def step_conv3(t, x):
x = x.reshape(shape_in)
nk, ni, nj = shape_in[-3:]
f = filters.shape[0]
sk, si, sj = filters.shape[-3:]
si2 = (si - 1) / 2
sj2 = (sj - 1) / 2
sk2 = (sk - 1) / 2
y = np.zeros(shape_out)
for k in range(nk):
for i in range(ni):
for j in range(nj):
i0, i1 = i - si2, i + si2 + 1
j0, j1 = j - sj2, j + sj2 + 1
k0, k1 = k - sk2, k + sk2 + 1
sli = slice(max(-i0, 0), min(ni + si - i1, si))
slj = slice(max(-j0, 0), min(nj + sj - j1, sj))
slk = slice(max(-k0, 0), min(nk + sk - k1, sk))
w = (filters[:, i, j, :, sli, slj] if local_filters else
filters[:, :,slk, sli, slj])
xkij = x[:,max(k0, 0):min(k1, nk), max(i0, 0):min(i1, ni), max(j0, 0):min(j1, nj)]
y[:,k, i, j] = np.dot(xkij.ravel(), w.reshape(f, -1).T)
if biases is not None:
y += biases
return y.ravel()
def cooccurr_ix(l, th=0.001):
ix = []
for i in range(len(l)):
for j in range(i+1, len(l)):
if abs(l[i] - l[j]) < th:
ix.append((i, j))
return ix
def get_syllables(n_syllables, minfreq, maxfreq, rng=np.random):
allpaths = []
for gdir in ['ges-de-ccv', 'ges-de-cv', 'ges-de-cvc', 'ges-de-v']:
allpaths.extend([ges_path(gdir, gfile)
for gfile in os.listdir(ges_path(gdir))])
indices = rng.permutation(len(allpaths))
return ([allpaths[indices[i]] for i in range(n_syllables)],
[rng.uniform(minfreq, maxfreq) for i in range(n_syllables)])
def cd_encoders_biases(n_encoders, trainX, trainY, rng=np.random, mask=None,
norm_min=0.05, norm_tries=10):
"""Constrained difference (CD) method for encoders from data [1]_.
Parameters
==========
n_encoders : int
Number of encoders to generate.
trainX : (n_samples, n_dimensions) array-like
Training features.
trainY : (n_samples,) array-like
Training labels.
Returns
=======
encoders : (n_encoders, n_dimensions) array
Generated encoders.
biases : (n_encoders,) array
Generated biases. These are biases assuming `f = G[E * X + b]`,
and are therefore more like Nengo's `intercepts`.
References
==========
.. [1] McDonnell, M. D., Tissera, M. D., Vladusich, T., Van Schaik, A.,
Tapson, J., & Schwenker, F. (2015). Fast, simple and accurate
handwritten digit classification by training shallow neural network
classifiers with the "Extreme learning machine" algorithm. PLoS ONE,
10(8), 1-20. doi:10.1371/journal.pone.0134254
"""
assert trainX.shape[0] == trainY.size
trainX = trainX.reshape(trainX.shape[0], -1)
trainY = trainY.ravel()
d = trainX.shape[1]
classes = np.unique(trainY)
assert mask is None or mask.shape == (n_encoders, d)
inds = [(trainY == label).nonzero()[0] for label in classes]
train_norm = npext.norm(trainX, axis=1).mean()
encoders = np.zeros((n_encoders, d))
biases = np.zeros(n_encoders)
for k in range(n_encoders):
for _ in range(norm_tries):
i, j = rng.choice(len(classes), size=2, replace=False)
a, b = trainX[rng.choice(inds[i])], trainX[rng.choice(inds[j])]
dab = a - b
if mask is not None:
dab *= mask[k]
ndab = npext.norm(dab)**2
if ndab >= norm_min * train_norm:
break
else:
raise ValueError("Cannot find valid encoder")
encoders[k] = (2. / ndab) * dab
biases[k] = np.dot(a + b, dab) / ndab
return encoders, biases
def test_lif_step(upsample, n_elements):
"""Test the lif nonlinearity, comparing one step with the Numpy version."""
rng = np.random
dt = 1e-3
n_neurons = [12345, 23456, 34567]
J = RA([rng.normal(scale=1.2, size=n) for n in n_neurons])
V = RA([rng.uniform(low=0, high=1, size=n) for n in n_neurons])
W = RA([rng.uniform(low=-5 * dt, high=5 * dt, size=n) for n in n_neurons])
OS = RA([np.zeros(n) for n in n_neurons])
ref = 2e-3
taus = list(rng.uniform(low=15e-3, high=80e-3, size=len(n_neurons)))
queue = cl.CommandQueue(ctx)
clJ = CLRA(queue, J)
clV = CLRA(queue, V)
clW = CLRA(queue, W)
clOS = CLRA(queue, OS)
clTau = CLRA(queue, RA(taus))
# simulate host
nls = [LIF(tau_ref=ref, tau_rc=taus[i])
for i, n in enumerate(n_neurons)]
for i, nl in enumerate(nls):
if upsample <= 1:
nl.step_math(dt, J[i], OS[i], V[i], W[i])
else:
s = np.zeros_like(OS[i])
for j in range(upsample):
nl.step_math(dt / upsample, J[i], s, V[i], W[i])
OS[i] = (1./dt) * ((OS[i] > 0) | (s > 0))
# simulate device
plan = plan_lif(queue, clJ, clV, clW, clV, clW, clOS, ref, clTau, dt,
n_elements=n_elements, upsample=upsample)
plan()
if 1:
a, b = V, clV
for i in range(len(a)):
nc, _ = not_close(a[i], b[i]).nonzero()
if len(nc) > 0:
j = nc[0]
print("i", i, "j", j)
print("J", J[i][j], clJ[i][j])
print("V", V[i][j], clV[i][j])
print("W", W[i][j], clW[i][j])
print("...", len(nc) - 1, "more")
n_spikes = np.sum([np.sum(os) for os in OS])
if n_spikes < 1.0:
logger.warn("LIF spiking mechanism was not tested!")
assert ra.allclose(J, clJ.to_host())
assert ra.allclose(V, clV.to_host())
assert ra.allclose(W, clW.to_host())
assert ra.allclose(OS, clOS.to_host())
def test_one_short_segment_many_dots(planner):
for ND in 2, 20, 100:
check_from_shapes(
planner,
0.5, 0.6, 0.7,
A_shapes=[(10, 1 + ii % 2) for ii in range(ND)],
X_shapes=[(1 + ii % 2, 1) for ii in range(ND)],
A_js=[range(ND)],
X_js=[range(ND)])
def test_one_short_segment_many_longer_dots(planner):
for ND in 2, 20, 100:
check_from_shapes(
planner,
0.5, 0.6, 0.7,
A_shapes=[(2000, ii + 1) for ii in range(ND)],
X_shapes=[(ii + 1, 1) for ii in range(ND)],
A_js=[range(ND)],
X_js=[range(ND)])
def rasterplot(time, spikes, ax=None, **kwargs):
"""Generate a raster plot of the provided spike data
Parameters
----------
time : array
Time data from the simulation
spikes: array
The spike data with columns for each neuron and 1s indicating spikes
ax: matplotlib.axes.Axes
The figure axes to plot into.
