def test_cont2discrete_zoh(dt, allclose):
"""test the function cont2discrete with zero-order hold"""
taus = np.logspace(-np.log10(dt) - 1, np.log10(dt) + 3, 30)
# test with lowpass filter, using analytic solution
for tau in taus:
num, den = [1], [tau, 1]
d = -np.expm1(-dt / tau)
num0, den0 = [0, d], [1, d - 1]
num1, den1, _ = cont2discrete((num, den), dt)
assert allclose(num0, num1)
assert allclose(den0, den1)
# test with alpha filter, using analytic solution
for tau in taus:
num, den = [1], [tau**2, 2 * tau, 1]
a = dt / tau
ea = np.exp(-a)
num0 = [0, -a * ea + (1 - ea), ea * (a + ea - 1)]
den0 = [1, -2 * ea, ea**2]
num1, den1, _ = cont2discrete((num, den), dt)
assert allclose(num0, num1)
assert allclose(den0, den1)
# test integrative filter, using analytic solution
num, den = [1], [1, 0]
num0, den0 = [0, dt], [1, -1]
num1, den1, _ = cont2discrete((num, den), dt)
assert allclose(num0, num1)
assert allclose(den0, den1)
def test_cont2discrete_zoh(dt):
taus = np.logspace(-np.log10(dt) - 1, np.log10(dt) + 3, 30)
# test with lowpass filter, using analytic solution
for tau in taus:
num, den = [1], [tau, 1]
d = -np.expm1(-dt / tau)
num0, den0 = [0, d], [1, d - 1]
num1, den1, _ = cont2discrete((num, den), dt)
assert np.allclose(num0, num1)
assert np.allclose(den0, den1)
# test with alpha filter, using analytic solution
for tau in taus:
num, den = [1], [tau**2, 2*tau, 1]
a = dt / tau
ea = np.exp(-a)
num0 = [0, -a*ea + (1 - ea), ea*(a + ea - 1)]
den0 = [1, -2 * ea, ea**2]
num1, den1, _ = cont2discrete((num, den), dt)
assert np.allclose(num0, num1)
assert np.allclose(den0, den1)
# test integrative filter, using analytic solution
num, den = [1], [1, 0]
num0, den0 = [0, dt], [1, -1]
num1, den1, _ = cont2discrete((num, den), dt)
assert np.allclose(num0, num1)
assert np.allclose(den0, den1)
def build_filter(model, synapse, owner, input_signal):
num, den, _ = cont2discrete(
(synapse.num, synapse.den), model.dt, method='zoh')
num = num.flatten()
num = num[1:] if num[0] == 0 else num
den = den[1:] # drop first element (equal to 1)
build_discrete_filter(model, synapse, owner, input_signal, num, den)
def make_step(self,
shape_in,
shape_out,
dt,
rng,
y0=None,
dtype=np.float64,
method='zoh'):
"""Returns a `.Step` instance that implements the linear filter."""
assert shape_in == shape_out
num, den = self.num, self.den
if self.analog:
num, den, _ = cont2discrete((num, den), dt, method=method)
num = num.flatten()
if den[0] != 1.:
raise ValidationError("First element of the denominator must be 1",
attr='den',
obj=self)
num = num[1:] if num[0] == 0 else num
den = den[1:] # drop first element (equal to 1)
num, den = num.astype(dtype), den.astype(dtype)
output = np.zeros(shape_out, dtype=dtype)
if len(num) == 1 and len(den) == 0:
return LinearFilter.NoDen(num, den, output)
elif len(num) == 1 and len(den) == 1:
return LinearFilter.Simple(num, den, output, y0=y0)
return LinearFilter.General(num, den, output, y0=y0)
def test_combined_delay(Simulator, allclose):
# ensure that two sequential filters has the same output
# as combining their individual discrete transfer functions
nt = 50
tau = 0.005
dt = 0.001
sys1 = nengo.Lowpass(tau)
(num,), den, _ = cont2discrete((sys1.num, sys1.den), dt=dt)
sys2 = nengo.LinearFilter(np.poly1d(num) ** 2, np.poly1d(den) ** 2, analog=False)
