Project information

Release History

2.2.2 (unreleased)

2.2.1 (October 2, 2019)


  • Update testing framework to use new nengo pytest ecosystem (pytest-rng, pytest-allclose, and pytest-nengo)

  • Disable TensorFlow 2.0 behaviour (e.g. control flow v2) by default. This will be re-enabled when full TensorFlow 2.0 support is added.


  • Fixed tensorflow-gpu installation check in pep517-style isolated build environments.

2.2.0 (July 24, 2019)



  • The default session will now be set to the NengoDL session before calling TensorNodes’ post_build function.

  • Renamed the pytest unroll_simulation argument to unroll-simulation.

  • Switched to nengo-bones templating system for TravisCI config/scripts.

  • NengoDL will disable eager execution on import (and will probably not work properly if it is manually re-enabled).

  • Increased minimum numpy version to 1.14.5 (required by TensorFlow 1.14).

  • Minimum Nengo version is now 2.8.0.

  • Update LinearFilter synapse implementation to match recent changes in Nengo core (see


  • Fixed TensorFlow seeding so that randomness can be reliably controlled by setting the Simulator seed.

  • Improved robustness of tensorflow-gpu installation check (in particular, it will now correctly detect GPU dists installed through conda).

  • Fixed inspection of TensorNode.tensor_func arguments for partial functions.

  • Simulator seed will now be deterministic for a given top-level Network seed.

  • Raise a more informative error if user attempts to pickle a Simulator (this is not possible to do with TensorFlow sessions; see the documentation for other methods of saving/loading a NengoDL model).


  • NengoDL no longer supports Python 3.4 (official support for 3.4 ended in March 2019).

2.1.1 (January 11, 2019)



  • Increased minimum progressbar2 version to 3.39.0.

  • We now only provide sdist releases, not bdist_wheel. Due to the way the TensorFlow packages are organized, bdist_wheel forces any existing TensorFlow installations (e.g. tensorflow-gpu or tf-nightly) to be overwritten by tensorflow, which we don’t want to do.


  • Removed the nef-init tutorial (replaced by the new from-nengo tutorial).

2.1.0 (December 5, 2018)


  • Added a built-in objective to assist in applying regularization during training.

  • Added keep_history config option, which can be set to False on Probes if only the data from the most recent simulation step is desired (as opposed to the default behaviour of keeping the data from all steps).


  • Moved utils.mse to objectives.mse.

  • sim.loss will now apply nengo_dl.objectives.mse to all probes in data if no explicit objective is given (mirroring the default behaviour in sim.train).

  • The Spaun benchmark network will now be installed through pip rather than manually cloning and importing the repo.


  • Fixed objective argument parsing if objective is a callable class or method.

  • Fixed bug in sim.train 1-step synapse warning when explicitly specifying n_steps (rather than passing in data).


  • Passing "mse" as the objective in sim.train/sim.loss is no longer supported. Use the function nengo_dl.objectives.mse instead.

2.0.0 (November 23, 2018)

Breaking API changes

  • sim.train and sim.loss now accept a single data argument, which combines the previous inputs and targets arguments. For example,

    sim.train({my_node: x}, {my_probe: y}, ...)

    is now equivalent to

    sim.train({my_node: x, my_probe: y}, ...)

    The motivation for this change is that not all objective functions require target values. Switching to the more generic data argument simplifies the API and makes it more flexible, allowing users to specify whatever training/loss data is actually required.

  • The objective argument in sim.train/sim.loss is now always specified as a dictionary mapping probes to objective functions. Note that this was available but optional previously; it was also possible to pass a single value for the objective function, which would be applied to all probes in targets. The latter is no longer supported. For example,

    sim.train(..., objective="mse")

    must now be explicitly specified as

    sim.train(..., objective={my_probe: "mse"})

    The motivation for this change is that, especially with the other new features introduced in the 2.0 update, there were a lot of different ways to specify the objective argument. This made it somewhat unclear how exactly this argument worked, and the automatic “broadcasting” was also ambiguous (e.g., should the single objective be applied to each probe individually, or to all of them together?). Making the argument explicit helps clarify the mental model.


  • An integer number of steps can now be passed for the sim.loss/sim.train data argument, if no input/target data is required.

  • The objective dict in sim.train/sim.loss can now contain tuples of probes as the keys, in which case the objective function will be called with a corresponding tuple of probe/target values as each argument.

  • Added the sim.run_batch function. This exposes all the functionality that the functions are based on, allowing advanced users full control over how to run a NengoDL simulation.

