Added

• Added groups parameter to nengo.Convolution. (#1675, #1684)

Changed

• Made NengoCore available under the GPLv2 license. (#1693)

Fixed

• Fixed an issue where nengo.LinearFilter and subclasses (e.g. Lowpass, Alpha) would fail when running on tensors with dimension >= 3. (#1687)

Added

• Added official support for Python 3.9. (#1660)
• Added ChannelShape.from_space_and_channels to easily construct a ChannelShape from a spatial shape and number of channels. (#1648)
• Added the ConvolutionTranspose transform to perform transposed convolution. It is commonly used for various forms of upsampling in deep networks. (#1648)
• Added Conv and ConvTranspose aliases for Convolution and ConvolutionTranspose. (#1648)

Changed

• The minimum supported NumPy version is now 1.19, as earlier versions are no longer officially supported. (NEP-29, #1683)

Removed

• Removed support for Python 3.5 (which reached its end of life in September 2020). (#1649)
• Removed nengo.utils.graphs.graph (this was a small utility function for building graphs that was only used in tests). (#1654)
• Removed simulator.ProbeDict alias; this was previously renamed to simulator.SimulationData. (#1649)

Fixed

• Fixed a bug with a problematic cache index breaking decoder solvers. The solver now avoids using the cache, rather than crashing. (#1649)
• Operator graph step order will now be deterministic. (#1654)
• Fixed an issue in which some simulators could not be reset due to signals not being marked as readonly. (#1676)
• Fixed an inconsistency in which normal Node output functions would receive a copy of the input signal, while Process step functions would not. Process step functions now also receive copies. (#1679)
• Duplicate keys in Neurons.probeable have been removed. (#1681)

Added

• Added a new example notebook for Legendre Memory Units. (#1589)
• Added the step_order attribute to nengo.Simulator, which contains an ordered list of the operations run on each timestep. (#1615)
• Added the make_state method to NeuronType, which initializes the neuron type's state variables. (#1609)
• Added the spiking attribute to NeuronType, which exposes whether a neuron type is spiking or non-spiking. (#1609)
• Added the negative attribute to NeuronType, which indicates whether the neuron type can have negative outputs. (#1609)
• Added the Tanh neuron type to simulate hyperbolic tangent neurons. (#1609)
• Added the RatesToSpikesNeuronType, which is a base class for neuron types that convert a rate-based type to a spiking one. (#1609)
• Added the RegularSpiking neuron type, which emits regularly-spaced spikes at the rate specified by its base type. (#1609)
• Added the StochasticSpiking neuron type, which emits spikes based on stochastic rounding to roughly match the rate specified by its base type. (#1609)
• Added the PoissonSpiking neuron type, which emits Poisson-distributed spikes, as are commonly used to match biological spiking statistics. (#1609)
• Added the PositiveNeuronType test argument to run tests on all neuron types for which negative is not True. (#1609)
• Added the QuasirandomSequence distribution, which is similar to Uniform but spreads points across the space evenly. (#1611)
• Added the ScatteredHypersphere distribution, which is similar to UniformHypersphere but spreads points across the space more evenly. (#1611)
• Added the RLS (recursive least-squares) learning rule, which is an online version of the least-squares method typically used for offline decoder-solving. (#1611, example)
• Added the SimProbe operator, which marks a signal as being probed. (#1653)

Changed

• Nengo is now compatible with Python 3.8. (#1628)
• The default Connection transform is now None, meaning that there will be no transform applied. This only changes behavior when learning on a neuron-neuron connection with the default scalar transform. In that situation there are now no weights to apply learning to, so this will result in an error. The old behaviour can be obtained by setting transform=1. (#1591)
• Network list attributes (e.g. .ensembles, .connections, .probes) are now read-only, to prevent users from accidentally overwriting them with their own data. (#1545, #1608)
• The NeuronType.step_math method has been renamed to NeuronType.step. (#1609)
• Neuron types can now create arbitrary state variables without needing to register a new build function. The state class attribute declares the neuron type's state variables and their default initial values. All __init__ methods accept an initial_state dictionary for users to override the default initial state values. (#1609)
• The nl and nl_nodirect test arguments have been renamed to AnyNeuronType and NonDirectNeuronType. (#1609)
• Weight solvers (i.e. those with weights=True) are now allowed on all connections. For connections that are not between Ensembles, though, weight solvers have the same effects as solvers with weights=False, and a warning will be raised. (#1626)
• Various improvements to simulation speed. (#1629)
• EnsembleArray now raises an error if add_output would overwrite an existing attribute. (#1611)
• The encoders and eval_points of Ensemble are now sampled from ScatteredHypersphere by default. (#1611)