with nengo.Network() as model:
u = nengo.Node(1)
x = nengo.Node(size_in=1)
nengo.Connection(u, x, synapse=sys1)
p1 = nengo.Probe(x, synapse=sys1)
p2 = nengo.Probe(u, synapse=sys2)
with Simulator(model, dt=dt) as sim:
sim.run_steps(nt)
assert allclose(sim.data[p1], sim.data[p2])
# Both have two time-step delays:
# for sys1, this comes from two levels of filtering
# for sys2, this comes from one level of filtering + delay in sys2
assert allclose(sim.data[p1][:2], 0)
assert not allclose(sim.data[p1][2], 0, record_rmse=False)
def _get_ss(self, dt):
A, B, C, D = tf2ss(self.num, self.den)
# discretize (if len(A) == 0, filter is stateless and already discrete)
if self.analog and len(A) > 0:
A, B, C, D, _ = cont2discrete((A, B, C, D), dt, method=self.method)
return A, B, C, D
def __init__(self,
vocab,
gain=5,
deriv_synapse=.1,
theta=.05,
order=30,
dt=.001,
**kwargs):
super().__init__(**kwargs)
self.vocab = vocab
# parameters of LMU
theta = theta # length of window
order = order # number of Legendre polynomials
dt = dt # simulation timestep
Q = np.arange(order, dtype=np.float64)
R = (2 * Q + 1)[:, None] / theta
j, i = np.meshgrid(Q, Q)
A = np.where(i < j, -1, (-1.0)**(i - j + 1)) * R
B = (-1.0)**Q[:, None] * R
C = np.ones((1, order))
D = np.zeros((1, ))
A, B, C, D, _ = cont2discrete((A, B, C, D), dt=dt, method="zoh")
with self:
self.input = nengo.Node(size_in=1)
self.deriv = nengo.Ensemble(100,
1,
intercepts=nengo.dists.Uniform(
0.1, .9),
encoders=[[1]] * 100)
gain = 5
nengo.Connection(self.input, self.deriv, transform=gain)
nengo.Connection(self.input,
self.deriv,
transform=-gain,
synapse=deriv_synapse)
self.lmu = nengo.Node(size_in=order)
nengo.Connection(self.deriv, self.lmu, transform=B, synapse=None)
nengo.Connection(self.lmu, self.lmu, transform=A, synapse=0)
self.output = nengo.Node(size_in=2)
nengo.Connection(self.deriv, self.output[0], synapse=None)
nengo.Connection(self.lmu,
self.output[1],
transform=C,
synapse=None)
self.declare_input(self.input, None)
self.declare_output(self.output, None)
def test_cont2discrete():
import scipy.signal
dt = 1e-3
tau = 0.03
num, den = [1], [tau**2, 2 * tau, 1]
num0, den0, _ = scipy.signal.cont2discrete((num, den), dt)
num1, den1, _ = cont2discrete((num, den), dt)
assert np.allclose(num0, num1)
assert np.allclose(den0, den1)
def test_cont2discrete():
import scipy.signal
dt = 1e-3
tau = 0.03
num, den = [1], [tau**2, 2*tau, 1]
num0, den0, _ = scipy.signal.cont2discrete((num, den), dt)
num1, den1, _ = cont2discrete((num, den), dt)
assert np.allclose(num0, num1)
assert np.allclose(den0, den1)
def pack_data(self, dt, buffer, offset=0):
"""Pack the struct describing the filter into the buffer."""
# Compute the filter coefficients
b, a, _ = cont2discrete((self.num, self.den), dt)
b = b.flatten()
# Strip out the first values
# `a` is negated so that it can be used with a multiply-accumulate
# instruction on chip.
assert b[0] == 0.0 # Oops!
ab = np.vstack((-a[1:], b[1:])).T.flatten()
# Convert the values to fixpoint and write into a data buffer
struct.pack_into("<I{}s".format(self.order * 2 * 4), buffer, offset,
self.order, tp.np_to_fix(ab).tostring())
def pack_data(self, dt, buffer, offset=0):
"""Pack the struct describing the filter into the buffer."""