  • Added option to disable progress bar in sim.train and sim.loss.

  • Added training argument to sim.loss to control whether the loss is evaluated in training or inference mode.

  • Added support for the new Nengo Transform API (see


  • Custom objective functions passed to sim.train/sim.loss can now accept a single argument (my_objective(outputs): ... instead of my_objective(outputs, targets): ...) if no target values are required.

  • utils.minibatch_generator now accepts a single data argument rather than inputs and targets (see discussion in “Breaking API changes”).

  • sim.training_step is now the same as tf.train.get_or_create_global_step().

  • Switched documentation to new nengo-sphinx-theme.

  • Reorganized documentation into “User guide” and “API reference” sections.

  • Improve build speed of models with large constants (#69)

  • Moved op-specific merge logic into the OpBuilder classes.


  • Ensure that training step is always updated before TensorBoard events are added (previously it could update before or after depending on the platform).


  • The input_feeds argument has been renamed to data (for consistency with other simulator functions).


1.2.1 (November 2, 2018)


  • Added a warning if users run one-timestep training with a network containing synaptic filters.


  • Test Simulator parameters are now controlled through pytest arguments, rather than environment variables.

  • Disable INFO-level TensorFlow logging (from C side) on import. Added a NengoDL log message indicating the device the simulation will run on, as a more concise replacement.

  • Boolean signals are now supported (#61)


  • Avoid backpropagating NaN gradients from spiking neurons.

  • Fixed an error that was thrown when calling get_tensor on a Signal that was first initialized inside the Simulation while loop (#56)

  • Allow TensorNodes to run in Nengo GUI.

  • Avoid bug in TensorFlow 1.11.0 that prevents certain models from running (see Note that this doesn’t prevent this from occurring in user models, as we cannot control the model structure there. If your model hangs indefinitely when you call sim.train, try downgrading to TensorFlow 1.10.0.

  • Ensure that sim.training_step is always updated after the optimization step (in certain race conditions it would sometimes update part-way through the optimization step).

1.2.0 (September 5, 2018)


  • NengoDL will now automatically use a rate-based approximation to compute the gradient for spiking neuron types, if one is known (no more need to manually swap neuron types for training and inference).

  • Added nengo_dl.configure_settings(inference_only=True) option, which will build the network in inference-only mode. This will slightly improve the inference speed of the simulation, but the network will not be trainable.

  • Added nengo_dl.configure_settings(lif_smoothing=x) option, which will control how much smoothing is applied to the LIF function during gradient calculations (if any).

  • Added documentation on the various NengoDL config options.

  • Added better validation for TensorNode output when size_out != None (#51)


  • More informative error message if the user tries to pass target values for a probe that isn’t used in the objective function.

  • Switched to ADD_N gradient accumulation (from TREE); this will increase the memory usage during training, but improve performance.

  • Revert to Timeline profiling method. tf.profiler can produce incorrect output, and isn’t maintained any more (

  • Reduce memory usage during training by caching temporary variables used when computing ScatterUpdate gradient.

  • Increase minimum TensorFlow version to 1.4.0.

  • Increased minimum NumPy version to 1.12.1 (required by TensorFlow)

  • Sort write signals as well as reads during graph optimization (encourages tighter partitioning, which can improve training/inference speed).

  • Moved configure_settings from to


  • Fixed a bug where nengo_dl.dists.VarianceScaling(..., distribution="normal") did not respect the seed if one was given.


  • The Simulator(dtype=...) argument has been deprecated; use nengo_dl.configure_settings(dtype=...) instead. Will be removed in 1.3.0.

1.1.0 (July 24, 2018)


  • The default TensorFlow Session is now set to the underlying Simulator session within the Simulator context.

  • Added CLI for

  • Added sim.freeze_params tool, to more easily extract model parameters for reuse in different Simulators.

  • Added documentation on saving and loading model parameters.

  • Added Spaun example in


  • Move tensorflow-gpu installation check to Simulator init, and only apply if device=None.

  • Switched to pylint for style checks.

  • TensorFlow INFO-level log messages are now disabled by default on import

  • All previous releases now tracked in documentation

  • Updated spiking MNIST example to simplify and improve performance.

  • Passing unknown configuration options to nengo_dl.configure_settings will now give a more explicit error message.