Deprecated

• NeuronType.step replaces the NeuronType.step_math method, which will be removed in Nengo 4.0.0. (#1609)
• Connection.is_decoded is deprecated, as the definition of whether a Connection is decoded or not was ambiguous. Instead we recommend directly checking the pre/post objects for the properties of interest. (#1640)

Fixed

• Fixed a bug when comparing equality with Ensemble.neurons or Connection.learning_rule objects. (#1588)
• Fixed a bug preventing unpickling an Ensemble. (#1598)
• Fixed a bug in which unpickling a Simulator would rerun the optimizer. (#1598)
• Fixed a bug where the LstsqDrop solver errored when solving for zero weights. (#1541, #1607)
• Fixed a bug in the validation of Choice distributions. (#1630)
• Fixed a bug where a Signal did not register as sharing memory with itself. (#1627)
• Fixed a shape error when applying PES learning to a neuron-to-neuron connection with a slice on the post-synaptic neurons. (#1640)
• Fixed a shape error when applying PES learning to a neuron->ensemble connection with a weight solver. (#1640)
• Fixed a shape error when applying PES learning to an ensemble->neuron connection. (#1640)
• Fixed a shape error when applying PES learning with a slice on the pre-synaptic object. (#1640)

Added

• Added progress bar support for Jupyter Lab >=0.32. (#1428, #1087)
• We now warn that the progress bar is not supported in Jupyter Notebook <5. (#1428, #1426)
• Added support for convolutional connections. (#1481)
• Added version tracking to documentation, so that documentation from old versions remains available. (#1488)
• Added support for sparse connections. (#1532)
• Added a fail_fast setting to test operators when they are first added to the model. See configuration options for details. (#1532)
• Added a --memory option for pytest that prints the total memory consumed by the tests when they complete (Linux and Mac OS X only). (#640)
• Added a bit precision setting to change the number of bits allocated to each value tracked by Nengo. (#640)
• Added a Simulator.clear_probes method to clear probe data. This method can be used before pickling to reduce the pickle file size. (#1387)
• Nengo tests now use the allclose fixture from pytest-allclose, which makes it possible for backends to change test tolerances. (#1563)
• Nengo tests now use the rng and seed fixtures from pytest-rng. (#1566)
• Nengo tests now use the plt fixture from pytest-plt. (#1566)
• Added a nengo_simloader pytest option for specifying a callable that takes a pytest request and returns a callable to be used as Simulator in the Nengo test suite. (#1566)
• Added more content to the API reference documentation. (#1578)

Changed

• Python 2 is no longer supported. The oldest supported Python version is 3.5. (#1520, python3statement.org)
• Nengo no longer supports Python 3.4. Official 3.4 support ended in March 2019. (PEP-429, #1514)
• Replaced the dt argument to Simulator.trange with sample_every because dt would return values that the simulator had not simulated. dt is now an alias for sample_every and will be removed in the future. (#1368, #1384)
• Dense connection transforms (this includes all previously supported values for Connection.transform) will now be represented internally as nengo.Dense objects. Arrays/scalars can still be passed as transform values, and they will be automatically converted to the equivalent nengo.Dense object. Retrieving the value of my_conn.transform will return that Dense object. The original input array can be retrieved through my_conn.transform.init. (#1481)
• nengo.solvers.NoSolver(w, weights=True) now expects w to have shape (pre.n_neurons, function_d), rather than pre.n_neurons, post.n_neurons). That is, with NoSolver you are always specifying the values for the decoders, and encoders/transform will be applied automatically to those decoders (as occurs with all other solvers). Note that this does not affect NoSolver(..., weights=False) (the default). (#1481)
• Increased minimum NumPy version to 1.11.0. See our instructions for installing NumPy if you need to upgrade. (#1481)
• Solvers are now explicitly marked as compositional or non-compositional depending on whether they must act on full connection weight matrices when solving for weights. (#1507)
• Solvers no longer take encoders as an argument. Instead, encoders will be applied to the targets before the solve function for non-compositional solvers and applied by the Transform builder for compositional solvers. (#1507)
• Example Jupyter notebooks have been upgraded to notebook format 4. (#1440)
• Switched documentation to new nengo-sphinx-theme. (#1489)
• The settled_firingrate function has been moved from nengo.utils.neurons to nengo.neurons. (#1187)
• Added new pytest config option, nengo_test_unsupported (replacing the previous Simulator.unsupported functionality). (#1521)
• Switched to nengo-bones templating system for TravisCI config/scripts. (#1514)
• The NeuronType.current and NeuronType.rates methods now document the supported shapes of parameters and return values. (#1437)
• PES learning updates are now applied on the next timestep rather than the current one. (#1398)
• The NdarrayParam now accepts a dtype argument to check that data assigned to that parameter matches the given Numpy dtype. DistOrArrayParam accepts an analogous sample_dtype argument. (#1532)
• We no longer test operators when they are initially added to the model, which speed up build times slightly. To re-enable this testing, enable the fail_fast RC setting. (#1532)
• LinearFilter now uses state space representations internally, which is faster and potentially more accurate. (#1535)
• The default value of y0 in Synapse.filt is now 0 instead of the initial value of the input signal. This allows unstable filters (e.g., integrators) to be used with filt. (#1535)
• LinearFilter now accepts the discretization method as an argument, rather than having it specified in make_step. (#1535)
• The synapse_kwargs argument to FilteredNoise has been removed. (#1535)
• Processes with internal state now declare that state by defining a make_state method and accepting a state parameter in make_step. (#1387)
• Simulator is now pickleable, allowing its state to be saved and loaded. (#1387)
• Renamed utils.testing.allclose to utils.testing.signals_allclose, to differentiate it from the allclose fixture. (#1563)
• The default intercepts value has been changed to Uniform(-1, 0.9) to avoid high gains when intercepts are close to 1. (#1534, #1561)
• The --simulator and --neurons pytest command line arguments are now specified by nengo_simulator and nengo_neurons entries in the pytest config file instead. (#1566)
• The nengo_test_unsupported option now uses pytest nodeids for the test names (the main change is that this means a double :: between file and function names). (#1566)
• Signals will now raise an error if their initial value contains NaNs. (#1571)
• The builder will now raise an error if any encoders are NaN, which can occur if an encoder has length zero. (#1571)
• Renamed simulator.ProbeDict to simulator.SimulationData. (#1574)
• Increased minimum numpy version to 1.13. (#1577)
• Documentation pages that had underscores in their filenames have been renamed to have hyphens instead. (#1585)

