def validate_block(block):
# -- Compartment
validate_compartment(block.compartment)
# -- Axons
OUT_AXONS_MAX = d(b'NDA5Ng==', int)
n_axons = sum(a.axon_slots() for a in block.axons)
if n_axons > OUT_AXONS_MAX:
raise BuildError("Output axons (%d) exceeded max (%d)" %
(n_axons, OUT_AXONS_MAX))
for axon in block.axons:
validate_axon(axon)
# -- Synapses
IN_AXONS_MAX = d(b'NDA5Ng==', int)
n_axons = sum(s.n_axons for s in block.synapses)
if n_axons > IN_AXONS_MAX:
raise BuildError("Input axons (%d) exceeded max (%d)" %
(n_axons, IN_AXONS_MAX))
MAX_SYNAPSE_BITS = d(b'MTA0ODU3Ng==', int)
synapse_bits = sum(s.bits() for s in block.synapses)
if synapse_bits > MAX_SYNAPSE_BITS:
raise BuildError("Total synapse bits (%d) exceeded max (%d)" %
(synapse_bits, MAX_SYNAPSE_BITS))
for synapse in block.synapses:
validate_synapse(synapse)
# -- Probes
for probe in block.probes:
validate_probe(probe)
def atom_bits_extra(self):
"""Number of extra bits needed for the atom for incoming pop16 spikes."""
if self.pop_type == 16:
atom_bits = self.atom_bits()
assert atom_bits <= d(b"OQ==", int), "Too many atom bits"
return max(atom_bits - d(b"NQ==", int), 0)
else:
return 0 # meaningless if pop_type != 16
def add_spikes_to_generator(cls, t, spikes, basic_spike_generator):
methods = {
0: getattr(basic_spike_generator, d(b"YWRkU3Bpa2U=")),
16: getattr(basic_spike_generator, d(b"YWRkUG9wMTZTcGlrZQ==")),
32: getattr(basic_spike_generator, d(b"YWRkUG9wMzJTcGlrZQ==")),
}
time = d(b"dGltZQ==")
chip_id = d(b"Y2hpcElk")
core_id = d(b"Y29yZUlk")
axon_id = d(b"YXhvbklk")
atom = d(b"c3JjQXRvbQ==")
atom_bits_extra = d(b"YXRvbUJpdHM=")
for spike in spikes:
axon_type = int(spike["axon_type"])
kwargs = {
time: t,
chip_id: spike["chip_id"],
core_id: spike["core_id"],
axon_id: spike["axon_id"],
}
if axon_type == 0:
assert spike[
"atom"] == 0, "Atom must be zero for discrete spikes"
else:
kwargs[atom] = spike["atom"]
if axon_type == 16:
kwargs[atom_bits_extra] = spike["atom_bits_extra"]
methods[axon_type](**kwargs)
class CompartmentConfig(Config):
DECAY_U_MAX = d(b"NDA5NQ==", int)
DECAY_V_MAX = d(b"NDA5NQ==", int)
REFRACT_DELAY_MAX = d(b"NjM=", int)
params = ("decay_u", "decay_v", "refract_delay", "enable_noise")
def __init__(self, decay_v, decay_u, refract_delay, enable_noise):
super(CompartmentConfig, self).__init__()
self.decay_v = decay_v
self.decay_u = decay_u
self.refract_delay = refract_delay
self.enable_noise = enable_noise
class CompartmentConfig(Config):
DECAY_U_MAX = d(b'NDA5NQ==', int)
DECAY_V_MAX = d(b'NDA5NQ==', int)
REFRACT_DELAY_MAX = d(b'NjM=', int)
params = ('decay_u', 'decay_v', 'refract_delay', 'enable_noise')
def __init__(self, decay_v, decay_u, refract_delay, enable_noise):
super(CompartmentConfig, self).__init__()
self.decay_v = decay_v
self.decay_u = decay_u
self.refract_delay = refract_delay
self.enable_noise = enable_noise
def validate_synapse(synapse):
validate_synapse_cfg(synapse.synapse_cfg)
if synapse.axon_compartment_bases is not None:
min_base = d(b'LTE=', int)
max_base = d(b'MjU2', int)
assert all(
min_base <= b < max_base for b in synapse.axon_compartment_bases
), ("compartment base must be >= %d and < %d (-1 indicating unused)" %
(min_base, max_base))
if synapse.pop_type == 16:
if synapse.axon_compartment_bases is not None:
assert all(b % d(b'NA==', int) == 0
for b in synapse.axon_compartment_bases)
def discretize_weights(synapse_cfg,
w,
dtype=np.int32,
lossy_shift=True,
check_result=True):
"""Takes weights and returns their quantized values with weight_exp.
