.2.devX in preparation for this release (no PR).

  • Fix test breakage caused by upstream dependency change in scipy.optimize, by Austin E. Soplata in #1076.

  • hnn_core/viz.py::plot_spikes_raster() now correctly titles the plot according to if dipoles are overlaid or not, by Dylan Daniels in #1074.

  • Update hnn_core/mod/vecevent.mod to a newer version from NEURON’s source code, enabling use of NEURON’s latest version (8.2.7) and Python’s latest stable version (3.13), by Austin E. Soplata in #1058.

  • Introduce concurrency controls to Github Actions, by Maira Usman in #1015.

  • Major bugfix: Cell positions were previously starting at origin (incorrect) instead of origin + their pos_dict locations (correct). This has been fixed. This change should not affect results from dipole or single-cell output, but will affect results from LFP or CSD output. Identified and fixed by Katharina Duecker in #1062.

  • Updates to Linkcheck Workflow frequency so it only runs on every merge to master and on the 1st and 15th of every month, by Maira Usman in #1043 (new contributor!!!).

  • Miscellaneous link updates/fixes needed due to our overhaul of our HNN Frontpage and our new HNN Textbook website, by Austin E. Soplata in #1055, #1059, #1063, and #1067.

  • Add support for ruff format, merge changes taken from running it, and require its use for new code contributions, by Austin E. Soplata in #934.

  • Add conda package installation instructions, by Austin E. Soplata in #1051.

  • Add and update “Cloud” usage instructions on install page to redirect to our new Textbook website, by Austin E. Soplata in #1046 and #1048.

  • Remove hardcoding of celltypes in :class:~hnn_core.CellResponse and add rename_cell(), by Mohamed W. ElSayed in #702 and #970.

0.4.1 Patch Notes#

  • Version 0.4.1 is a bug-fixing patch release for version 0.4. This includes changes to importing of BatchSimulate due to previously-undetected install/import issues (#1034), configuration of packaging metadata format (same PR), and elimination of a discrepancy in our method of cleaning local compiled files that led to architecture-specific files being included in the Pypi 0.4 release, which caused simulations on some platforms to fail (#1035). The public Pypi version has already been updated to 0.4.1.

0.4 Release Notes#

v0.4 represents a major milestone in development of hnn_core and the HNN ecosystem as a whole. v0.4 includes over two years of active development work by many people (>800 commits!), and brings with it many new and exciting features, including significant improvements to robustness, testing, and bug-fixing.

New Features#

  • hnn_core now includes a fully-tested and robust GUI of its own. The hnn_core GUI was present as a prototype in v0.3, but it is now ready for production. New features and visual improvements will still be coming to it in the future, such as the ability to use optimization. See our new Install page for ways to install it, and we have already begun incorporating it into a new, fresh series of tutorials for our upcoming revamp of the HNN website. If you have installed it, you can start the GUI using hnn-gui in your terminal/command prompt window.

  • The BatchSimulate class: Thanks to Abdul Samad Siddiqui and Google Summer of Code 2024, there is now the capability to run “batches” of simulations across multiple parameter sets, enabling easy analysis and simulation of behavior across parameter sweeps. See our example for more details. Note that currently, only its loky backend is supported, and the "hnn-core[parallel]" dependencies must be installed for it to be used.

  • Significant improvements to the API, documentation, and pedagogical examples especially for Optimization, among others.

  • Calcium concentration can now be recorded: recorded calcium concentration from either the soma, or all sections, are enabled by setting record_ca to soma or all in simulate_dipole(). Recordings are accessed through ca.

  • There is now a new class NetworkPlotter which can be used to visualize an entire network in 3D, including firing animations; see our example of how to use it here.

  • There is now a new function plot_drive_strength() for illustrating the absolute or relative amount of strength that a particular drive provides to different cell types.

  • A very large amount of polishing, bug fixes, general improvements, etc.

Deprecations#

  • The new Python 3.13 is not supported by hnn_core at this time, due to NEURON’s current lack of support for it. This will change in the near future. We still support 3.8 through 3.12 (inclusively).

