Lava Backend

isort:skip_file

fugu.backends.lava_backend.warnIfValueExceedsPrecision(value, precision, value_name)

fugu.backends.lava_backend.warnIfFeatureNotAvailable(feature)

exception fugu.backends.lava_backend.FeatureNotAvailableException

Bases: Exception

exception fugu.backends.lava_backend.NonIntegerDelayValueException

Bases: Exception

fugu.backends.lava_backend.reduce_factor_by_bits(value, num_bits)

fugu.backends.lava_backend.calculate_loihi_scale_factor(max_threshold=0, max_weight=0, max_bias=0, starting_scale=1048576, min_scale=64, threshold_bit_limit=16)

class fugu.backends.lava_backend.InputIterator

Bases: object

Supporting class for lava_Backend. Feeds the inputs to Dataloader process. Note that the output of this iterator is a tuple (input values, ground truth). Ground truth can be a scalar or vector. Since we don’t care, we output the scalar 0.

shape()

class fugu.backends.lava_backend.lava_Backend

Bases: Backend

_allocate(v, dv, vth, b, count=1)

Lava LIF has single fixed {du, dv, vth} for entire population. We don’t do anything with du, but every (dv, vth) combination requires a separate population (or “process” in Lava terminology). We also need to assemble a list of initial voltages for each population.

compile(scaffold, compile_args={})

creates neuron populations and synapses

run(n_steps=10, return_potentials=False)

Runs circuit for n_steps then returns data

cleanup()

Deletes/frees neurons and synapses

reset()

Resets time-step to 0 and resets neuron/synapse properties

set_properties(properties={})

Set properties for specific neurons and synapses :param properties: dictionary of parameter for bricks

Example

for brick in properties:

neuron_props, synapse_props = self.circuit[brick].get_changes(properties[brick]) for neuron in neuron_props:

set neuron properties

for synapse in synapse_props:

set synapse properties

@NOTE: Currently, this function behaves differently for Input Bricks

• Instead of returning the changes, they change internally and reset the iterator

• This is because of how initial spike times are calculated using said bricks

• I have not yet found a way of incorporating my proposed method (above) into these bricks yet

set_input_spikes()