The actual weight to be put on the chip is this returned value
divided by the ``scale`` attribute.
Parameters
----------
w : float ndarray
Weights to be discretized, in the range -255 to 255.
dtype : np.dtype, optional (Default: np.int32)
Data type for discretized weights.
lossy_shift : bool, optional (Default: True)
Whether to mimic the two-part weight shift that currently happens
on the chip, which can lose information for small weight_exp.
check_results : bool, optional (Default: True)
Whether to check that the discretized weights fall in
the valid range for weights on the chip (-256 to 255).
"""
s = synapse_cfg.shift_bits
m = 2**(d(b'OA==', int) - s) - 1
w = np.round(w / 2.**s).clip(-m, m).astype(dtype)
s2 = s + synapse_cfg.weight_exp
if lossy_shift:
if s2 < 0:
warnings.warn("Lost %d extra bits in weight rounding" % (-s2, ))
# Round before `s2` right shift. Just shifting would floor
# everything resulting in weights biased towards being smaller.
w = (np.round(w * 2.**s2) / 2**s2).clip(-m, m).astype(dtype)
shift(w, s2, out=w)
np.left_shift(w, d(b'Ng==', int), out=w)
else:
shift(w, d(b'Ng==', int) + s2, out=w)
if check_result:
ws = w // synapse_cfg.scale
assert (np.all(ws <= d(b'MjU1', int))
and np.all(ws >= d(b'LTI1Ng==', int)))
return w
def scale_pes_errors(error, scale=1.0):
"""Scale PES errors based on a scaling factor, round and clip."""
error = scale * error
error = np.round(error).astype(np.int32)
max_err = d(b"MTI3", int)
q = error > max_err
if np.any(q):
warnings.warn(
"Received PES error greater than chip max (%0.2e). "
"Consider changing `Model.pes_error_scale`." % (max_err / scale,)
)
logger.debug(
"PES error %0.2e > %0.2e (chip max)", np.max(error) / scale, max_err / scale
)
error[q] = max_err
q = error < -max_err
if np.any(q):
warnings.warn(
"Received PES error less than chip min (%0.2e). "
"Consider changing `Model.pes_error_scale`." % (-max_err / scale,)
)
logger.debug(
"PES error %0.2e < %0.2e (chip min)",
np.min(error) / scale,
-max_err / scale,
)
error[q] = -max_err
return error
def set_learning(
self, learning_rate=1., tracing_tau=2, tracing_mag=1.0, wgt_exp=4):
assert tracing_tau == int(tracing_tau), "tracing_tau must be integer"
self.learning = True
self.tracing_tau = int(tracing_tau)
self.tracing_mag = tracing_mag
# stdp_cfg hard-coded for now (see hardware.builder)
self.format(learning_cfg=d(b'MQ==', int), stdp_cfg=d(b'MA==', int))
self.train_epoch = 2
self.learn_epoch_k = 1
self.learn_epoch = self.train_epoch * 2**self.learn_epoch_k
self.learning_rate = learning_rate * self.learn_epoch
self.learning_wgt_exp = wgt_exp
def _chip2host_snips(self, probes_receivers):
count = self.nengo_io_c2h_count
data = self.nengo_io_c2h.read(count)
time_step, data = data[0], np.array(data[1:], dtype=np.int32)
snip_range = self.nengo_io_snip_range
for probe in self._snip_probe_data:
assert probe.use_snip
x = data[snip_range[probe]]
assert x.ndim == 1
if probe.key == 'spiked':
assert isinstance(probe.target, LoihiBlock)