Deprecated

• Deprecated the nengo.spa module. Use the Nengo SPA project instead. (#1465)
• The A and B inputs to the Product and CircularConvolution networks are officially deprecated. Use input_a and input_b instead. (#887, #1179)
• nengo.utils.compat will be removed in the next minor release. (#1520)
• Deprecated utils.numpy.rmse. Call utils.numpy.rms on the difference between two arrays instead. (#1563)

Removed

• Networks no longer accept the net argument. To set network arguments like label, pass them as keyword arguments instead. (#1179)
• Removed generate_graphviz utility function. It can now be found in nengo_extras. (#1187)
• Removed functions for estimating firing rates from spikes. They can now be found in nengo_extras. ([#1187](https://...

Added

• Added a warning when setting gain and bias along with either of max_rates or intercepts, as the latter two parameters are ignored. (#1431, #1433)

Changed

• Learning rules can now be sliced when providing error input. (#1365, #1385)

• The order of parameters in learning rules has changed such that learning_rate always comes first. (#1095)

• Learning rules take pre_synapse, post_synapse, and theta_synapse instead of pre_tau, post_tau, and theta_tau respectively. This allows arbitrary Synapse objects to be used as filters on learning signals. (#1095)

Deprecated

• The nengo.ipynb IPython extension and the IPython2ProgressBar have been deprecated and replaced by the IPython5ProgressBar. This progress bar will be automatically activated in IPython and Jupyter notebooks from IPython version 5.0 onwards. (#1087, #1375)
• The pre_tau, post_tau, and theta_tau parameters for learning rules are deprecated. Instead, use pre_synapse, post_synapse, and theta_synapse respectively. (#1095)

Removed

• Removed nengo.utils.docutils in favor of using nbsphinx. (#1349)

Added

• Added amplitude parameter to LIF, LIFRate, and RectifiedLinear which scale the output amplitude. (#1325, #1391)
• Added the SpikingRectifiedLinear neuron model. (#1391)

Changed

• Default values can no longer be set for Ensemble.n_neurons or Ensemble.dimensions. (#1372)
• If the simulator seed is not specified, it will now be set from the network seed if a network seed is specified. (#980, #1386)

Fixed

• Fixed an issue in which signals could not be pickled, making it impossible to pickle Model instances. (#1135)
• Better error message for invalid return values in nengo.Node functions. (#1317)
• Fixed an issue in which accepting and passing (*args, **kwargs) could not be used in custom solvers. (#1358, #1359)
• Fixed an issue in which the cache would not release its index lock on abnormal termination of the Nengo process. (#1364)
• Fixed validation checks that prevented the default from being set on certain parameters. (#1372)
• Fixed an issue with repeated elements in slices in which a positive and negative index referred to the same dimension. (#1395)
• The Simulator.n_steps and Simulator.time properties now return scalars, as was stated in the documentation. (#1406)
• Fixed the --seed-offset option of the test suite. (#1409)

Added

• Added a NoSolver solver that can be used to manually pass in a predefined set of decoders or weights to a connection. (#1352)
• Added a Piecewise process, which replaces the now deprecated piecewise function. (#1036, #1100, #1355, #1362)