# Compute the filter coefficients
b, a, _ = cont2discrete((self.num, self.den), dt)
b = b.flatten()
# Strip out the first values
# `a` is negated so that it can be used with a multiply-accumulate
# instruction on chip.
assert b[0] == 0.0 # Oops!
ab = np.vstack((-a[1:], b[1:])).T.flatten()
# Convert the values to fixpoint and write into a data buffer
struct.pack_into("<I", buffer, offset, self.order)
buffer[offset + 4:4+self.order*8] = tp.np_to_fix(ab).tostring()
def make_step(self, dt, output, method='zoh'):
num, den = self.num, self.den
if self.analog:
num, den, _ = cont2discrete((num, den), dt, method=method)
num = num.flatten()
if den[0] != 1.:
raise ValueError("First element of the denominator must be 1")
num = num[1:] if num[0] == 0 else num
den = den[1:] # drop first element (equal to 1)
if len(num) == 1 and len(den) == 0:
return LinearFilter.NoDen(num, den, output)
elif len(num) == 1 and len(den) == 1:
return LinearFilter.Simple(num, den, output)
return LinearFilter.General(num, den, output)
def make_step(self, dt, output, method='zoh'):
num, den = self.num, self.den
if self.analog:
num, den, _ = cont2discrete((num, den), dt, method=method)
num = num.flatten()
if den[0] != 1.:
raise ValueError("First element of the denominator must be 1")
num = num[1:] if num[0] == 0 else num
den = den[1:] # drop first element (equal to 1)
if len(num) == 1 and len(den) == 0:
return LinearFilter.NoDen(num, den, output)
elif len(num) == 1 and len(den) == 1:
return LinearFilter.Simple(num, den, output)
return LinearFilter.General(num, den, output)
def make_step(self, dt, output, method='zoh'):
num, den, _ = cont2discrete((self.num, self.den), dt, method=method)
num = num.flatten()
num = num[1:] if num[0] == 0 else num
den = den[1:] # drop first element (equal to 1)
if len(num) == 1 and len(den) == 0:
return functools.partial(
LinearFilter.no_den_step, output=output, b=num[0])
elif len(num) == 1 and len(den) == 1:
return functools.partial(
LinearFilter.simple_step, output=output, a=den[0], b=num[0])
else:
x = collections.deque(maxlen=len(num))
y = collections.deque(maxlen=len(den))
return functools.partial(LinearFilter.general_step,
output=output, x=x, y=y, num=num, den=den)
def make_step(self, dt, output, method='zoh'):
num, den, _ = cont2discrete((self.num, self.den), dt, method=method)
num = num.flatten()
num = num[1:] if num[0] == 0 else num
den = den[1:] # drop first element (equal to 1)
if len(num) == 1 and len(den) == 0:
return functools.partial(
LinearFilter.no_den_step, output=output, b=num[0])
elif len(num) == 1 and len(den) == 1:
return functools.partial(
LinearFilter.simple_step, output=output, a=den[0], b=num[0])
else:
x = collections.deque(maxlen=len(num))
y = collections.deque(maxlen=len(den))
return functools.partial(LinearFilter.general_step,
output=output, x=x, y=y, num=num, den=den)
def lti(u, system, state=lambda x: x, dt=0.001, method="zoh"):
r"""Operator that solves \dot{x} = A.state(x) + B.u. where A, B = system.
The state parameter can be any callable function that consumes a Stimulus operator
and produces a Gyrus operator that consumes said operator as an input. For instance,
nonlinear dynamical systems may be implemented by specifying a nonlinear function
for the ``state``.
"""
if not is_iterable(system) or len(system) != 2:
raise NotImplementedError(
f"lti currently only supports systems as two-tuples: (A, B); not {system}"
)
# Reshape and validate A, B matrices.
A, B = system
A = np.asarray(A)
if A.ndim != 2 or A.shape[0] != A.shape[1]:
raise ValueError(f"A ({A}) must be a square matrix")
size_out = A.shape[0]
B = np.asarray(B)
if B.ndim == 1:
B = B[:, None]
if B.ndim != 2:
raise ValueError(f"B ({B}) must be 1D or 2D, but is {B.ndim}")
if B.shape[0] != size_out:
raise ValueError(
f"B ({B}) must be an array of length {size_out}, not {B.shape[0]}")
C = np.zeros_like(B).T
D = 0
# Discretize the dynamical system, \dot{x} = Ax + Bu, to
# x[t + dt] = Ax[t] + Bu[t], using some method (ZOH recommended for stability).
Abar, Bbar, _, _, _ = cont2discrete((A, B, C, D), dt=dt, method=method)
# Apply Voelker (2019) equation 5.30 with H(z) = dt / (z - 1).
# This compensates for the discretized integrator such that the resulting
# system is the one that was requested. In this particular case (with the
# synapse being the discretized integrator) this reduces to the inverse
# of Euler's method.
Amap = (Abar - np.eye(len(Abar))) / dt
Bmap = Bbar / dt
# Finally express the Amap, Bmap system using vectorized Gyrus operators.
return u.transform(Bmap).integrate(
integrand=lambda x: state(x).transform(Amap))
def write_spec(self, spec, dt, width):
BasicFilterImpl.write_basic_spec(self, spec, width)
"""Pack the struct describing the filter into the buffer."""