refract_delays = probe.target.compartment.refract_delay
# Loihi uses the voltage value to indicate where we
# are in the refractory period. We want to find neurons
# starting their refractory period.
x = (x == refract_delays * d(b'MTI4', int))
if probe.weights is not None:
x = np.dot(x, probe.weights)
receiver = probes_receivers.get(probe, None)
if receiver is not None:
# chip->host
receiver.receive(self.model.dt * time_step, x)
else:
# onchip probes
self._snip_probe_data[probe].append(x)
self._chip2host_sent_steps += 1
def validate_synapse(synapse):
validate_synapse_cfg(synapse.synapse_cfg)
if synapse.axon_compartment_bases is not None:
min_base = d(b"LTE=", int)
max_base = d(b"MjU2", int)
assert all(
min_base <= b < max_base for b in synapse.axon_compartment_bases
), ("compartment base must be >= %d and < %d (-1 indicating unused)" %
(min_base, max_base))
if synapse.pop_type == 16:
if synapse.axon_compartment_bases is not None:
assert all(
b % 4 == 0 for b in synapse.axon_compartment_bases
if b >= 0), (
"Pop16 axons must have all compartment bases modulo 4: %s"
% synapse.axon_compartment_bases)
def block_to_chip(self, block, chip):
if block.compartment.n_compartments > d(b'MTAyNA==', int):
raise ValidationError("Segment does not fit on one chip",
"n_neurons")
core = chip.new_core()
core.add_block(block)
compartment_cfgs, compartment_cfg_idxs = core_compartment_cfgs(core)
[core.add_compartment_cfg(cfg) for cfg in compartment_cfgs]
core.compartment_cfg_idxs = compartment_cfg_idxs
vth_cfgs, vth_cfg_idxs = core_vth_cfgs(core)
[core.add_vth_cfg(cfg) for cfg in vth_cfgs]
core.vth_cfg_idxs = vth_cfg_idxs
for synapse in block.synapses:
core.add_synapse(synapse)
stdp_pre_cfgs, stdp_pre_cfg_idxs = core_stdp_pre_cfgs(core)
[core.add_stdp_pre_cfg(stdp_pre_cfg) for stdp_pre_cfg in stdp_pre_cfgs]
core.stdp_pre_cfg_idxs = stdp_pre_cfg_idxs
core.stdp_pre_cfg_idx = None # hardware.builder will set
core.stdp_cfg_idx = None # hardware.builder will set
def prepare_for_probe(self, block, pinfo, target_idx):
chip_idx = pinfo.chip_idx[target_idx]
core_id = pinfo.core_id[target_idx]
compartment_idxs = pinfo.compartment_idxs[target_idx]
self.cores.add(core_id)
key = pinfo.key
if key == "spike":
refract_delay = block.compartment.refract_delay[0]
assert np.all(block.compartment.refract_delay == refract_delay)
key = refract_delay * d(b"MTI4", int)
n_comps = len(compartment_idxs)
logger.info(n_comps)
comp0 = compartment_idxs[0]
comp_diff = np.diff(compartment_idxs)
is_ranged_comps = np.all(
comp_diff == comp_diff[0] if len(comp_diff) > 0 else False
)
is_packed_spikes = is_ranged_comps and (pinfo.key == "spike")
n_packed_spikes = n_comps if is_packed_spikes else 0
output_len = ceil_div(n_comps, 32) if is_packed_spikes else n_comps
output_slice = slice(self.last_output, self.last_output + output_len)
pinfo.snip_range.append((chip_idx, output_slice, n_packed_spikes))
offset = self.output_offset + self.last_output
if is_ranged_comps:
self.probes.append((offset, key, core_id, comp0, comp_diff[0], n_comps))
else:
for i, comp in enumerate(compartment_idxs):
self.probes.append((offset + i, key, core_id, comp, 0, 1))
self.last_output += output_len
def run_steps(self, steps, blocking=True):
if self.use_snips and self.nengo_io_h2c is None:
self.create_io_snip()
# NOTE: we need to call connect() after snips are created
self.connect()
d_get(self.nxsdk_board, b'cnVu')(steps, **{
d(b'YVN5bmM='): not blocking
})
def run_steps(self, steps, blocking=True):
assert self.connected, "Interface is not built"
# start the board running the desired number of steps
d_get(self.nxsdk_board, b"cnVu")(steps, **{d(b"YVN5bmM="): not blocking})
# connect snips
if self.use_snips and not self.snips.connected:
self.snips.connect(self.nxsdk_board)
def set_population_weights(
self, weights, indices, axon_to_weight_map, compartment_bases, pop_type=None
):
"""Set population weights on this Synapse."""