  • Improved speed of parameter fetching though get_nengo_params

  • Raise a warning if user tries to train a network with non-differentiable elements (requires tensorflow>=1.9.0)

  • Improved accuracy of SoftLIFRate implementation for small values (#45)

  • Simplified how TensorSignals are loaded into the TensorFlow graph


  • Better handling of Simulator errors not associated with a specific op (fixes #41)

  • Fixed node outputs changing after simulator is built (fixes #4)

  • Fixed some broken cross references in the documentation

  • Fixed several edge cases for get_nengo_params; don’t use trained gains for direct neuron connections, error raised if get_nengo_params applied to an Ensemble with Direct neurons

  • Compatible with tensorflow==1.9.0 release

  • Fixed bug in nengo_dl.configure_settings(session_config=...) when passing a pre-build model to the Simulator instead of a Network

  • Fixed TensorFlow version comparisons for 1.10.0



  • Removed nengo_dl.DATA_DIR constant

  • Removed benchmarks.compare_backends (use instead)

  • Removed ghp-import dependency

1.0.0 (May 30, 2018)


  • User can now directly specify the output error gradient, rather than using targets/objective (useful for when you have some external process for computing error that is not easy to implement as an objective function). See the documentation for details.

  • Added NengoDL white paper


  • Extra requirements for documentation/testing are now stored in’s extra_requires instead of requirements-*.txt. For example, instead of doing pip install -r requirements-test.txt, instead use pip install nengo-dl[tests] (or pip install -e .[tests] for a developer installation).

  • Improved efficiency of PES implementation


  • Removed sphinxcontrib-versioning dependency for building documentation

0.6.2 (May 4, 2018)


  • Added sim.get_nengo_params function to more easily extract model parameters for reuse when building different models.

  • Added Simulator(..., progress_bar=False) option to disable the progress information printed to console when the network is building.

  • TensorFlow session config options can now be set using nengo_dl.configure_settings (e.g., nengo_dl.configure_settings(session_config={"gpu_options.allow_growth": True}))

  • The signal sorting/graph simplificaton functions can now be configured through nengo_dl.configure_settings

  • Added extra_feeds parameter to, which can be used to feed Tensor values directly into the TensorFlow session


  • Improved speed of PES implementation by adding a custom operator.

  • Renamed project from nengo_dl to nengo-dl (to be more consistent with standard conventions). This only affects the display name of the project on PyPI/GitHub, and the documentation now resides at; there are no functional changes to user code.

  • Minor efficiency improvements to graph planner

  • Avoid using tf.constant, to get around TensorFlow’s 2GB limit on graph size when building large models


  • Checking nengo_dl version without nengo installed will no longer result in an error.

  • Updated progress bar to work with progressbar2>=3.37.0

  • Updated PES implementation to work with generic synapse types (see

  • Fixed installation to work with pip>=10.0

  • Fixed bug when using a TensorNode with a pre_build function and size_in==0

0.6.1 (March 7, 2018)


  • Added TensorFlow implementation for nengo.SpikingRectifiedLinear neuron type.


  • Optimizer variables (e.g., momentum values) will only be initialized the first time that optimizer is passed to sim.train. Subsequent calls to sim.train will resume with the values from the previous call.

  • Low-level simulation input/output formats have been reworked to make them slightly easier to use (for users who want to bypass or sim.train and access the TensorFlow session directly).

  • Batch dimension will always be first (if present) when checking model parameters via

  • TensorFlow ops created within the Simulator context will now default to the same device as the Simulator.

  • Update minimum Nengo version to 2.7.0


  • Better error message if training data has incorrect rank

  • Avoid reinstalling TensorFlow if one of the nightly build packages is already installed

  • Lowpass synapse can now be applied to multidimensional inputs

  • TensorNodes will no longer be built into the default graph when checking their output dimensionality.


  • Removed utils.cast_dtype function

0.6.0 (December 13, 2017)


  • The SoftLIFRate neuron type now has an amplitude parameter, which scales the output in the same way as the new amplitude parameter in LIF/LIFRate (see Nengo PR #1325).

  • Added progress_bar=False option to, which will disable the information about the simulation status printed to standard output (#17).

  • Added progress bars for the build/simulation process.

  • Added truncated backpropagation option to sim.train (useful for reducing memory usage during training). See the documentation for details.


  • Changed the default tensorboard argument in Simulator from False to None

  • Use the new tf.profiler tool to collect profiling data in sim.run_steps and sim.train when profile=True.

  • Minor improvements to efficiency of build process.

  • Minor improvements to simulation efficiency targeting small ops (tf.reshape/identity/constant).