# Compute the filter coefficients
b, a, _ = cont2discrete((self.num, self.den), dt)
b = b.flatten()
# Strip out the first values
# `a` is negated so that it can be used with a multiply-accumulate
# instruction on chip.
assert b[0] == 0.0 # Oops!
ab = numpy.vstack((-a[1:], b[1:])).T.flatten()
# Convert the values to fixpoint and write into a data buffer
print "other {}".format(self._other)
spec.write_value(self._order)
print "ab {}".format(ab)
spec.write_array(helpful_functions.convert_numpy_array_to_s16_15(ab))
def __init__(self, ops, signals, config):
super().__init__(ops, signals, config)
# the main difference between this and the general linearfilter
# OneX implementation is that this version allows us to merge
# synapses with different input dimensionality (by duplicating
# the synapse parameters for every input, rather than using
# broadcasting)
self.input_data = signals.combine([op.input for op in ops])
self.output_data = signals.combine([op.output for op in ops])
nums = []
dens = []
for op in ops:
if op.process.tau <= 0.03 * signals.dt_val:
num = 1
den = 0
else:
num, den, _ = cont2discrete((op.process.num, op.process.den),
signals.dt_val,
method="zoh")
num = num.flatten()
num = num[1:] if num[0] == 0 else num
assert len(num) == 1
num = num[0]
assert len(den) == 2
den = den[1]
nums += [num] * op.input.shape[0]
dens += [den] * op.input.shape[0]
if self.input_data.minibatched:
# add batch dimension for broadcasting
nums = np.expand_dims(nums, 0)
dens = np.expand_dims(dens, 0)
# apply the negative here
dens = -np.asarray(dens)
self.nums = tf.constant(nums, dtype=self.output_data.dtype)
self.dens = tf.constant(dens, dtype=self.output_data.dtype)
def __init__(self, units, order, theta, **kwargs):
super().__init__(**kwargs)
self.units = units
self.order = order
self.theta = theta
Q = np.arange(order, dtype=np.float64)
R = (2 * Q + 1)[:, None] / theta
j, i = np.meshgrid(Q, Q)
A = np.where(i < j, -1, (-1.0) ** (i - j + 1)) * R
B = (-1.0) ** Q[:, None] * R
C = np.ones((1, order))
D = np.zeros((1,))
self._A, self._B, _, _, _ = cont2discrete(
(A, B, C, D), dt=1.0, method="zoh"
)
def __init__(self, ops, signals, config):
super(LowpassBuilder, self).__init__(ops, signals, config)
self.input_data = signals.combine([op.input for op in ops])
self.output_data = signals.combine([op.output for op in ops])
nums = []
dens = []
for op in ops:
if op.process.tau <= 0.03 * signals.dt_val:
num = 1
den = 0
else:
num, den, _ = cont2discrete((op.process.num, op.process.den),
signals.dt_val,
method="zoh")
num = num.flatten()
num = num[1:] if num[0] == 0 else num
assert len(num) == 1
num = num[0]
assert len(den) == 2
den = den[1]
nums += [num] * op.input.shape[0]
dens += [den] * op.input.shape[0]
nums = np.asarray(nums)
while nums.ndim < len(self.input_data.full_shape):
nums = np.expand_dims(nums, -1)
# note: applying the negative here
dens = -np.asarray(dens)
while dens.ndim < len(self.input_data.full_shape):
dens = np.expand_dims(dens, -1)
# need to manually broadcast for scatter_mul
# dens = np.tile(dens, (1, signals.minibatch_size))
self.nums = signals.constant(nums, dtype=self.output_data.dtype)
self.dens = signals.constant(dens, dtype=self.output_data.dtype)
def __init__(self, ops, signals, config):
super(LowpassBuilder, self).__init__(ops, signals, config)
self.input_data = signals.combine([op.input for op in ops])
self.output_data = signals.combine([op.output for op in ops])
nums = []
dens = []
for op in ops:
if op.process.tau <= 0.03 * signals.dt_val:
num = 1
den = 0
else:
num, den, _ = cont2discrete((op.process.num, op.process.den),
signals.dt_val, method="zoh")
num = num.flatten()
num = num[1:] if num[0] == 0 else num
assert len(num) == 1
num = num[0]
assert len(den) == 2
den = den[1]
nums += [num] * op.input.shape[0]
dens += [den] * op.input.shape[0]
nums = np.asarray(nums)
while nums.ndim < len(self.input_data.full_shape):
nums = np.expand_dims(nums, -1)
# note: applying the negative here
dens = -np.asarray(dens)
while dens.ndim < len(self.input_data.full_shape):
dens = np.expand_dims(dens, -1)
# need to manually broadcast for scatter_mul
# dens = np.tile(dens, (1, signals.minibatch_size))
self.nums = signals.constant(nums, dtype=self.output_data.dtype)
self.dens = signals.constant(dens, dtype=self.output_data.dtype)
def make_step(self, shape_in, shape_out, dt, rng, y0=None, dtype=np.float64, method="zoh"):
"""Returns a `.Step` instance that implements the linear filter."""