Returns
-------
ax: matplotlib.axes.Axes
The axes that were plotted into
Examples
--------
>>> import nengo
>>> model = nengo.Model("Raster")
>>> A = nengo.Ensemble(nengo.LIF(20), dimensions=1)
>>> A_spikes = nengo.Probe(A, "spikes")
>>> sim = nengo.Simulator(model)
>>> sim.run(1)
>>> rasterplot(sim.trange(), sim.data[A_spikes])
"""
if ax is None:
ax = plt.gca()
colors = kwargs.pop('colors', None)
if colors is None:
color_cycle = plt.rcParams['axes.color_cycle']
colors = [color_cycle[ix % len(color_cycle)]
for ix in range(spikes.shape[1])]
if hasattr(ax, 'eventplot'):
spikes = [time[spikes[:, i] > 0].flatten()
for i in range(spikes.shape[1])]
for ix in range(len(spikes)):
if spikes[ix].shape == (0,):
spikes[ix] = np.array([-1])
ax.eventplot(spikes, colors=colors, **kwargs)
ax.set_ylim(len(spikes) - 0.5, -0.5)
if len(spikes) == 1:
ax.set_ylim(0.4, 1.6) # eventplot plots different for len==1
ax.set_xlim(left=0, right=max(time))
else:
# Older Matplotlib, doesn't have eventplot
for i in range(spikes.shape[1]):
ax.plot(time[spikes[:, i] > 0],
np.ones_like(np.where(spikes[:, i] > 0)).T + i, ',',
color=colors[i], **kwargs)
return ax
def get_convolution_matrix(self):
"""Return the matrix that does a circular convolution by this vector.
This should be such that ``A*B == dot(A.get_convolution_matrix, B.v)``.
"""
D = len(self.v)
T = []
for i in range(D):
T.append([self.v[(i - j) % D] for j in range(D)])
return np.array(T)
def check_from_shapes(
planner,
alpha, beta, gamma,
A_shapes, X_shapes,
A_js,
X_js,
):
rng = np.random.RandomState(1234)
A = RA([0.1 + rng.rand(*shp) for shp in A_shapes])
X = RA([0.1 + rng.rand(*shp) for shp in X_shapes])
Y = RA([0.1 + rng.rand(
A_shapes[A_js[ii][0]][0],
X_shapes[X_js[ii][0]][1])
for ii in range(len(A_js))])
A_js = RA(A_js)
X_js = RA(X_js)
# -- prepare initial conditions on device
queue = cl.CommandQueue(ctx)
clA = CLRA(queue, A)
clX = CLRA(queue, X)
clY = CLRA(queue, Y)
clA_js = CLRA(queue, A_js)
clX_js = CLRA(queue, X_js)
assert allclose(A, clA)
assert allclose(X, clX)
assert allclose(Y, clY)
assert allclose(A_js, clA_js)
assert allclose(X_js, clX_js)
# -- run cl computation
prog = planner(
queue, alpha, clA, clA_js, clX, clX_js, beta, clY, gamma=gamma)
plans = prog.plans
assert len(plans) == 1
plans[0]()
# -- ensure they match
for i in range(len(A_js)):
ref = gamma + beta * Y[i] + alpha * sum(
[np.dot(A[aj], X[xj])
for aj, xj in zip(A_js[i].ravel(), X_js[i].ravel())])
sim = clY[i]
if not np.allclose(ref, sim, atol=1e-3, rtol=1e-3):
print('A_shapes', A_shapes)
print('X_shapes', X_shapes)
if len(ref) > 20:
print('ref', ref[:10], '...', ref[-10:])
print('sim', sim[:10], '...', sim[-10:])
else:
print('ref', ref)
print('sim', sim)
assert 0
def _mmul_transforms(A_shape, B_shape, C_dim):
transformA = np.zeros((C_dim, A_shape[0] * A_shape[1]))
transformB = np.zeros((C_dim, B_shape[0] * B_shape[1]))
for i in range(A_shape[0]):
for j in range(A_shape[1]):
for k in range(B_shape[1]):
tmp = (j + k * A_shape[1] + i * B_shape[0] * B_shape[1])
transformA[tmp * 2][j + i * A_shape[1]] = 1
transformB[tmp * 2 + 1][k + j * B_shape[1]] = 1
return transformA, transformB
def test_lif_speed(rng, heterogeneous):
"""Test the speed of the lif nonlinearity
heterogeneous: if true, use a wide range of population sizes.
"""
dt = 1e-3
ref = 2e-3
tau = 20e-3
n_iters = 1000
if heterogeneous:
n_neurons = [1.0e5] * 50 + [1e3] * 500
else:
n_neurons = [1.1e5] * 50
n_neurons = list(map(int, n_neurons))
J = RA([rng.randn(n) for n in n_neurons], dtype=np.float32)
V = RA([rng.uniform(low=0, high=1, size=n) for n in n_neurons],
dtype=np.float32)
W = RA([rng.uniform(low=-10 * dt, high=10 * dt, size=n)
for n in n_neurons], dtype=np.float32)
OS = RA([np.zeros(n) for n in n_neurons], dtype=np.float32)
queue = cl.CommandQueue(
ctx, properties=cl.command_queue_properties.PROFILING_ENABLE)
clJ = CLRA(queue, J)
clV = CLRA(queue, V)
clW = CLRA(queue, W)
clOS = CLRA(queue, OS)
for i, blockify in enumerate([False, True]):
plan = plan_lif(queue, dt, clJ, clV, clW, clOS, ref, tau,
blockify=blockify)
with Timer() as timer:
for j in range(n_iters):
plan()
print("plan %d: blockify = %s, dur = %0.3f"
% (i, blockify, timer.duration))
print "Original LIF impl"
for i, blockify in enumerate([False, True]):
plan = plan_lif_old(queue, dt, clJ, clV, clW, clOS, ref, tau,
blockify=blockify)
with Timer() as timer:
for j in range(n_iters):
plan()
print("plan %d: blockify = %s, dur = %0.3f"
% (i, blockify, timer.duration))
def conjgrad_scipy(A, Y, sigma, tol=1e-4):
"""Solve the least-squares system using Scipy's conjugate gradient."""