self.axon_to_weight_map = axon_to_weight_map
self.axon_compartment_bases = compartment_bases
self.pop_type = 16 if pop_type is None else pop_type
self._set_weights_indices(weights, indices=indices, compression=d(b"MA==", int))
def set_learning(
self, learning_rate=1.0, tracing_tau=2, tracing_mag=1.0, wgt_exp=4
):
"""Set the learning parameters for this Synapse."""
assert tracing_tau == int(tracing_tau), "tracing_tau must be integer"
self.learning = True
self.tracing_tau = int(tracing_tau)
self.tracing_mag = tracing_mag
# stdp_cfg hard-coded for now (see hardware.builder)
self.format(learning_cfg=d(b"MQ==", int), stdp_cfg=d(b"MA==", int))
self.train_epoch = 2
self.learn_epoch_k = 1
self.learn_epoch = self.train_epoch * 2 ** self.learn_epoch_k
self.learning_rate = learning_rate * self.learn_epoch
self.learning_wgt_exp = wgt_exp
def set_full_weights(self, weights):
weights = np.array(weights, copy=False, dtype=np.float32)
assert weights.ndim in (1, 2)
if weights.ndim == 1:
indices = np.arange(weights.size)
else:
indices = None
self._set_weights_indices(weights, indices=indices)
assert len(self.weights) == self.n_axons, (
"Full weights must have different weights for each axon")
self.format(
compression=d(b'Mw==', int),
idx_bits=self.idx_bits(),
fanout_type=d(b'MQ==', int),
n_synapses=d(b'NjM=', int),
weight_bits=d(b'Nw==', int),
)
def _set_weights_indices(
self,
weights,
indices=None,
weight_dtype=np.float32,
compression=d(b"MA==", int),
):
weights = [
np.array(w, copy=False, dtype=weight_dtype, ndmin=2) for w in weights
]
assert all(
w.ndim == 2 for w in weights
), "Weights must be shape (n_axons,) (n_populations, n_compartments)"
assert all(
w.shape[0] == weights[0].shape[0] for w in weights
), "All axon weights must have the same number of populations"
self.weights = weights
if indices is None:
indices = [
np.zeros((w.shape[0], 1), dtype=np.int32)
+ np.arange(w.shape[1], dtype=np.int32)
for w in self.weights
]
indices = [np.array(i, copy=False, dtype=np.int32, ndmin=2) for i in indices]
assert all(
i.ndim == 2 for i in indices
), "Indices must be shape (n_axons,) (n_populations, n_compartments)"
assert all(
i.shape == w.shape for i, w in zip(indices, weights)
), "Indices shapes must match weights shapes"
assert len(weights) == len(indices)
self.indices = indices
self.format(
compression=compression,
idx_bits=self.idx_bits(),
fanout_type=d(b"MQ==", int),
n_synapses=d(b"NjM=", int),
weight_bits=d(b"Nw==", int),
)
def set_population_weights(
self,
weights,
indices,
axon_to_weight_map,
compartment_bases,
pop_type=None
):
self._set_weights_indices(weights, indices)
self.axon_to_weight_map = axon_to_weight_map
self.axon_compartment_bases = compartment_bases
self.pop_type = 16 if pop_type is None else pop_type
idx_bits = self.idx_bits()
self.format(
compression=d(b'MA==', int),
idx_bits=idx_bits,
fanout_type=d(b'MQ==', int),
n_synapses=d(b'NjM=', int),
weight_bits=d(b'Nw==', int),
)
def decay_int(x, decay, bits=None, offset=0, out=None):
"""Decay integer values using a decay constant.
