Configuration options

NengoDL uses Nengo’s config system to allow users to control more fine-grained aspects of the simulation. In general, most users will not need to worry about these options, and can leave them at their default settings. However, these options may be useful in some scenarios.

configure_settings is a utility function that can be used to set these configuration options. It needs to be called within a Network context, as in:

with nengo.Network() as net:
    nengo_dl.configure_settings(config_option=config_value, ...)
    ...

Each call to configure_settings only sets the configuration options specified in that call. That is,

nengo_dl.configure_settings(option0=val0)
nengo_dl.configure_settings(option1=val1)

is equivalent to

nengo_dl.configure_settings(option0=val0, option1=val1)

Under the hood, configure_settings is setting Network.config attributes on the top-level network. All of the same effects could be achieved by setting those Network.config attributes directly, configure_settings simply makes this easier.

Note that TensorFlow also has its own config system, which can be used as normal with NengoDL to control the underlying TensorFlow behaviour.

trainable

The trainable config attribute can be used to control which parts of a model will be optimized by the Simulator.fit process. configure_settings(trainable=None) will add a configurable trainable attribute to the objects in a network. Setting trainable=None will use the default trainability settings, or trainable=True/False can be used to override the default for all objects.

Once the trainable attribute has been added to all the objects in a model, the Network.config system can then be used to control the trainability of individual objects.

For example, suppose we only want to optimize one connection in our network, while leaving everything else unchanged. This could be achieved via

with nengo.Network() as net:
    # this adds the `trainable` attribute to all the trainable objects
    # in the network, and initializes it to `False`
    nengo_dl.configure_settings(trainable=False)

    a = nengo.Node([0])
    b = nengo.Ensemble(10, 1)
    c = nengo.Node(size_in=1)

    nengo.Connection(a, b)

    # make this specific connection trainable
    conn = nengo.Connection(b, c)
    net.config[conn].trainable = True

Or if we wanted to disable training for some subnetwork:

with nengo.Network() as net:
    nengo_dl.configure_settings(trainable=None)
    ...
    with nengo.Network() as subnet:
        nengo_dl.configure_settings(trainable=False)
        ...

Note that config[nengo.Ensemble].trainable controls both encoders and biases, as both are properties of an Ensemble. However, it is possible to separately control the biases via config[nengo.ensemble.Neurons].trainable or config[my_ensemble.neurons].trainable.

There is one important caveat to keep in mind when configuring trainable, which differ from the standard config behaviour. trainable applies to all objects in a network, regardless of whether they were created before or after trainable is set. For example,

with nengo.Network() as net:
    nengo_dl.configure_settings(trainable=None)
    ...
    net.config[nengo.Ensemble].trainable = False
    a = nengo.Ensemble(10, 1)
    ...

is the same as

with nengo.Network() as net:
    nengo_dl.configure_settings(trainable=None)
    ...
    a = nengo.Ensemble(10, 1)
    net.config[nengo.Ensemble].trainable = False
    ...

Trainability settings are prioritized according to the following rules:

  1. Settings in lower subnetworks take priority.

    with nengo.Network() as net:
        nengo_dl.configure_settings(trainable=True)
        with nengo.Network() as subnet:
            nengo_dl.configure_settings(trainable=False)
    
            # this ensemble will not be trainable, because settings on
            # `subnet` override settings on `net`
            a = nengo.Ensemble(10, 1)
    
  2. Settings on instances take priority over classes.

    with nengo.Network() as net:
        nengo_dl.configure_settings(trainable=None)
    
        a = nengo.Ensemble(10, 1)
    
        # this will make `a` trainable (even though Ensembles in general
        # are not trainable)
        net.config[a].trainable = True
        net.config[nengo.Ensemble].trainable = False
    

planner

This option can be used to change the algorithm used for assigning an order to simulation operations during the graph optimization stage. For example, we could disable operator merging by using the noop_planner.

from nengo_dl.graph_optimizer import noop_planner

with nengo.Network() as net:
    nengo_dl.configure_settings(planner=noop_planner)

sorter

This option can be used to change the algorithm used for sorting signals/operators during the graph optimization stage. For example, we could disable sorting via

from nengo_dl.graph_optimizer import noop_order_signals

with nengo.Network() as net:
    nengo_dl.configure_settings(sorter=noop_order_signals)

simplifications

This option can be used to change the simplification transformations applied during the graph optimization stage. This takes a list of transformation functions, where each will be applied in sequence. For example, we could apply only two of the default simplifications via

from nengo_dl.graph_optimizer import remove_identity_muls, remove_zero_incs

with nengo.Network() as net:
    nengo_dl.configure_settings(simplifications=[remove_identity_muls,
                                                 remove_zero_incs])

inference_only

By default, NengoDL models are built to support both training and inference. However, sometimes we may know that we’ll only be using a simulation for inference (for example, if we want to take advantage of the batching/GPU acceleration of NengoDL, but don’t need the sim.fit functionality). In that case we can improve the simulation speed of the model by omitting some of the aspects related to training. Setting nengo_dl.configure_settings(inference_only=True) will cause the network to be built in inference-only mode.