import scipy.sparse.linalg
Y, m, n, d, matrix_in = _format_system(A, Y)
damp = m * sigma**2
calcAA = lambda x: np.dot(A.T, np.dot(A, x)) + damp * x
G = scipy.sparse.linalg.LinearOperator((n, n),
matvec=calcAA,
matmat=calcAA,
dtype=A.dtype)
B = np.dot(A.T, Y)
X = np.zeros((n, d), dtype=B.dtype)
infos = np.zeros(d, dtype='int')
itns = np.zeros(d, dtype='int')
for i in range(d):
def callback(x):
itns[i] += 1 # use the callback to count the number of iterations
X[:, i], infos[i] = scipy.sparse.linalg.cg(G,
B[:, i],
tol=tol,
callback=callback)
info = {'rmses': _rmses(A, X, Y), 'iterations': itns, 'info': infos}
return X if matrix_in else X.flatten(), info
def _conjgrad_iters(calcAx, b, x, maxiters=None, rtol=1e-6):
"""Solve the single-RHS linear system using conjugate gradient."""
if maxiters is None:
maxiters = b.shape[0]
r = b - calcAx(x)
p = r.copy()
rsold = np.dot(r, r)
for i in range(maxiters):
Ap = calcAx(p)
alpha = rsold / np.dot(p, Ap)
x += alpha * p
r -= alpha * Ap
rsnew = np.dot(r, r)
beta = rsnew / rsold
if np.sqrt(rsnew) < rtol:
break
if beta < 1e-12: # no perceptible change in p
break
# p = r + beta*p
p *= beta
p += r
rsold = rsnew
return x, i + 1
def transform_in(dims, align, invert):
"""Create a transform to map the input into the Fourier domain.
See CircularConvolution docstring for more details.
Parameters
----------
dims : int
Input dimensions.
align : 'A' or 'B'
How to align the real and imaginary components; the alignment
depends on whether we're doing transformA or transformB.
invert : bool
Whether to reverse the order of elements.
"""
if align not in ('A', 'B'):
raise ValueError("'align' must be either 'A' or 'B'")
dims2 = 4 * (dims // 2 + 1)
tr = np.zeros((dims2, dims))
dft = dft_half(dims)
for i in range(dims2):
row = dft[i // 4] if not invert else dft[i // 4].conj()
if align == 'A':
tr[i] = row.real if i % 2 == 0 else row.imag
else: # align == 'B'
tr[i] = row.real if i % 4 == 0 or i % 4 == 3 else row.imag
remove_imag_rows(tr)
return tr.reshape((-1, dims))
def run_steps(self, n_steps, d=None, dt=None, rng=np.random, **kwargs):
"""Run process without input for given number of steps.
Keyword arguments that do not appear in the parameter list below
will be passed to the ``make_step`` function of this process.
Parameters
----------
n_steps : int
The number of steps to run.
d : int, optional (Default: None)
Output dimensionality. If None, ``default_size_out`` will be used.
dt : float, optional (Default: None)
Simulation timestep. If None, ``default_dt`` will be used.
rng : `numpy.random.RandomState` (Default: ``numpy.random``)
Random number generator used for stochstic processes.
"""
shape_in = as_shape(0)
shape_out = as_shape(self.default_size_out if d is None else d)
dt = self.default_dt if dt is None else dt
rng = self.get_rng(rng)
step = self.make_step(shape_in, shape_out, dt, rng, **kwargs)
output = np.zeros((n_steps, ) + shape_out)
for i in range(n_steps):
output[i] = step((i + 1) * dt)
return output
def test_sine_waves(Simulator, nl, plt):
radius = 2
dim = 5
product = nengo.networks.Product(
200, dim, radius, neuron_type=nl(), seed=63)
func_A = lambda t: np.sqrt(radius)*np.sin(np.arange(1, dim+1)*2*np.pi*t)
func_B = lambda t: np.sqrt(radius)*np.sin(np.arange(dim, 0, -1)*2*np.pi*t)
with product:
input_A = nengo.Node(func_A)
input_B = nengo.Node(func_B)
nengo.Connection(input_A, product.A)
nengo.Connection(input_B, product.B)
p = nengo.Probe(product.output, synapse=0.005)
sim = Simulator(product)
sim.run(1.0)
t = sim.trange()
AB = np.asarray(list(map(func_A, t))) * np.asarray(list(map(func_B, t)))
delay = 0.013
offset = np.where(t >= delay)[0]
for i in range(dim):
plt.subplot(dim+1, 1, i+1)
plt.plot(t + delay, AB[:, i])
plt.plot(t, sim.data[p][:, i])
assert rmse(AB[:len(offset), :], sim.data[p][offset, :]) < 0.2
def test_sine_waves(Simulator, nl):
radius = 2
dim = 5
product = nengo.networks.Product(
200, dim, radius, neuron_type=nl(), seed=63)
func_A = lambda t: radius*np.sin(np.arange(1, dim+1)*2*np.pi*t)
func_B = lambda t: radius*np.sin(np.arange(dim, 0, -1)*2*np.pi*t)
pstc = 0.003
with product:
input_A = nengo.Node(func_A)
input_B = nengo.Node(func_B)
nengo.Connection(input_A, product.A)
nengo.Connection(input_B, product.B)
p = nengo.Probe(product.output, synapse=pstc)
sim = Simulator(product)
sim.run(1.0)
t = sim.trange()
AB = np.asarray(list(map(func_A, t))) * np.asarray(list(map(func_B, t)))
delay = 0.011
offset = np.where(t > delay)[0]
with Plotter(Simulator) as plt:
for i in range(dim):
plt.subplot(dim+1, 1, i+1)
plt.plot(t + delay, AB[:, i], label="$A \cdot B$")
plt.plot(t, sim.data[p][:, i], label="Output")
plt.legend()
plt.savefig('test_product.test_sine_waves.pdf')
plt.close()
assert rmse(AB[:len(offset), :], sim.data[p][offset, :]) < 0.3
def reset(self, seed=None):
if self.closed:
raise SimulatorClosed("Cannot reset closed Simulator.")