Upcoming Deprecations#

  • Both plot_laminar_lfp() and plot_dipole() will have their tmin and tmax arguments removed in the future. Please set the x-axis limits using methods called directly on the existing matplotlib objects, or using matplotlib.pyplot.xlim.

  • Network’s argument of legacy_mode for importing old param files will be removed in the future.

  • add_tonic_bias()’s argument of cell_type, along with setting the argument amplitude to a single float, will be removed in the future. Instead, set the amplitude argument to a dictionary as described in the docstring.

  • simulate_dipole()’s argument of postproc for post-processing will be removed in the future. Instead, use explicit smoothing and scaling via Dipole methods.

API Changes#

  • New argument to Network initialization: you can now set mesh_shape to easily make a grid of different sizes of Networks.

  • Cell initialization argument topology has had both its name changed to cell_tree and its data type significantly changed; see the API docs of Cell for details.

  • jones_2009_model() and other built-in Network Models including law_2021_model() and calcium_model() all accept the aforementioned mesh_shape argument like Network.

  • The API for optimization has changed significantly. Instead of running the function optimize_evoked obtained using from hnn_core.optimization import optimize_evoked, you should use the new Optimizer class and its methods; see our example of evoked-response optimization here.

  • plot_spikes_hist() now accepts more arguments, including invert_spike_types, color, and any **kwargs_hist which can be applied to matplotlib.axes.Axes.hist. See the docstring for details.

  • plot_spikes_raster() now accepts many more arguments, including cell_types, colors, show_legend, marker_size, dpl, and overlay_dipoles. See the docstring for details.

  • plot_cell_morphology() now accepts more arguments, including color and several arguments related to its position and viewing window, including pos, xlim, ylim, and zlim. See the docstring for details.

  • plot_laminar_csd() now accepts more arguments, including vmin, vmax, sink, and interpolation. See the docstring for details.

  • MPIBackend now accepts many more arguments, including use_hwthreading_if_found, sensible_default_cores, override_hwthreading_option, and override_oversubscribe_option. See the docstring for details; the ability to customize it has been greatly increased.

  • read_params() now accepts a new argument file_contents which lets you pass in network configuration contents using a string.

People who contributed to this release (in alphabetical order of family name):#

PRs merged (API)#

PRs merged (Bug fixes and corrections)#

PRs merged (GUI changes)#

  • Add RMSE calculation and plotting to GUI, by Huzi Cheng in #636

  • Update GUI to use ipywidgets v8.0.0+ API, by George Dang in #696

  • Add GUI visualization testing, by Abdul Samad Siddiqui in #726

  • Add pre defined plot sets for simulated data in GUI, by Camilo Diaz in #746

  • Add GUI widget to enable/disable synchronous input in simulations, by Camilo Diaz in #750

  • Add GUI widgets to save simulation as csv and updated the file upload to support csv data, by Camilo Diaz in #753

  • Refactor GUI tests, by George Dang in #765

  • Refactor GUI import of _read_dipole_text function, by George Dang in #771

  • Add GUI feature to include Tonic input drives in simulations, by Camilo Diaz #773

  • Add GUI feature to read and modify cell parameters, by Camilo Diaz in #806

  • Refactor GUI read_network_configuration, by George Dang in #833

  • Change the configuration/parameter file format support of the GUI. Loading of connectivity and drives use a new multi-level json structure that mirrors the structure of the Network object. Flat parameter and json configuration files are no longer supported by the GUI, by George Dang in #837

  • GUI load confirmation message, by George Dang in #846

  • Differentiate L5/L2 Pyr geometry options in GUI, by Nick Tolley in #848

  • Updated the GUI load drive widget to be able to load tonic biases from a network configuration file. George Dang in #852

  • Update GUI initialization of network, by George Dang in #853

  • Update GUI color, by Nick Tolley in #855

  • Added “No. Drive Cells” input widget to the GUI and changed the “Synchronous Input” checkbox to “Cell-Specific” to align with the API George Dang in #861