lif_smoothing

During training, NengoDL automatically replaces the non-differentiable spiking LIF neuron model with the differentiable LIFRate approximation. However, although LIFRate is generally differentiable, it has a sharp discontinuity at the firing threshold. In some cases this can lead to difficulties during the training process, and performance can be improved by smoothing the LIFRate response around the firing threshold. This is known as the SoftLIFRate neuron model.

SoftLIFRate has a parameter sigma that controls the degree of smoothing (SoftLIFRate approaches LIFRate as sigma goes to zero). Setting nengo_dl.configure_settings(lif_smoothing=x) will cause the LIF gradients to be approximated by SoftLIFRate instead of LIFRate, with sigma=x.

dtype

This specifies the floating point precision to be used for the simulator’s internal computations. It can be either "float32" or "float64", for 32 or 64-bit precision, respectively. 32-bit precision is the default, as it is faster, will use less memory, and in most cases will not make a difference in the results of the simulation. However, if very precise outputs are required then this can be changed to "float64".

keep_history

By default, a nengo.Probe stores the probed output from every simulation timestep. However, sometimes in NengoDL we want to add a probe to something for other reasons, and don’t necessarily care about all of that data (which can consume a lot of memory). For example, we might want to apply a probe to some connection weights so that we can apply a regularization penalty, but since the weights aren’t changing during a simulation run we don’t need to keep the value from every simulation step.

The keep_history config option allows Probes to be configured so that they only store the output of the probed signal from the last simulation timestep. Calling

with nengo.Network() as net:
    nengo_dl.configure_settings(keep_history=False)

will set the default value for all probes in the simulation, which can then be further configured on a per-probe basis, e.g.

with nengo.Network() as net:
   nengo_dl.configure_settings(keep_history=True)

   my_ens = nengo.Ensemble(10, 1)
   my_probe = nengo.Probe(my_ens)
   net.config[my_probe].keep_history = False

stateful

By default, a NengoDL simulator is built to be stateful (meaning that internal simulation state can be preserved between runs). However, if you know that you will not need this functionality (i.e. you want all Simulator executions to begin from the default initial conditions) it can be disabled by setting nengo_dl.configure_settings(stateful=False). This may slightly improve the simulation speed.

Note that in any case the internal state of the Simulation will be tracked within a given call (e.g. within one call to Simulator.run). This only affects whether state is preserved between calls.

use_loop

By default, NengoDL models run inside a loop within TensorFlow; this is what allows us to flexibly simulate a model for any number of timesteps. However, in some cases we may not need this functionality (for example, if we have a simple feedforward network that will only ever be simulated for a single timestep). In that case we can set nengo_dl.configure_settings(use_loop=False) to build the model without the outer loop, which can improve the simulation speed.

Note that it is still possible to have a model that simulates multiple timesteps by setting nengo_dl.Simulator(..., unroll_simulation=x). This will explicitly build x timesteps into the model (without using a loop). So if we use unroll_simulation=x and use_loop=False, then the simulation will always run for exactly x timesteps.

learning_phase

NengoDL (and Keras in general) can run models in two different modes, or “phases”: training and inference. Typically the mode is set automatically based on the function used to execute the model; sim.fit will run in training mode, and all other functions (e.g., sim.predict, sim.run, and sim.evaluate) run in inference mode.

The most important way that the learning phase affects a NengoDL model is that it controls whether spiking neurons are simulated in spiking or non-spiking mode. Normally during training spiking neurons will automatically swap their behaviour to a non-spiking equivalent, and use the spiking behaviour during inference.

However, sometimes it can be useful to override this default behaviour. For example, we might want to evaluate a spiking model in training mode, to get more insight into how it is behaving during training. That is the role of the learning_phase config option. It can be specified during Network construction to make a network that will always run in training/inference mode, regardless of what function is being called. For example:

with nengo.Network() as net:
    nengo_dl.configure_settings(learning_phase=True)

    # this ensemble will always use the "training" mode of LIF
    # (equivalent to LIFRate)
    ens = nengo.Ensemble(10, 1, neuron_type=nengo.LIF())