if seed is not None:
raise NotImplementedError("Seed changing not implemented")
# reset signals
for base in self.all_bases:
# TODO: copy all data on at once
if not base.readonly:
self.all_data[self.sidx[base]] = base.initial_value
for clra, ra in iteritems(self._raggedarrays_to_reset):
# TODO: copy all data on at once
for i in range(len(clra)):
clra[i] = ra[i]
# clear probe data
if self._cl_probe_plan is not None:
self._cl_probe_plan.cl_bufpositions.fill(0)
for probe in self.model.probes:
del self._probe_outputs[probe][:]
self._reset_rng()
self._reset_cl_rngs()
self._probe_step_time()
def test_sine_waves(Simulator, plt, seed):
radius = 2
dim = 5
product = nengo.networks.Product(
200, dim, radius, net=nengo.Network(seed=seed))
func_A = lambda t: np.sqrt(radius)*np.sin(np.arange(1, dim+1)*2*np.pi*t)
func_B = lambda t: np.sqrt(radius)*np.sin(np.arange(dim, 0, -1)*2*np.pi*t)
with product:
input_A = nengo.Node(func_A)
input_B = nengo.Node(func_B)
nengo.Connection(input_A, product.A)
nengo.Connection(input_B, product.B)
p = nengo.Probe(product.output, synapse=0.005)
sim = Simulator(product)
sim.run(1.0)
t = sim.trange()
AB = np.asarray(list(map(func_A, t))) * np.asarray(list(map(func_B, t)))
delay = 0.013
offset = np.where(t >= delay)[0]
for i in range(dim):
plt.subplot(dim+1, 1, i+1)
plt.plot(t + delay, AB[:, i])
plt.plot(t, sim.data[p][:, i])
plt.xlim(right=t[-1])
assert rmse(AB[:len(offset), :], sim.data[p][offset, :]) < 0.2
def test_sine_waves(Simulator, plt, seed):
radius = 2
dim = 5
product = nengo.networks.Product(200, dim, radius, seed=seed)
func_a = lambda t: np.sqrt(radius) * np.sin(
np.arange(1, dim + 1) * 2 * np.pi * t)
func_b = lambda t: np.sqrt(radius) * np.sin(
np.arange(dim, 0, -1) * 2 * np.pi * t)
with product:
input_a = nengo.Node(func_a)
input_b = nengo.Node(func_b)
nengo.Connection(input_a, product.input_a)
nengo.Connection(input_b, product.input_b)
p = nengo.Probe(product.output, synapse=0.005)
with Simulator(product) as sim:
sim.run(1.0)
t = sim.trange()
ideal = np.asarray(list(map(func_a, t))) * np.asarray(list(map(func_b, t)))
delay = 0.013
offset = np.where(t >= delay)[0]
for i in range(dim):
plt.subplot(dim + 1, 1, i + 1)
plt.plot(t + delay, ideal[:, i])
plt.plot(t, sim.data[p][:, i])
plt.xlim(right=t[-1])
plt.yticks((-2, 0, 2))
assert rmse(ideal[:len(offset), :], sim.data[p][offset, :]) < 0.2
def run_steps(self, n_steps, d=None, dt=None, rng=np.random, **kwargs):
"""Run process without input for given number of steps.
Keyword arguments that do not appear in the parameter list below
will be passed to the ``make_step`` function of this process.
Parameters
----------
n_steps : int
The number of steps to run.
d : int, optional (Default: None)
Output dimensionality. If None, ``default_size_out`` will be used.
dt : float, optional (Default: None)
Simulation timestep. If None, ``default_dt`` will be used.
rng : `numpy.random.RandomState` (Default: ``numpy.random``)
Random number generator used for stochstic processes.
"""
shape_in = as_shape(0)
shape_out = as_shape(self.default_size_out if d is None else d)
dt = self.default_dt if dt is None else dt
rng = self.get_rng(rng)
step = self.make_step(shape_in, shape_out, dt, rng, **kwargs)
output = np.zeros((n_steps,) + shape_out)
for i in range(n_steps):
output[i] = step((i+1) * dt)
return output
def __call__(self, A, Y, rng=None, E=None):
tstart = time.time()
Y, m, n, _, matrix_in = format_system(A, Y)
# solve for coefficients using standard solver
X, info0 = self.solver1(A, Y, rng=rng)
X = self.mul_encoders(X, E)
# drop weights close to zero, based on `drop` ratio
Xabs = np.sort(np.abs(X.flat))
threshold = Xabs[int(np.round(self.drop * Xabs.size))]
X[np.abs(X) < threshold] = 0
# retrain nonzero weights
Y = self.mul_encoders(Y, E)
for i in range(X.shape[1]):
nonzero = X[:, i] != 0
if nonzero.sum() > 0:
X[nonzero, i], info1 = self.solver2(
A[:, nonzero], Y[:, i], rng=rng)
t = time.time() - tstart
info = {'rmses': rmses(A, X, Y), 'info0': info0, 'info1': info1,
'time': t}
return X if matrix_in or X.shape[1] > 1 else X.ravel(), info
def test_noise(RefSimulator, seed):
"""Make sure that we can generate noise properly."""
n = 1000
mean, std = 0.1, 0.8
noise = Signal(np.zeros(n), name="noise")
process = nengo.processes.StochasticProcess(nengo.dists.Gaussian(
mean, std))
m = Model(dt=0.001)
m.operators += [Reset(noise), SimNoise(noise, process)]
sim = RefSimulator(None, model=m, seed=seed)
samples = np.zeros((100, n))
for i in range(100):
sim.step()
samples[i] = sim.signals[noise]
h, xedges = np.histogram(samples.flat, bins=51)
x = 0.5 * (xedges[:-1] + xedges[1:])
dx = np.diff(xedges)
z = 1. / np.sqrt(2 * np.pi * std**2) * np.exp(-0.5 *
(x - mean)**2 / std**2)
y = h / float(h.sum()) / dx
assert np.allclose(y, z, atol=0.02)
def test_large(Simulator):
"""Test with a lot of big probes. Can also be used for speed."""