  • Process inputs are now reseeded for each input when batch processing (if seed is not manually set).

  • Users can pass a dict of config options for the profile argument in run_steps/train, which will be passed on to the TensorFlow profiler; see the tf.profiler documentation for the available options.


  • Removed backports.print_function dependency


  • Fixed a bug where input nodes that were only read as a view were not feedable

  • Updated tensorflow-gpu installation check

  • Improved numerical stability of LIFRate gradients (#26)

  • Added more informative error message when data is provided with fewer items than sim.minibatch_size (#30)

0.5.2 (October 11, 2017)


  • TensorNode outputs can now define a post_build function that will be executed after the simulation is initialized (see the TensorNode documentation for details).

  • Added functionality for outputting summary data during the training process that can be viewed in TensorBoard (see the sim.train documentation).

  • Added some examples demonstrating how to use Nengo DL in a more complicated task using semantic pointers to encode/retrieve information

  • Added sim.training_step variable which will track the current training iteration (can be used, e.g., for TensorFlow’s variable learning rate operations).

  • Users can manually create tf.summary ops and pass them to sim.train summaries

  • The Simulator context will now also set the default TensorFlow graph to the one associated with the Simulator (so any TensorFlow ops created within the Simulator context will automatically be added to the correct graph)

  • Users can now specify a different objective for each output probe during training/loss calculation (see the sim.train documentation).


  • Resetting the simulator now only rebuilds the necessary components in the graph (as opposed to rebuilding the whole graph)

  • The default "mse" loss implementation will now automatically convert np.nan values in the target to zero error

  • If there are multiple target probes given to sim.train/sim.loss the total error will now be summed across probes (instead of averaged)


  • now implements the full collections.Mapping interface

  • Fixed bug where signal order was non-deterministic for Networks containing objects with duplicate names (#9)

  • Fixed bug where non-slot optimizer variables were not initialized (#11)

  • Implemented a modified PES builder in order to avoid slicing encoders on non-decoded PES connections

  • TensorBoard output directory will be automatically created if it doesn’t exist

0.5.1 (August 28, 2017)


  •[obj] will now return live parameter values from the simulation, rather than initial values from the build process. That means that it can be used to get the values of object parameters after training, e.g.[my_conn].weights.

  • Increased minimum Nengo version to 2.5.0.

  • Increased minimum TensorFlow version to 1.3.0.

0.5.0 (July 11, 2017)


  • Added nengo_dl.tensor_layer to help with the construction of layer-style TensorNodes (see the TensorNode documentation)

  • Added an example demonstrating how to train a neural network that can run in spiking neurons

  • Added some distributions for weight initialization to nengo_dl.dists

  • Added sim.train(..., profile=True) option to collect profiling information during training

  • Added new methods to simplify the Nengo operation graph, resulting in faster simulation/training speed

  • The default graph planner can now be modified by setting the planner attribute on the top-level Network config

  • Added TensorFlow implementation for general linear synapses

  • Added backports.tempfile and backports.print_function requirement for Python 2.7 systems


  • Increased minimum TensorFlow version to 1.2.0

  • Improved error checking for input/target data

  • Improved efficiency of stateful gradient operations, resulting in faster training speed

  • The functionality for nengo_dl.configure_trainable has been subsumed into the more general nengo_dl.configure_settings(trainable=x). This has resulted in some small changes to how trainability is controlled within subnetworks; see the updated documentation for details.

  • Calling Simulator.train/Simulator.loss no longer resets the internal state of the simulation (so they can be safely intermixed with calls to


  • The old step_blocks/unroll_simulation syntax has been fully deprecated, and will result in errors if used


  • Fixed bug related to changing the output of a Node after the model is constructed (#4)

  • Order of variable creation is now deterministic (helps make saving/loading parameters more reliable)

  • Configuring whether or not a model element is trainable does not affect whether or not that element is minibatched

  • Correctly reuse variables created inside a TensorNode when unroll_simulation > 1

  • Correctly handle probes that aren’t connected to any ops

  • Swapped fan_in/fan_out in dists.VarianceScaling to align with the standard definitions

  • Temporary patch to fix memory leak in TensorFlow (see #11273)

  • Fixed bug related to nodes that had matching output functions but different size_out

  • Fixed bug related to probes that do not contain any data yet

0.4.0 (June 8, 2017)



  • Updated TensorFuncParam to new Nengo Param syntax

  • The interface for Simulator step_blocks/unroll_simulation has been changed. Now unroll_simulation takes an integer as argument which is equivalent to the old step_blocks value, and unroll_simulation=1 is equivalent to the old unroll_simulation=False. For example, Simulator(..., unroll_simulation=True, step_blocks=10) is now equivalent to Simulator(..., unroll_simulation=10).

