Developer guide#

Networks#

Network is the mutable authoring object. Compilation turns it into immutable graph data. Every graph has a name and timestep.

net = snn.Network("example", dt=0.1 * snn.ms)

Names must be unique and contain no whitespace. They become stable identifiers in graphs, recordings, and checkpoints.

Signals and inputs#

Signals carry an identifier, shape, physical unit, and signal type. Time-varying inputs use the canonical (time, batch, channels) axis order.

image = net.input(
    "image",
    shape=("time", "batch", 784),
    signal_type="spikes",
    unit="spike",
)

This declares what the graph accepts. It does not decide whether the spikes came from MNIST, SHD, a Poisson encoder, or a recorded event stream.

Populations#

A population contains equally configured neurons and exposes named ports such as spikes, voltage, excitatory, and inhibitory.

excitatory = net.population(
    "E",
    size=80,
    neuron=snn.COBA_LIF(
        capacitance_nf=1.0,
        leak_us=0.05,
        resting_mv=-65.0,
        threshold_mv=-50.0,
        reset_mv=-65.0,
        refractory_steps=12,
    ),
)

The authoring API describes COBA-LIF, LIF, and non-spiking leaky-integrator populations. The graph executor currently supports COBA-LIF and leaky integrators. General current-based LIF execution is not implemented.

Parameters#

Parameters have stable names, shapes, units, initializers, and optional constraints. Projections normally create their own dense weight parameter, but a named ParameterRef may be supplied explicitly.

Initializers explicitly distinguish lower-clamped normal, signed normal, uniform, constant, and zero distributions. Lower-clamped normal may request Bernoulli or exact-fan-in initial zeroing; those zeros remain ordinary parameters rather than a permanent connectivity mask. Projection weights record fan-in normalization, while direct readout parameters record direct storage. The execution result reports each parameter’s initializer, constraint, unit, runtime shape, scaling rule, and realized count, zero fraction, mean, standard deviation, minimum, and maximum.

API reference#

Network#

Network(name: str, *, dt: Quantity = 0.1 * ms)

name becomes the graph identifier. dt is a time Quantity. Names claimed inside one network must be non-empty, unique, and contain no whitespace.

Network.input#

net.input(name, *, shape, signal_type, unit="1") -> Signal

shape is a tuple of integers and symbolic axes. Time-varying dense inputs begin with ("time", "batch", ...). The returned signal identifier is <name>.value.

Network.population#

net.population(name, *, size, neuron, spiking=True) -> Population

size must be positive. A spiking population exposes .spikes and .voltage; a non-spiking population exposes .voltage. Target ports are .excitatory, .inhibitory, and .modulatory.

Network.parameter and Network.constant#

net.parameter(name, *, shape, initializer, unit="1", constraint=None) -> ParameterRef
net.constant(name, value, *, unit="1") -> str

Parameter constructors are LowerClampedNormal(mean, std, initial_zero_fraction=0, zeroing="bernoulli"), its compatibility spelling Normal(mean, std), SignedNormal(mean, std), Uniform(low, high), Constant(value), and Zeros(). Exact fan-in zeroing uses zeroing="exact_k". NonNegative() is the implemented constraint. Neuron constructors are COBA_LIF(**values), LIF(**values), and LeakyIntegrator(**values). Unit helpers are ms, mV, nS, uS, and Hz.

Core return types#

Signal exposes id, shape, unit, signal_type, owner, and port. Population exposes id, size, neuron, spiking, group, and its signal or target-port properties. ParameterRef contains the stable parameter id.

exp106 — Components, projections, and delays