n = 10
def input_fn(t):
return list(range(1, 10))
model = nengo.Network(label='test_large_probes', seed=3249)
with model:
probes = []
for i in range(n):
xi = nengo.Node(label='x%d' % i, output=input_fn)
probes.append(nengo.Probe(xi, 'output'))
sim = Simulator(model)
simtime = 2.483
with Timer() as timer:
sim.run(simtime)
logger.debug("Ran %d probes for %f sec simtime in %0.3f sec",
n, simtime, timer.duration)
t = sim.dt * np.arange(int(np.round(simtime / sim.dt)))
x = np.asarray([input_fn(ti) for ti in t])
for p in probes:
y = sim.data[p]
assert np.allclose(y[1:], x[:-1]) # 1-step delay
def test_eval_points(Simulator, nl_nodirect, plt, seed, rng):
n = 100
d = 5
filter = 0.08
eval_points = np.logspace(np.log10(300), np.log10(5000), 11)
eval_points = np.round(eval_points).astype('int')
max_points = eval_points.max()
n_trials = 1
rmses = np.nan * np.zeros((len(eval_points), n_trials))
for j in range(n_trials):
points = rng.normal(size=(max_points, d))
points *= (rng.uniform(size=max_points) / norm(points, axis=-1))[:,
None]
rng_j = np.random.RandomState(348 + j)
seed = 903824 + j
# generate random input in unit hypersphere
x = rng_j.normal(size=d)
x *= rng_j.uniform() / norm(x)
for i, n_points in enumerate(eval_points):
model = nengo.Network(seed=seed)
with model:
model.config[nengo.Ensemble].neuron_type = nl_nodirect()
u = nengo.Node(output=x)
a = nengo.Ensemble(n * d,
dimensions=d,
eval_points=points[:n_points])
nengo.Connection(u, a, synapse=0)
up = nengo.Probe(u)
ap = nengo.Probe(a)
with Timer() as timer:
sim = Simulator(model)
sim.run(10 * filter)
t = sim.trange()
xt = nengo.synapses.filtfilt(sim.data[up], filter, dt=sim.dt)
yt = nengo.synapses.filtfilt(sim.data[ap], filter, dt=sim.dt)
t0 = 5 * filter
t1 = 7 * filter
tmask = (t > t0) & (t < t1)
rmses[i, j] = rms(yt[tmask] - xt[tmask])
print("done %d (%d) in %0.3f s" % (n_points, j, timer.duration))
# subtract out mean for each model
rmses_norm = rmses - rmses.mean(0, keepdims=True)
mean = rmses_norm.mean(1)
low = rmses_norm.min(1)
high = rmses_norm.max(1)
plt.semilogx(eval_points, mean, 'k-')
plt.semilogx(eval_points, high, 'r-')
plt.semilogx(eval_points, low, 'b-')
plt.xlim([eval_points[0], eval_points[-1]])
plt.xticks(eval_points, eval_points)
def test_signal_slicing(rng):
slices = [
0, 1,
slice(None, -1),
slice(1, None),
slice(1, -1),
slice(None, None, 3),
slice(1, -1, 2)
]
x = np.arange(12, dtype=float)
y = np.arange(24, dtype=float).reshape(4, 6)
a = Signal(x.copy())
b = Signal(y.copy())
for i in range(100):
si0, si1 = rng.randint(0, len(slices), size=2)
s0, s1 = slices[si0], slices[si1]
assert np.array_equiv(a[s0].initial_value, x[s0])
assert np.array_equiv(b[s0, s1].initial_value, y[s0, s1])
with pytest.raises(ValueError):
a[[0, 2]]
with pytest.raises(ValueError):
b[[0, 1], [3, 4]]
def test_init():
a = SemanticPointer([1, 2, 3, 4])
assert len(a) == 4
a = SemanticPointer([1, 2, 3, 4, 5])
assert len(a) == 5
a = SemanticPointer(list(range(100)))
assert len(a) == 100
a = SemanticPointer(27)
assert len(a) == 27
assert np.allclose(a.length(), 1)
with pytest.raises(Exception):
a = SemanticPointer(np.zeros(2, 2))
with pytest.raises(Exception):
a = SemanticPointer(-1)
with pytest.raises(Exception):
a = SemanticPointer(0)
with pytest.raises(Exception):
a = SemanticPointer(1.7)
with pytest.raises(Exception):
a = SemanticPointer(None)
with pytest.raises(Exception):
a = SemanticPointer(int)
def _conjgrad_iters(calcAx, b, x, maxiters=None, rtol=1e-6):
"""Solve the single-RHS linear system using conjugate gradient."""
if maxiters is None:
maxiters = b.shape[0]
r = b - calcAx(x)
p = r.copy()
rsold = np.dot(r, r)
for i in range(maxiters):
Ap = calcAx(p)
alpha = rsold / np.dot(p, Ap)
x += alpha * p
r -= alpha * Ap
rsnew = np.dot(r, r)
beta = rsnew / rsold
if np.sqrt(rsnew) < rtol:
break
if beta < 1e-12: # no perceptible change in p
break
# p = r + beta*p
p *= beta
p += r
rsold = rsnew
return x, i+1
def test_direct_mode_with_single_neuron(Simulator, plt, seed):
radius = 2
dim = 5
config = nengo.Config(nengo.Ensemble)
config[nengo.Ensemble].neuron_type = nengo.Direct()
with config:
product = nengo.networks.Product(
1, dim, radius, net=nengo.Network(seed=seed))
func_A = lambda t: np.sqrt(radius)*np.sin(np.arange(1, dim+1)*2*np.pi*t)
func_B = lambda t: np.sqrt(radius)*np.sin(np.arange(dim, 0, -1)*2*np.pi*t)
with product:
input_A = nengo.Node(func_A)
input_B = nengo.Node(func_B)
nengo.Connection(input_A, product.A)
nengo.Connection(input_B, product.B)
p = nengo.Probe(product.output, synapse=0.005)
sim = Simulator(product)
sim.run(1.0)
t = sim.trange()
AB = np.asarray(list(map(func_A, t))) * np.asarray(list(map(func_B, t)))
delay = 0.013
offset = np.where(t >= delay)[0]
for i in range(dim):
plt.subplot(dim+1, 1, i+1)
plt.plot(t + delay, AB[:, i])
plt.plot(t, sim.data[p][:, i])
plt.xlim(right=t[-1])
assert rmse(AB[:len(offset), :], sim.data[p][offset, :]) < 0.2
def conjgrad_scipy(A, Y, sigma, tol=1e-4):
"""Solve the least-squares system using Scipy's conjugate gradient."""
import scipy.sparse.linalg
Y, m, n, d, matrix_in = _format_system(A, Y)
damp = m * sigma**2
calcAA = lambda x: np.dot(A.T, np.dot(A, x)) + damp * x
G = scipy.sparse.linalg.LinearOperator(
(n, n), matvec=calcAA, matmat=calcAA, dtype=A.dtype)
B = np.dot(A.T, Y)
X = np.zeros((n, d), dtype=B.dtype)
infos = np.zeros(d, dtype='int')
itns = np.zeros(d, dtype='int')
for i in range(d):
def callback(x):
itns[i] += 1 # use the callback to count the number of iterations
X[:, i], infos[i] = scipy.sparse.linalg.cg(
G, B[:, i], tol=tol, callback=callback)
info = {'rmses': npext.rms(Y - np.dot(A, X), axis=0),
'iterations': itns,
'info': infos}
return X if matrix_in else X.flatten(), info
def __call__(self, A, Y, rng=None, E=None):
Y, m, n, d, matrix_in = _format_system(A, Y)
# solve for coefficients using standard solver
X, info0 = self.solver1(A, Y, rng=rng)
X = self.mul_encoders(X, E)
# drop weights close to zero, based on `drop` ratio
Xabs = np.sort(np.abs(X.flat))
threshold = Xabs[int(np.round(self.drop * Xabs.size))]
X[np.abs(X) < threshold] = 0
# retrain nonzero weights
Y = self.mul_encoders(Y, E)
for i in range(X.shape[1]):
nonzero = X[:, i] != 0
if nonzero.sum() > 0:
X[nonzero, i], info1 = self.solver2(A[:, nonzero],
Y[:, i],
rng=rng)
info = {
'rmses': npext.rms(Y - np.dot(A, X), axis=0),
'info0': info0,
'info1': info1
}
return X if matrix_in else X.flatten(), info
def __call__(self, A, Y, rng=None, E=None):
tstart = time.time()
Y, m, n, d, matrix_in = format_system(A, Y)
# solve for coefficients using standard solver
X, info0 = self.solver1(A, Y, rng=rng)
X = self.mul_encoders(X, E)
# drop weights close to zero, based on `drop` ratio
Xabs = np.sort(np.abs(X.flat))
threshold = Xabs[int(np.round(self.drop * Xabs.size))]
X[np.abs(X) < threshold] = 0
# retrain nonzero weights
Y = self.mul_encoders(Y, E)
for i in range(X.shape[1]):
nonzero = X[:, i] != 0
if nonzero.sum() > 0:
X[nonzero, i], info1 = self.solver2(
A[:, nonzero], Y[:, i], rng=rng)
t = time.time() - tstart
info = {'rmses': rmses(A, X, Y), 'info0': info0, 'info1': info1,
'time': t}
return X if matrix_in else X.flatten(), info
def zpk2tf(z, p, k):
"""Return polynomial transfer function representation from zeros
and poles
Parameters
----------
z : ndarray
Zeros of the transfer function.