The decayed value is given by::
sign(x) * floor(abs(x) * (2**bits - offset - decay) / 2**bits)
"""
if out is None:
out = np.zeros_like(x)
if bits is None:
bits = d(b'MTI=', int)
r = (2**bits - offset - np.asarray(decay)).astype(np.int64)
np.right_shift(np.abs(x) * r, bits, out=out)
return np.sign(x) * out
def validate_core(core):
# TODO: check these numbers are correct
assert len(core.compartment_cfgs) <= d(b'MzI=', int)
assert len(core.vth_cfgs) <= d(b'MTY=', int)
assert len(core.synapse_cfgs) <= d(b'MTY=', int)
assert len(core.stdp_pre_cfgs) <= d(b'Mw==', int)
for cfg in core.compartment_cfgs:
validate_compartment_cfg(cfg)
for cfg in core.vth_cfgs:
validate_vth_cfg(cfg, core=core)
for cfg in core.synapse_cfgs:
if cfg is not None:
validate_synapse_cfg(cfg)
for cfg in core.stdp_pre_cfgs:
validate_trace_cfg(cfg)
for synapse in core.synapse_axons:
cfg = core.get_synapse_cfg(synapse)
idxbits = cfg.real_idx_bits
for i in synapse.indices:
assert np.all(i >= 0)
assert np.all(i < 2**idxbits)
def set_weights(self, weights):
"""Set dense or sparse weights on this Synapse."""
if isinstance(weights, scipy.sparse.spmatrix):
csr = weights.tocsr()
weights_by_row, idxs_by_row = [], []
for i in range(weights.shape[0]):
i0, i1 = csr.indptr[i : i + 2]
weights_by_row.append(csr.data[i0:i1])
idxs_by_row.append(csr.indices[i0:i1])
weights = weights_by_row
indices = idxs_by_row
else:
weights = np.array(weights, copy=False, dtype=np.float32)
assert weights.ndim == 2
indices = None
assert len(weights) == self.n_axons, "Must have different weights for each axon"
self._set_weights_indices(weights, indices=indices, compression=d(b"Mw==", int))
def build_block(nxsdk_core, core, block, compartment_idxs, ax_range):
assert block.compartment.scale_u is False
assert block.compartment.scale_v is False
logger.debug("Building %s on core.id=%d", block, nxsdk_core.id)
for i, bias in enumerate(block.compartment.bias):
bman, bexp = bias_to_manexp(bias)
icomp = core.compartment_cfg_idxs[block][i]
ivth = core.vth_cfg_idxs[block][i]
ii = compartment_idxs[i]
d_func(d_get(nxsdk_core, b'Y3hDZmc=')[ii],
b'Y29uZmlndXJl',
kwargs={
b'Ymlhcw==': bman,
b'Ymlhc0V4cA==': bexp,
b'dnRoUHJvZmlsZQ==': ivth,
b'Y3hQcm9maWxl': icomp,
})
phasex = d(b'cGhhc2UlZA==') % (ii % 4, )
d_get(
d_get(nxsdk_core, b'Y3hNZXRhU3RhdGU=')[ii // 4],
b'Y29uZmlndXJl')(**{
phasex: 2
})
logger.debug("- Building %d synapses", len(block.synapses))
for synapse in block.synapses:
build_synapse(nxsdk_core, core, block, synapse, compartment_idxs)
logger.debug("- Building %d axons", len(block.axons))
pop_id_map = {}
for axon in block.axons:
build_axons(nxsdk_core, core, block, axon, compartment_idxs,
pop_id_map)
logger.debug("- Building %d probes", len(block.probes))
for probe in block.probes:
build_probe(nxsdk_core, core, block, probe, compartment_idxs)
def test_interface_connection_errors(Simulator, monkeypatch):
with nengo.Network() as net:
nengo.Ensemble(2, 1)
# test opening closed interface error
sim = Simulator(net)
interface = sim.sims["loihi"]
interface.close()
with pytest.raises(SimulationError, match="cannot be reopened"):
with interface:
pass
sim.close()
# test failed connection error
def start(*args, **kwargs):
raise Exception("Mock failure to connect")
monkeypatch.setattr(NxsdkBoard, d(b"c3RhcnQ="), start)
with pytest.raises(SimulationError, match="Mock failure to connect"):
with Simulator(net):
pass
def bits_per_axon(self, n_weights):
"""For an axon with n weights, compute the weight memory bits used"""
bits_per_weight = self.real_weight_bits + self.delay_bits + self.tag_bits