Changed

• The minimum required version of NumPy has been raised to 1.8. (#947)
• Learning rules can now have a learning rate of 0. (#1356)
• Running the simulator for zero timesteps will now issue a warning, and running for negative time will error. (#1354, #1357)

Fixed

• Fixed an issue in which the PES learning rule could not be used on connections to an ObjView when using a weight solver. (#1317)
• The progress bar that can appear when building a large model will now appear earlier in the build process. (#1340)
• Fixed an issue in which ShapeParam would always store None. (#1342)
• Fixed an issue in which multiple identical indices in a slice were ignored. (#947, #1361)

Deprecated

• The piecewise function in nengo.utils.functions has been deprecated. Please use the Piecewise process instead. (#1100)

Added

• Added a n_neurons property to Network, which gives the number of neurons in the network, including all subnetworks. (#435, #1186)
• Added a new example showing how adjusting ensemble tuning curves can improve function approximation. (#1129)
• Added a minimum magnitude option to UniformHypersphere. (#799)
• Added documentation on RC settings. (#1130)
• Added documentation on improving performance. (#1119, #1130)
• Added LinearFilter.combine method to combine two LinearFilter instances. (#1312)
• Added a method to all neuron types to compute ensemble max_rates and intercepts given gain and bias. (#1334)

Changed

• Learning rules now have a size_in parameter and attribute, allowing both integers and strings to define the dimensionality of the learning rule. This replaces the error_type attribute. (#1307, #1310)
• EnsembleArray.n_neurons now gives the total number of neurons in all ensembles, including those in subnetworks. To get the number of neurons in each ensemble, use EnsembleArray.n_neurons_per_ensemble. (#1186)
• The Nengo modelling API document now has summaries to help navigate the page. (#1304)
• The error raised when a Connection function returns None is now more clear. (#1319)
• We now raise an error when a Connection transform is set to None. (#1326)

Fixed

• Probe cache is now cleared on simulator reset. (#1324)
• Neural gains are now always applied after the synapse model. Previously, this was the case for decoded connections but not neuron-to-neuron connections. (#1330)
• Fixed a crash when a lock cannot be acquired while shrinking the cache. (#1335, #1336)

Added

• Added an optimizer that reduces simulation time for common types of models. The optimizer can be turned off by passing optimize=False to Simulator. (#1035)
• Added the option to not normalize encoders by setting Ensemble.normalize_encoders to False. (#1191, #1267)
• Added the Samples distribution to allow raw NumPy arrays to be passed in situations where a distribution is required. (#1233)

Changed

• We now raise an error when an ensemble is assigned a negative gain. This can occur when solving for gains with intercepts greater than 1. (#1212, #1231, #1248)
• We now raise an error when a Node or Direct ensemble produces a non-finite value. (#1178, #1280, #1286)
• We now enforce that the label of a network must be a string or None, and that the seed of a network must be an int or None. This helps avoid situations where the seed would mistakenly be passed as the label. (#1277, #1275)
• It is now possible to pass NumPy arrays in the ens_kwargs argument of EnsembleArray. Arrays are wrapped in a Samples distribution internally. (#691, #766, #1233)
• The default refractory period (tau_ref) for the Sigmoid neuron type has changed to 2.5 ms (from 2 ms) for better compatibility with the default maximum firing rates of 200-400 Hz. (#1248)
• Inputs to the Product and CircularConvolution networks have been renamed from A and B to input_a and input_b for consistency. The old names are still available, but should be considered deprecated. (#887, #1296)

Fixed

• Properly handle non C-contiguous node outputs. (#1184, #1185)

Deprecated

• The net argument to networks has been deprecated. This argument existed so that network components could be added to an existing network instead of constructing a new network. However, this feature is rarely used, and makes the code more complicated for complex networks. (#1296)

Added

• Added documentation on config system quirks. (#1224)
• Added nengo.utils.network.activate_direct_mode function to make it easier to activate direct mode in networks where some parts require neurons. (#1111, #1168)

Fixed

• The matrix multiplication example will now work with matrices of any size and uses the product network for clarity. (#1159)
• Fixed instances in which passing a callable class as a function could fail. (#1245)
• Fixed an issue in which probing some attributes would be one timestep faster than other attributes. (#1234, #1245)
• Fixed an issue in which SPA models could not be copied. (#1266, #1271)
• Fixed an issue in which Nengo would crash if other programs had locks on Nengo cache files in Windows. (#1200, #1235)

Changed

• Integer indexing of Nengo objects out of range raises an IndexError now to be consistent with standard Python behaviour. (#1176, #1183)
• Documentation that applies to all Nengo projects has been moved to https://nengo.github.io/. (#1251)