p : ndarray
Poles of the transfer function.
k : float
System gain.
Returns
-------
b : ndarray
Numerator polynomial.
a : ndarray
Denominator polynomial.
"""
z = atleast_1d(z)
k = atleast_1d(k)
if len(z.shape) > 1:
temp = poly(z[0])
b = zeros((z.shape[0], z.shape[1] + 1), temp.dtype.char)
if len(k) == 1:
k = [k[0]] * z.shape[0]
for i in range(z.shape[0]):
b[i] = k[i] * poly(z[i])
else:
b = k * poly(z)
a = atleast_1d(poly(p))
return b, a
def transform_in(dims, align, invert):
"""Create a transform to map the input into the Fourier domain.
See CircularConvolution docstring for more details.
Parameters
----------
dims : int
Input dimensions.
align : 'A' or 'B'
How to align the real and imaginary components; the alignment
depends on whether we're doing transformA or transformB.
invert : bool
Whether to reverse the order of elements.
"""
if align not in ('A', 'B'):
raise ValidationError("'align' must be either 'A' or 'B'", 'align')
dims2 = 4 * (dims // 2 + 1)
tr = np.zeros((dims2, dims))
dft = dft_half(dims)
for i in range(dims2):
row = dft[i // 4] if not invert else dft[i // 4].conj()
if align == 'A':
tr[i] = row.real if i % 2 == 0 else row.imag
else: # align == 'B'
tr[i] = row.real if i % 4 == 0 or i % 4 == 3 else row.imag
remove_imag_rows(tr)
return tr.reshape((-1, dims))
def test_groupby(hashable, force_list, rng):
if hashable:
keys = list(range(1, 5))
else:
keys = [[0, 0], [0, 1], [1, 0], [1, 1]]
keys = sorted(keys)
# make groups and pairs
groups = [rng.randn(rng.randint(5, 10)) for _ in keys]
pairs = []
for key, group in zip(keys, groups):
pairs.extend((key, value) for value in group)
# shuffle pairs
pairs = [pairs[i] for i in rng.permutation(len(pairs))]
# call groupby
keygroups = groupby(pairs, lambda p: p[0], force_list=force_list)
keys2 = sorted(map(lambda x: x[0], keygroups))
assert keys2 == keys
for key2, keygroup2 in keygroups:
group = groups[keys.index(key2)]
group2 = map(lambda x: x[1], keygroup2)
assert sorted(group2) == sorted(group)
def test_groupby(hashable, force_list, rng):
if hashable:
keys = list(range(1, 5))
else:
keys = [[0, 0], [0, 1], [1, 0], [1, 1]]
keys = sorted(keys)
# make groups and pairs
groups = [rng.randn(rng.randint(5, 10)) for _ in keys]
pairs = []
for key, group in zip(keys, groups):
pairs.extend((key, value) for value in group)
# shuffle pairs
pairs = [pairs[i] for i in rng.permutation(len(pairs))]
# call groupby
keygroups = groupby(pairs, lambda p: p[0], force_list=force_list)
keys2 = sorted(map(lambda x: x[0], keygroups))
assert keys2 == keys
for key2, keygroup2 in keygroups:
group = groups[keys.index(key2)]
group2 = map(lambda x: x[1], keygroup2)
assert sorted(group2) == sorted(group)
def test_init():
a = SemanticPointer([1, 2, 3, 4])
assert len(a) == 4
a = SemanticPointer([1, 2, 3, 4, 5])
assert len(a) == 5
a = SemanticPointer(list(range(100)))
assert len(a) == 100
a = SemanticPointer(27)
assert len(a) == 27
assert np.allclose(a.length(), 1)
with pytest.raises(Exception):
a = SemanticPointer(np.zeros(2, 2))
with pytest.raises(Exception):
a = SemanticPointer(-1)
with pytest.raises(Exception):
a = SemanticPointer(0)
with pytest.raises(Exception):
a = SemanticPointer(1.7)
with pytest.raises(Exception):
a = SemanticPointer(None)
with pytest.raises(Exception):
a = SemanticPointer(int)
def reset(self, seed=None):
if self.closed:
raise SimulatorClosed("Cannot reset closed Simulator.")
if seed is not None:
raise NotImplementedError("Seed changing not implemented")
# reset signals
for base in self.all_bases:
# TODO: copy all data on at once
if not base.readonly:
self.all_data[self.sidx[base]] = base.initial_value
for clra, ra in iteritems(self._raggedarrays_to_reset):
# TODO: copy all data on at once
for i in range(len(clra)):
clra[i] = ra[i]
# clear probe data
if self._cl_probe_plan is not None:
self._cl_probe_plan.cl_bufpositions.fill(0)
for probe in self.model.probes:
del self._probe_outputs[probe][:]
self._reset_rng()
self._reset_cl_rngs()
self._probe_step_time()
def test_basic():
# -- prepare initial conditions on host
A = RA([[[0.1, .2], [.3, .4]], [[.5, .6]]])
X = RA([[3, 5]])
Y = RA([[0.0], [2, 3], ])
A_js = RA([[1], [0]], dtype=np.int32)
X_js = RA([[0], [0]], dtype=np.int32)
# alpha = 0.5
alpha = 1.0
# beta = 0.1
beta = 1.0
# -- prepare initial conditions on device
queue = cl.CommandQueue(ctx)
clA = CLRA(queue, A)
clX = CLRA(queue, X)
clY = CLRA(queue, Y)
assert allclose(A, clA)
assert allclose(X, clX)
assert allclose(Y, clY)
# -- run cl computation
prog = plan_ragged_gather_gemv(
queue, alpha, clA, A_js, clX, X_js, beta, clY)
# plans = prog.choose_plans()
# assert len(plans) == 1
for plan in prog.plans:
plan()
# -- ensure they match
for i in range(len(A_js)):
aj, xj = int(A_js[i]), int(X_js[i])
ref = alpha * np.dot(A[aj], X[xj]) + beta * Y[i]
sim = clY[i]
assert np.allclose(ref, sim)
def test_large(Simulator, seed, logger):
"""Test with a lot of big probes. Can also be used for speed."""