if self.compression == d(b"MA==", int):
bits_per_weight += self.real_idx_bits
elif self.compression == d(b"Mw==", int):
pass
else:
raise NotImplementedError("Compression %s" % (self.compression,))
synapse_idx_bits = d(b"NA==", int)
n_synapses_bits = d(b"Ng==", int)
bits = 0
synapses_per_block = self.n_synapses + 1
for i in range(0, n_weights, synapses_per_block):
n = min(n_weights - i, synapses_per_block)
bits_i = n * bits_per_weight + synapse_idx_bits + n_synapses_bits
# round up to nearest memory unit
bits_i = -d(b"NjQ=", int) * (-bits_i // d(b"NjQ=", int))
bits += bits_i
return bits
class Compartment:
"""Stores information for configuring Loihi compartments.
The information stored here will be associated with some block,
and all compartments will share certain information.
While compartments are usually thought of neurons, we use compartments
to implement Nengo ensembles, nodes, and connection through special
decode neurons.
Before `.discretize_compartment` has been called, most attributes in
this class are floating-point values. Calling `.discretize_compartment`
converts them to integer values inplace for use on Loihi.
Attributes
----------
bias : (n,) ndarray
Compartment biases.
enable_noise : (n,) ndarray
Whether to enable noise for each compartment.
decay_u : (n,) ndarray
Input (synapse) decay constant for each compartment.
decay_v : (n,) ndarray
Voltage decay constant for each compartment.
label : string
A label for the block (for debugging purposes).
n_compartments : int
The number of compartments in the block.
noise_at_membrane : {0, 1}
Inject noise into current (0) or voltage (1).
noise_exp : float or int
Exponent for noise generation. Floating point values are base 10
in units of current or voltage. Integer values are in base 2.
noise_offset : float or int
Offset for noise generation.
refract_delay : (n,) ndarray
Compartment refractory delays, in time steps.
scale_u : bool
Scale input (U) by decay_u so that the integral of U is
the same before and after filtering.
scale_v : bool
Scale voltage (V) by decay_v so that the integral of V is
the same before and after filtering.
tau_s : float or None
Time constant used to set decay_u. None if decay_u has not been set.
vmax : float or int (range [2**9 - 1, 2**23 - 1])
Maximum voltage for all compartments, in the same units as ``vth``.
vmin : float or int (range [-2**23 + 1, 0])
Minimum voltage for all compartments, in the same units as ``vth``.
vth : (n,) ndarray
Compartment voltage thresholds.
"""
# threshold at which U/V scaling is allowed
DECAY_SCALE_TH = 0.5 / d(b'NDA5Ng==', int) # half of decay scaling unit
def __init__(self, n_compartments, label=None):
self.n_compartments = n_compartments
self.label = label
# parameters specific to compartments/block
self.decay_u = np.ones(n_compartments, dtype=np.float32)
# ^ default to no filter
self.decay_v = np.zeros(n_compartments, dtype=np.float32)
# ^ default to integration
self.tau_s = None
self.scale_u = True
self.scale_v = False
self.refract_delay = np.zeros(n_compartments, dtype=np.int32)
self.vth = np.zeros(n_compartments, dtype=np.float32)
self.bias = np.zeros(n_compartments, dtype=np.float32)
self.enable_noise = np.zeros(n_compartments, dtype=bool)
# parameters common to core
self.vmin = 0
self.vmax = np.inf
self.noise_offset = 0
self.noise_exp = 0
self.noise_at_membrane = 0
def __str__(self):
return "%s(%s)" % (
type(self).__name__, self.label if self.label else '')
def configure_default_filter(self, tau_s, dt=0.001):
"""Set the default Lowpass synaptic input filter for compartments.