  • Add GUI export of configurations, George Dang in #862

  • Add screenshot of GUI to README, George Dang in #865 and #866

  • Add button to delete a single drive on GUI drive windows, by George Dang in #890

  • Add post-processing for GUI figures, by George Dang in #893

  • Add minimum spectral frequency widget to GUI for adjusting spectrogram frequency axis, by George Dang in #894

  • Update GUI to display “L2/3”, by Austin E. Soplata in #904

  • Update PSD plot in GUI to use plot config provided frequencies instead of hard-coded values, by Dylan Daniels in #914

  • Flip drives in input histogram based on position in GUI by Dylan Daniels in #923

  • Add GUI widget to adjust default smoothing value, by Dylan Daniels in #924

  • Change Morlet cycles divisor for better alpha spectral plotting by Austin E. Soplata in #928

  • Add GUI log error message if spectral arguments are invalid by Austin E. Soplata in #944

  • Move GUI log messages to bottom of output by George Dang in #946

  • Add GUI frequency default visualization parameters and many other smaller visual changes by Dylan Daniels in #952

  • Capture printed messages to logger in GUI by Dylan Daniels in #956

  • Correctly set layer-specific dipole axes limits in GUI, by Dylan Daniels in #1022

  • Improve spike raster plot by overlaying dipoles, by Dylan Daniels in #1026

PRs merged (Other)#

0.3#

Changelog#

Bug#

  • Fix bugs in drives API to enable: rate constant argument as float; evoked drive with connection probability, by Nick Tolley in #458

  • Allow regular strings as filenames in write() by Mainak Jas in #456.

  • Fix to make network output independent of the order in which drives are added to the network by making the seed of the random process generating spike times in drives use the offset of the gid with respect to the first gid in the population by Mainak Jas in #462.

  • Negative event_seed is no longer allowed by Mainak Jas in #462.

  • Evoked drive optimization no longer assigns a default timing sigma value to a drive if it is not already specified, by Ryan Thorpe in #446.

  • Subsets of trials can be indexed when using plot_spikes_raster() and plot_spikes_hist(), by Nick Tolley in #472.

  • Add option to plot the averaged dipole in plot_dipole() when dpl is a list of dipoles, by Huzi Cheng in #475.

  • Fix bug where plot_morphology() did not accurately reflect the shape of the cell being simulated, by Nick Tolley in #481

  • Fix bug where pick_connection() did not return an empty list when searching non existing connections, by Nick Tolley in #515

  • Fix bug in MPIBackend that caused an MPI runtime error (RuntimeError: MPI simulation failed. Return code: 143), when running a simulation with an oversubscribed MPI session on a reduced network, by Ryan Thorpe in #545.

  • Fix bug where pick_connection() failed when searching for connections with a list of cell types, by Nick Tolley in #559

  • Fix bug where add_evoked_drive() failed when adding a drive with just NMDA weights, by Nick Tolley in #611

  • Fix bug where read_params() failed to create a network when legacy mode is False, by Nick Tolley in #614

  • Fix bug where plot_dipole() failed to check the instance type of Dipole, by Rajat Partani in #606

API#

  • Optimization of the evoked drives can be conducted on any Network template model by passing a Network instance directly into optimize_evoked(). Simulations run during optimization can now consist of multiple trials over which the simulated dipole is averaged, by Ryan Thorpe in #446.

  • plot_dipole() now supports separate visualizations of different layers, by Huzi Cheng in #479.

  • Current source density (CSD) can now be calculated with calculate_csd2d() and plotted with plot_csd(). The method for plotting local field potential (LFP) is now found at plot_lfp(), by Steven Brandt and Ryan Thorpe in #517.

  • Recorded voltages/currents from the soma, as well as all sections, are enabled by setting either record_vsec or record_isec to 'all' or 'soma' in simulate_dipole(). Recordings are now accessed through vsec and isec, by Nick Tolley in #502.

  • legacy_mode is now set to False by default in all for all Network objects, by Nick Tolley and Ryan Thorpe in #619.