n = 10
def input_fn(t):
return list(range(1, 10))
model = nengo.Network(label="test_large_probes", seed=seed)
with model:
probes = []
for i in range(n):
xi = nengo.Node(label="x%d" % i, output=input_fn)
probes.append(nengo.Probe(xi, "output"))
sim = Simulator(model)
simtime = 2.483
with Timer() as timer:
sim.run(simtime)
logger.info("Ran %d probes for %f sec simtime in %0.3f sec", n, simtime, timer.duration)
t = sim.trange()
x = np.asarray([input_fn(ti) for ti in t])
for p in probes:
y = sim.data[p]
assert np.allclose(y[1:], x[:-1]) # 1-step delay
def test_linearfilter(ctx, n_per_kind, rng):
kinds = (
nengo.synapses.LinearFilter((2.,), (1.,), analog=False),
nengo.synapses.Lowpass(0.005),
nengo.synapses.Alpha(0.005),
)
assert len(n_per_kind) == len(kinds)
kinds_n = [(kind, n) for kind, n in zip(kinds, n_per_kind) if n > 0]
dt = 0.001
steps = [kind.make_step((n,), (n,), dt, None, dtype=np.float32)
for kind, n in kinds_n]
A = RA([step.den for step in steps])
B = RA([step.num for step in steps])
X = RA([rng.normal(size=n) for kind, n in kinds_n])
Y = RA([np.zeros(n) for kind, n in kinds_n])
Xbuf = RA([np.zeros(shape) for shape in zip(B.sizes, X.sizes)])
Ybuf = RA([np.zeros(shape) for shape in zip(A.sizes, Y.sizes)])
queue = cl.CommandQueue(ctx)
clA = CLRA(queue, A)
clB = CLRA(queue, B)
clX = CLRA(queue, X)
clY = CLRA(queue, Y)
clXbuf = CLRA(queue, Xbuf)
clYbuf = CLRA(queue, Ybuf)
n_calls = 3
plans = plan_linearfilter(queue, clX, clY, clA, clB, clXbuf, clYbuf)
with Timer() as timer:
for _ in range(n_calls):
[plan() for plan in plans]
print(timer.duration)
for i, [kind, n] in enumerate(kinds_n):
n = min(n, 100)
step = kind.make_step((n, 1), (n, 1), dt, None, dtype=np.float32)
x = X[i][:n]
y = np.zeros_like(x)
for _ in range(n_calls):
y[:] = step(0, x)
z = clY[i][:n]
assert np.allclose(z, y, atol=1e-7, rtol=1e-5), kind
def test_linearfilter(n_per_kind, rng):
kinds = (
nengo.synapses.LinearFilter((2.,), (1.,), analog=False),
nengo.synapses.Lowpass(0.005),
nengo.synapses.Alpha(0.005),
)
assert len(n_per_kind) == len(kinds)
kinds_n = [(kind, n) for kind, n in zip(kinds, n_per_kind) if n > 0]
dt = 0.001
steps = [kind.make_step((n,), (n,), dt, None, dtype=np.float32)
for kind, n in kinds_n]
A = RA([step.den for step in steps])
B = RA([step.num for step in steps])
X = RA([rng.normal(size=n) for kind, n in kinds_n])
Y = RA([np.zeros(n) for kind, n in kinds_n])
Xbuf = RA([np.zeros(shape) for shape in zip(B.sizes, X.sizes)])
Ybuf = RA([np.zeros(shape) for shape in zip(A.sizes, Y.sizes)])
queue = cl.CommandQueue(ctx)
clA = CLRA(queue, A)
clB = CLRA(queue, B)
clX = CLRA(queue, X)
clY = CLRA(queue, Y)
clXbuf = CLRA(queue, Xbuf)
clYbuf = CLRA(queue, Ybuf)
n_calls = 3
plan = plan_linearfilter(queue, clX, clY, clA, clB, clXbuf, clYbuf)
with Timer() as timer:
for _ in range(n_calls):
plan()
print(timer.duration)
for i, [kind, n] in enumerate(kinds_n):
n = min(n, 100)
step = kind.make_step((n, 1), (n, 1), dt, None, dtype=np.float32)
x = X[i][:n]
y = np.zeros_like(x)
for _ in range(n_calls):
y[:] = step(0, x)
z = clY[i][:n]
assert np.allclose(z, y, atol=1e-7, rtol=1e-5), kind
def plot(sim, a, p, title=""):
a_ref = np.tile(a, (len(t), 1))
a_sim = sim.data[p]
colors = ['b', 'g', 'r', 'c', 'm', 'y']
for i in range(a_sim.shape[1]):
plt.plot(t, a_ref[:, i], '--', color=colors[i % 6])
plt.plot(t, a_sim[:, i], '-', color=colors[i % 6])
plt.title(title)
def cl_timing(recorded, target):
recalign, tgtalign, rectimes, tgttimes = cl2string(recorded, target)
correct_ix = [i for i in range(len(recalign))
if recalign[i] == tgtalign[i]]
t_diff = [rectimes[i] - tgttimes[i] for i in correct_ix]
return np.mean(t_diff), np.var(t_diff)
def plot(actual, probe, title=""):
ref_y = np.tile(actual, (len(t), 1))
sim_y = sim.data[probe]
colors = ['b', 'g', 'r', 'c', 'm', 'y']
for i in range(min(dims, len(colors))):
plt.plot(t, ref_y[:, i], '--', color=colors[i])
plt.plot(t, sim_y[:, i], '-', color=colors[i])
plt.title(title)
def plot(sim, a, p, title=""):
a_ref = np.tile(a, (len(t), 1))
a_sim = sim.data[p]
colors = ['b', 'g', 'r', 'c', 'm', 'y']
for i in range(a_sim.shape[1]):
plt.plot(t, a_ref[:, i], '--', color=colors[i % 6])
plt.plot(t, a_sim[:, i], '-', color=colors[i % 6])
plt.title(title)
def __init__(self,
n_neurons,
n_ensembles,
ens_dimensions=1,
neuron_nodes=False,
label=None,
seed=None,
add_to_container=None,
**ens_kwargs):
if "dimensions" in ens_kwargs:
raise ValidationError(
"'dimensions' is not a valid argument to EnsembleArray. "
"To set the number of ensembles, use 'n_ensembles'. To set "
"the number of dimensions per ensemble, use 'ens_dimensions'.",
attr='dimensions',
obj=self)
super(EnsembleArray, self).__init__(label, seed, add_to_container)
for param in ens_kwargs:
if is_iterable(ens_kwargs[param]):
ens_kwargs[param] = nengo.dists.Samples(ens_kwargs[param])
self.config[nengo.Ensemble].update(ens_kwargs)
label_prefix = "" if label is None else label + "_"
self.n_neurons = n_neurons
self.n_ensembles = n_ensembles
self.dimensions_per_ensemble = ens_dimensions
# These may be set in add_neuron_input and add_neuron_output
self.neuron_input, self.neuron_output = None, None
self.ea_ensembles = []
with self:
self.input = nengo.Node(size_in=self.dimensions, label="input")
for i in range(n_ensembles):
e = nengo.Ensemble(n_neurons,
self.dimensions_per_ensemble,
label="%s%d" % (label_prefix, i))
nengo.Connection(self.input[i * ens_dimensions:(i + 1) *
ens_dimensions],
e,
synapse=None)
self.ea_ensembles.append(e)
if neuron_nodes:
self.add_neuron_input()
self.add_neuron_output()
warnings.warn(
"'neuron_nodes' argument will be removed in Nengo 2.2. Use "
"'add_neuron_input' and 'add_neuron_output' methods instead.",
DeprecationWarning)
self.add_output('output', function=None)
def gain_bias(self, max_rates, intercepts):
"""Compute the gain and bias needed to satisfy max_rates, intercepts.