assert shape_in == shape_out
num, den = self.num, self.den
if self.analog:
num, den, _ = cont2discrete((num, den), dt, method=method)
num = num.flatten()
if den[0] != 1.0:
raise ValidationError("First element of the denominator must be 1", attr="den", obj=self)
num = num[1:] if num[0] == 0 else num
den = den[1:] # drop first element (equal to 1)
num, den = num.astype(dtype), den.astype(dtype)
output = np.zeros(shape_out, dtype=dtype)
if len(num) == 1 and len(den) == 0:
return LinearFilter.NoDen(num, den, output)
elif len(num) == 1 and len(den) == 1:
return LinearFilter.Simple(num, den, output, y0=y0)
return LinearFilter.General(num, den, output, y0=y0)
def test_pack_data(self, num, den, dt, order):
# Create the filter
lf = LinearFilter(0, False, num, den)
# Create a buffer to pack data into
data = bytearray((order * 2 + 1) * 4)
# Pack the parameters
lf.pack_data(dt, data, 0)
# Generate what we expect the data to look like
numd, dend, _ = cont2discrete((num, den), dt)
numd = numd.flatten()
exp = list()
for a, b in zip(dend[1:], numd[1:]):
exp.append(-a)
exp.append(b)
expected_data = tp.np_to_fix(np.array(exp)).tostring()
# Check that's what we get
assert struct.unpack_from("<I", data, 0)[0] == order
assert data[4:] == expected_data
def __init__(self, ops, signals):
# TODO: implement general linear filter (using queues?)
self.input_data = (None if ops[0].input is None else signals.combine(
[op.input for op in ops]))
self.output_data = signals.combine([op.output for op in ops])
nums = []
dens = []
for op in ops:
if op.process.tau <= 0.03 * signals.dt_val:
num = 1
den = 0
else:
num, den, _ = cont2discrete((op.process.num, op.process.den),
signals.dt_val,
method="zoh")
num = num.flatten()
num = num[1:] if num[0] == 0 else num
assert len(num) == 1
num = num[0]
assert len(den) == 2
den = den[1]
nums += [num] * op.input.shape[0]
dens += [den] * op.input.shape[0]
nums = np.asarray(nums)[:, None]
# note: applying the negative here
dens = -np.asarray(dens)[:, None]
# need to manually broadcast for scatter_mul
# dens = np.tile(dens, (1, signals.minibatch_size))
self.nums = tf.constant(nums, dtype=self.output_data.dtype)
self.dens = tf.constant(dens, dtype=self.output_data.dtype)
def test_pack_data(self, num, den, dt, order):
# Create the filter
lf = LinearFilter(0, False, num, den)
# Create a buffer to pack data into
data = bytearray((order*2 + 1)*4)
# Pack the parameters
lf.pack_data(dt, data, 0)
# Generate what we expect the data to look like
numd, dend, _ = cont2discrete((num, den), dt)
numd = numd.flatten()
exp = list()
for a, b in zip(dend[1:], numd[1:]):
exp.append(-a)
exp.append(b)
expected_data = tp.np_to_fix(np.array(exp)).tostring()
# Check that's what we get
assert struct.unpack_from("<I", data, 0)[0] == order
assert data[4:] == expected_data
def test_cont2discrete_other_methods():
dt = 1e-3
# test with len(sys) == 3
assert (repr(cont2discrete(
([1], [1], [1]),
dt)) == "(array([1.0010005]), array([1.0010005]), 1.0, 0.001)")
# test with len(sys) == 5
with pytest.raises(ValueError):
cont2discrete(([1], [1], [1], [1], [1]), dt)
# test method gbt and alpha None
with pytest.raises(ValueError):
cont2discrete(([1], [1], [1]), dt, method="gbt", alpha=None)
# test method gbt and alpha invalid
with pytest.raises(ValueError):
cont2discrete(([1], [1], [1]), dt, method="gbt", alpha=2)