Parameters
----------
tau_s : float
`nengo.Lowpass` synapse time constant for filtering.
dt : float
Simulator time step.
"""
if self.tau_s is None: # don't overwrite a non-default filter
self._configure_filter(tau_s, dt=dt)
def configure_filter(self, tau_s, dt=0.001):
"""Set Lowpass synaptic input filter for compartments.
Parameters
----------
tau_s : float
`nengo.Lowpass` synapse time constant for filtering.
dt : float
Simulator time step.
"""
if self.tau_s is not None and tau_s < self.tau_s:
warnings.warn("tau_s is already set to %g, which is larger than "
"%g. Using %g." % (self.tau_s, tau_s, self.tau_s))
return
elif self.tau_s is not None and tau_s > self.tau_s:
warnings.warn(
"tau_s is currently %g, which is smaller than %g. Overwriting "
"tau_s with %g." % (self.tau_s, tau_s, tau_s))
self._configure_filter(tau_s, dt=dt)
self.tau_s = tau_s
def _configure_filter(self, tau_s, dt):
decay_u = 1 if tau_s == 0 else -np.expm1(-dt/np.asarray(tau_s))
self.decay_u[:] = decay_u
self.scale_u = decay_u > self.DECAY_SCALE_TH
if not self.scale_u:
raise BuildError(
"Current (U) scaling is required. Perhaps a synapse time "
"constant is too large in your model.")
def configure_lif(self, tau_rc=0.02, tau_ref=0.001, vth=1, dt=0.001,
min_voltage=0):
"""Configure these compartments as individual LIF neurons.
Parameters
----------
tau_rc : float
Membrane time constant (in seconds) of the neurons.
tau_ref : float
Refractory period (in seconds) of the neurons.
vth : float
Voltage firing threshold of the neurons.
dt : float
Simulator time step length (in seconds).
min_voltage : float
The minimum voltage for the neurons.
"""
self.decay_v[:] = -np.expm1(-dt / np.asarray(tau_rc))
self.refract_delay[:] = np.round(tau_ref / dt) + 1
self.vth[:] = vth
self.vmin = min_voltage
self.vmax = np.inf
self.scale_v = np.all(self.decay_v > self.DECAY_SCALE_TH)
if not self.scale_v:
raise BuildError(
"Voltage (V) scaling is required with LIF neurons. Perhaps "
"the neuron tau_rc time constant is too large.")
def configure_nonspiking(self, tau_ref=0.0, vth=1, dt=0.001):
"""Configure these compartments as individual non-spiking neurons.
Parameters
----------
tau_ref : float
Refractory period (in seconds) of the neurons.
vth : float
Voltage firing threshold of the neurons.
dt : float
Simulator time step length (in seconds).
"""
self.decay_v[:] = 1.
self.refract_delay[:] = 1
self.vth[:] = vth
self.vmin = 0
self.vmax = np.inf
self.scale_v = False
def configure_relu(self, tau_ref=0.0, vth=1, dt=0.001):
"""Configure these compartments as individual Rectified Linear neurons.
These are also known as non-leaky integrate-and-fire neurons. The
voltage is the integral of the input current.
Parameters
----------
tau_ref : float
Refractory period (in seconds) of the neurons.
vth : float
Voltage firing threshold of the neurons.
dt : float
Simulator time step length (in seconds).
"""
self.decay_v[:] = 0.
self.refract_delay[:] = np.round(tau_ref / dt) + 1
self.vth[:] = vth
self.vmin = 0
self.vmax = np.inf
self.scale_v = False
def axon_bits(self):
if self.pop_type == 16:
return d(b'MTA=', int) - self.atom_bits_extra()
else:
return d(b'MTI=', int)
def atom_bits_extra(self):
atom_bits = self.atom_bits()
assert atom_bits <= d(b'OQ==', int), "Too many atom bits"
return max(atom_bits - d(b'NQ==', int), 0)
def idxs_per_synapse(self):
return d(b'Mg==', int) if self.learning else d(b'MQ==', int)