  • Simulator.train/Simulator.loss no longer require step_blocks (or the new unroll_simulation) to be specified; the number of steps to train across will now be inferred from the input data.

0.3.1 (May 12, 2017)


  • Added more documentation on Simulator arguments


  • Improved efficiency of tree_planner, made it the new default planner


  • Correctly handle input feeds when n_steps > step_blocks

  • Detect cycles in transitive planner

  • Fix bug in uneven step_blocks rounding

  • Fix bug in Simulator.print_params

  • Fix bug related to merging of learning rule with different dimensionality

  • Use tf.Session instead of tf.InteractiveSession, to avoid strange side effects if the simulator isn’t closed properly

0.3.0 (April 25, 2017)


  • Use logger for debug/builder output

  • Implemented TensorFlow gradients for sparse Variable update Ops, to allow models with those elements to be trained

  • Added tutorial/examples on using Simulator.train

  • Added support for training models when unroll_simulation=False

  • Compatibility changes for Nengo 2.4.0

  • Added a new graph planner algorithm, which can improve simulation speed at the cost of build time


  • Significant improvements to simulation speed

    • Use sparse Variable updates for signals.scatter/gather

    • Improved graph optimizer memory organization

    • Implemented sparse matrix multiplication op, to allow more aggressive merging of DotInc operators

  • Significant improvements to build speed

    • Added early termination to graph optimization

    • Algorithmic improvements to graph optimization functions

  • Reorganized documentation to more clearly direct new users to relevant material


  • Fix bug where passing a built model to the Simulator more than once would result in an error

  • Cache result of calls to tensor_graph.build_loss/build_optimizer, so that we don’t unnecessarily create duplicate elements in the graph on repeated calls

  • Fix support for Variables on GPU when unroll_simulation=False

  • SimPyFunc operators will always be assigned to CPU, even when device="/gpu:0", since there is no GPU kernel

  • Fix bug where Simulator.loss was not being computed correctly for models with internal state

  • Data/targets passed to Simulator.train will be truncated if not evenly divisible by the specified minibatch size

  • Fixed bug where in some cases Nodes with side effects would not be run if their output was not used in the simulation

  • Fixed bug where strided reads that cover a full array would be interpreted as non-strided reads of the full array

0.2.0 (March 13, 2017)

Initial release of TensorFlow-based NengoDL

0.1.0 (June 12, 2016)

Initial release of Lasagne-based NengoDL

Contributing to NengoDL

Issues and pull requests are always welcome! We appreciate help from the community to make NengoDL better.

Filing issues

If you find a bug in NengoDL, or think that a certain feature is missing, please consider filing an issue! Please search the currently open issues first to see if your bug or feature request already exists. If so, feel free to add a comment to the issue so that we know that multiple people are affected.

Making pull requests

If you want to fix a bug or add a feature to NengoDL, we welcome pull requests. Ensure that you fill out all sections of the pull request template, deleting the comments as you go. We check most aspects of code style automatically. Please refer to our code style guide for things that we check manually.

Contributor agreement

We require that all contributions be covered under our contributor assignment agreement. Please see the agreement for instructions on how to sign.

More details

For more details on how to contribute to Nengo, please see the developer guide.

NengoDL license

Copyright (c) 2015-2019 Applied Brain Research

NengoDL is made available under a proprietary license that permits using, copying, sharing, and making derivative works from NengoDL and its source code for any non-commercial purpose, as long as the above copyright notice and this permission notice are included in all copies or substantial portions of the software.

If you would like to use NengoDL commercially, licenses can be purchased from Applied Brain Research. Please contact for more information.


Licensed code

NengoDL imports several open source libraries:

To build the documentation, NengoDL uses:

To run the tests, NengoDL uses:


If you would like to cite NengoDL in your research, please cite the white paper:

Rasmussen, D. (2018). NengoDL: Combining deep learning and neuromorphic
modelling methods. arXiv:1805.11144, 1–22.
  archivePrefix = {arXiv},
  arxivId = {1805.11144},
  author = {Rasmussen, Daniel},
  journal = {arXiv},
  pages = {1--22},
  title = {{NengoDL}: Combining deep learning and neuromorphic modelling
  url = {},
  volume = {1805.11144},
  year = {2018}