# test method bilinear
assert (repr(cont2discrete(([1], [1], [1]), dt, method="bilinear")) ==
"(array([1.0010005]), array([1.0010005]), 1.0, 0.001)")
# test method backward_diff
assert (repr(cont2discrete(([1], [1], [1]), dt, method="backward_diff")) ==
"(array([1.001001]), array([1.001001]), 1.0, 0.001)")
# test bad method
with pytest.raises(ValueError):
cont2discrete(([1], [1], [1]), dt, method="not_a_method")
def __init__(self, ops, signals, config):
super(LinearFilterBuilder, self).__init__(ops, signals, config)
self.input_data = signals.combine([op.input for op in ops])
self.output_data = signals.combine([op.output for op in ops])
self.n_ops = len(ops)
self.signal_d = ops[0].input.shape[0]
As = []
Cs = []
Ds = []
# compute the A/C/D matrices for each operator
for op in ops:
A, B, C, D = tf2ss(op.process.num, op.process.den)
if op.process.analog:
# convert to discrete system
A, B, C, D, _ = cont2discrete((A, B, C, D), signals.dt_val,
method="zoh")
# convert to controllable form
num, den = ss2tf(A, B, C, D)
if op.process.analog:
# add shift
num = np.concatenate((num, [[0]]), axis=1)
with warnings.catch_warnings():
# ignore the warning about B, since we aren't using it anyway
warnings.simplefilter("ignore", BadCoefficients)
A, _, C, D = tf2ss(num, den)
As.append(A)
Cs.append(C[0])
Ds.append(D.item())
self.state_d = sum(x.shape[0] for x in Cs)
# build a sparse matrix containing the A matrices as blocks
# along the diagonal
sparse_indices = []
corner = np.zeros(2, dtype=np.int64)
for A in As:
idxs = np.reshape(np.dstack(np.meshgrid(
np.arange(A.shape[0]), np.arange(A.shape[1]),
indexing="ij")), (-1, 2))
idxs += corner
corner += A.shape
sparse_indices += [idxs]
sparse_indices = np.concatenate(sparse_indices, axis=0)
self.A = signals.constant(np.concatenate(As, axis=0).flatten(),
dtype=signals.dtype)
self.A_indices = signals.constant(sparse_indices, dtype=(
tf.int32 if np.all(sparse_indices < np.iinfo(np.int32).max)
else tf.int64))
self.A_shape = tf.constant(corner, dtype=tf.int64)
if np.allclose(Cs, 0):
self.C = None
else:
# add empty dimension for broadcasting
self.C = signals.constant(np.concatenate(Cs)[:, None],
dtype=signals.dtype)
if np.allclose(Ds, 0):
self.D = None
else:
# add empty dimension for broadcasting
self.D = signals.constant(np.asarray(Ds)[:, None],
dtype=signals.dtype)
self.offsets = tf.expand_dims(
tf.range(0, len(ops) * As[0].shape[0], As[0].shape[0]),
axis=1)
# create a variable to represent the internal state of the filter
self.state_sig = signals.make_internal(
"state", (self.state_d, signals.minibatch_size * self.signal_d),
minibatched=False)
def __init__(self, ops, signals, config):
super(LinearFilterBuilder, self).__init__(ops, signals, config)
self.input_data = signals.combine([op.input for op in ops])
self.output_data = signals.combine([op.output for op in ops])
self.n_ops = len(ops)
self.signal_d = ops[0].input.shape[0]
As = []
Cs = []
Ds = []
# compute the A/C/D matrices for each operator
for op in ops:
A, B, C, D = tf2ss(op.process.num, op.process.den)
if op.process.analog:
# convert to discrete system
A, B, C, D, _ = cont2discrete((A, B, C, D),
signals.dt_val,
method="zoh")
# convert to controllable form
num, den = ss2tf(A, B, C, D)
if op.process.analog:
# add shift
num = np.concatenate((num, [[0]]), axis=1)
with warnings.catch_warnings():