People who contributed to this release (in alphabetical order):#

0.2#

Notable Changes#

  • Local field potentials can now be recorded during simulations Example

  • Ability to optimize parameters to reproduce event related potentials from real data Example

  • Published models using HNN were added and can be loaded via dedicated functions

  • Several improvements enabling easy modification of connectivity and cell properties Example

  • Improved visualization including spectral analysis, connectivity, and cell morphology

Changelog#

Bug#

API#

  • New API for defining cell-cell connections. Custom connections can be added with add_connection(), by Nick Tolley in #276

  • Remove L2Pyr, L5Pyr, L2Basket, and L5Basket classes in favor of instantiation through functions and a more consistent Cell class by Mainak Jas in #322

  • Remove parameter distribution in add_bursty_drive(). The distribution is now Gaussian by default, by Mainak Jas in #330

  • New API for accessing and modifying Cell attributes (e.g., synapse and biophysics parameters) as cells are now instantiated from template cells specified in a Network instance’s cell_types attribute by Ryan Thorpe in #321

  • New API for network creation. The default network is now created with net = jones_2009_model(params), by Nick Tolley in #318

  • Replace parameter T with tstop in add_tonic_bias() and create_tonic_bias() to be more consistent with other functions and improve readability, by Kenneth Loi in #354

  • Deprecated postproc argument in simulate_dipole(), whereby user should explicitly smooth and scale resulting dipoles, by Christopher Bailey in #372

  • Number of drive cells and their connectivity can now be specified through the n_drive_cells and cell_specific arguments in Network.add_xxx_drive() methods, replacing use of repeats and sync_within_trial, by Ryan Thorpe in #383

  • Simulation end time and integration time have to be specified now with tstop and dt in simulate_dipole(), by Mainak Jas in #397

  • CellResponse.reset() method is not supported any more, by Mainak Jas in #397

  • Target cell types and their connections are created for each drive according to the synaptic weight and delay dictionaries assigned in Network.add_xxx_drive(), by Ryan Thorpe in #369

  • Cell objects can no longer be accessed from Network as the cells attribute has been removed, by Ryan Thorpe in #436

People who contributed to this release (in alphabetical order):#

0.1#

Changelog#

Bug#

  • Fix missing autapses in network construction, by Mainak Jas in #50

  • Fix rhythmic input feed, by Ryan Thorpe in #98

  • Fix bug introduced into rhythmic input feed and add test, by Christopher Bailey in #102

  • Fix bug in amplitude of delay (for connection between L2 Basket and Gaussian feed) being passed incorrectly, by Mainak Jas in #146

  • Connections now cannot be removed by setting the weights to 0., by Mainak Jas and Ryan Thorpe in #162

  • MPI and Joblib backends now apply jitter across multiple trials identically, by Ryan Thorpe in #171

  • Fix bug in Poisson input where the first spike was being missed after the start time, by Mainak Jas in #204

  • Fix bug in network to add empty spike when empty file is read in, by Samika Kanekar and Ryan Thorpe in #207

API#

  • Make a context manager for Network class, by Mainak Jas and Blake Caldwell in #86

  • Create Spikes class, add write methods and read functions for Spikes and Dipole classes, by Ryan Thorpe in #96

  • Only specify n_jobs when instantiating the JoblibBackend, by Blake Caldwell in #79

  • Make a context manager for parallel backends (JoblibBackend, MPIBackend), by Blake Caldwell in #79

  • Add average_dipoles() function, by Blake Caldwell in #156

  • New API for defining external drives and biases to network. By default, a Network is created without drives, which are added using class methods. The argument add_drives_from_params controls this behavior, by Christopher Bailey in #221

  • Examples apply random state seeds that reproduce the output of HNN GUI documentation, by Christopher Bailey in #221

  • Force conversion to nAm (from fAm) for output of simulate_dipole() regardless of postproc-argument, which now only controls parameter file-based smoothing and scaling, by Christopher Bailey in #264

People who contributed to this release (in alphabetical order):#