# ignore the warning about B, since we aren't using it anyway
warnings.simplefilter("ignore", BadCoefficients)
A, _, C, D = tf2ss(num, den)
As.append(A)
Cs.append(C[0])
Ds.append(D.item())
self.state_d = sum(x.shape[0] for x in Cs)
# build a sparse matrix containing the A matrices as blocks
# along the diagonal
sparse_indices = []
corner = np.zeros(2, dtype=np.int64)
for A in As:
idxs = np.reshape(
np.dstack(
np.meshgrid(np.arange(A.shape[0]),
np.arange(A.shape[1]),
indexing="ij")), (-1, 2))
idxs += corner
corner += A.shape
sparse_indices += [idxs]
sparse_indices = np.concatenate(sparse_indices, axis=0)
self.A = signals.constant(np.concatenate(As, axis=0).flatten(),
dtype=signals.dtype)
self.A_indices = signals.constant(
sparse_indices,
dtype=(tf.int32 if np.all(
sparse_indices < np.iinfo(np.int32).max) else tf.int64))
self.A_shape = tf.constant(corner, dtype=tf.int64)
if np.allclose(Cs, 0):
self.C = None
else:
# add empty dimension for broadcasting
self.C = signals.constant(np.concatenate(Cs)[:, None],
dtype=signals.dtype)
if np.allclose(Ds, 0):
self.D = None
else:
# add empty dimension for broadcasting
self.D = signals.constant(np.asarray(Ds)[:, None],
dtype=signals.dtype)
self.offsets = tf.expand_dims(tf.range(0,
len(ops) * As[0].shape[0],
As[0].shape[0]),
axis=1)
# create a variable to represent the internal state of the filter
self.state_sig = signals.make_internal(
"state", (self.state_d, signals.minibatch_size * self.signal_d),
minibatched=False)
def __init__(self, units, order, theta, input_d, **kwargs):
super().__init__(**kwargs)
# compute the A and B matrices according to the LMU's mathematical
# derivation (see the paper for details)
Q = np.arange(order, dtype=np.float64)
R = (2 * Q + 1)[:, None] / theta
j, i = np.meshgrid(Q, Q)
A = np.where(i < j, -1, (-1.0) ** (i - j + 1)) * R
B = (-1.0) ** Q[:, None] * R
C = np.ones((1, order))
D = np.zeros((1,))
A, B, _, _, _ = cont2discrete((A, B, C, D), dt=1.0, method="zoh")
with self:
nengo_dl.configure_settings(trainable=None)
# create objects corresponding to the x/u/m/h variables in LMU
self.x = nengo.Node(size_in=input_d)
self.u = nengo.Node(size_in=1)
self.m = nengo.Node(size_in=order)
self.h = nengo_dl.TensorNode(
tf.nn.tanh, shape_in=(units,), pass_time=False
)
# compute u_t
# note that setting synapse=0 (versus synapse=None) adds a
# one-timestep delay, so we can think of any connections with
# synapse=0 as representing value_{t-1}
nengo.Connection(
self.x, self.u, transform=np.ones((1, input_d)), synapse=None
)
nengo.Connection(
self.h, self.u, transform=np.zeros((1, units)), synapse=0
)
nengo.Connection(
self.m, self.u, transform=np.zeros((1, order)), synapse=0
)
# compute m_t
# in this implementation we'll make A and B non-trainable, but they
# could also be optimized in the same way as the other parameters
conn = nengo.Connection(self.m, self.m, transform=A, synapse=0)
self.config[conn].trainable = False
conn = nengo.Connection(self.u, self.m, transform=B, synapse=None)
self.config[conn].trainable = False
# compute h_t
nengo.Connection(
self.x,
self.h,
transform=np.zeros((units, input_d)),
synapse=None,
)
nengo.Connection(
self.h, self.h, transform=np.zeros((units, units)), synapse=0
)
nengo.Connection(
self.m,
self.h,
transform=nengo_dl.dists.Glorot(distribution="normal"),
synapse=None,
)