This takes the neurons, approximates their response function, and then
uses that approximation to find the gain and bias value that will give
the requested intercepts and max_rates.
Note that this default implementation is very slow! Whenever possible,
subclasses should override this with a neuron-specific implementation.
Parameters
----------
max_rates : ndarray(dtype=float64)
Maximum firing rates of neurons.
intercepts : ndarray(dtype=float64)
X-intercepts of neurons.
Returns
-------
gain : ndarray(dtype=float64)
Gain associated with each neuron. Sometimes denoted alpha.
bias : ndarray(dtype=float64)
Bias current associated with each neuron.
"""
J_max = 0
J_steps = 101 # Odd number so that 0 is a sample
max_rate = max_rates.max()
# Start with dummy gain and bias so x == J in rate calculation
gain = np.ones(J_steps)
bias = np.zeros(J_steps)
rate = np.zeros(J_steps)
# Find range of J that will achieve max rates
while rate[-1] < max_rate and J_max < 100:
J_max += 10
J = np.linspace(-J_max, J_max, J_steps)
rate = self.rates(J, gain, bias)
J_threshold = J[np.where(rate <= 1e-16)[0][-1]]
gain = np.zeros_like(max_rates)
bias = np.zeros_like(max_rates)
for i in range(intercepts.size):
ix = np.where(rate > max_rates[i])[0]
if len(ix) == 0:
ix = -1
else:
ix = ix[0]
if rate[ix] == rate[ix - 1]:
p = 1
else:
p = (max_rates[i] - rate[ix - 1]) / (rate[ix] - rate[ix - 1])
J_top = p * J[ix] + (1 - p) * J[ix - 1]
gain[i] = (J_threshold - J_top) / (intercepts[i] - 1)
bias[i] = J_top - gain[i]
return gain, bias
def _test_conn(OclOnlySimulator, fn, size_in, dist_in=None, n=1):
seed = sum(map(ord, fn.__name__)) % 2**30
rng = np.random.RandomState(seed + 1)
# make input
if dist_in is None:
dist_in = [Uniform(-10, 10) for i in range(size_in)]
elif not isinstance(dist_in, (list, tuple)):
dist_in = [dist_in]
assert len(dist_in) == size_in
x = list(zip(*[d.sample(n, rng=rng) for d in dist_in])) # preserve types
# x = zip(*[arggen.gen(n, rng=rng) for arggen in arggens])
y = [fn(xx) for xx in x]
y = np.array([fn(xx) for xx in x])
if y.ndim < 2:
y.shape = (n, 1)
size_out = y.shape[1]
# make model
model = nengo.Network("test_%s" % fn.__name__, seed=seed)
with model:
probes = []
for i in range(n):
u = nengo.Node(output=x[i])
v = nengo.Ensemble(1,
dimensions=size_in,
neuron_type=nengo.Direct())
w = nengo.Ensemble(1,
dimensions=size_out,
neuron_type=nengo.Direct())
nengo.Connection(u, v, synapse=None)
nengo.Connection(v, w, synapse=None, function=fn, eval_points=x)
probes.append(nengo.Probe(w))
# run model
with OclOnlySimulator(model) as sim:
sim.step()
# sim.step()
# sim.step()
# compare output
z = np.array([sim.data[p][-1] for p in probes])
assert np.allclose(z, y, atol=2e-7)
def cl_timing(recorded, target):
recalign, tgtalign, rectimes, tgttimes = cl2string(recorded, target)
correct_ix = [
i for i in range(len(recalign)) if recalign[i] == tgtalign[i]
]
t_diff = [rectimes[i] - tgttimes[i] for i in correct_ix]
return np.mean(t_diff), np.var(t_diff)
def run_steps(self, n_steps, d=None, dt=None, rng=np.random):
d = self.default_size_out if d is None else d
dt = self.default_dt if dt is None else dt
rng = self.get_rng(rng)
step = self.make_step(0, d, dt, rng)
output = np.zeros((n_steps, d))
for i in range(n_steps):
output[i] = step(i * dt)
return output
def __init__(self,
n_neurons,
n_ensembles,
ens_dimensions=1,
neuron_nodes=False,
label=None,
**ens_kwargs):
if "dimensions" in ens_kwargs:
raise TypeError(
"'dimensions' is not a valid argument to EnsembleArray. "
"To set the number of ensembles, use 'n_ensembles'. To set "
"the number of dimensions per ensemble, use 'ens_dimensions'.")
self.config[nengo.Ensemble].update(ens_kwargs)
label_prefix = "" if label is None else label + "_"
self.n_neurons = n_neurons
self.n_ensembles = n_ensembles
self.dimensions_per_ensemble = ens_dimensions
self.input = nengo.Node(size_in=self.dimensions, label="input")
if neuron_nodes:
self.neuron_input = nengo.Node(size_in=n_neurons * n_ensembles,
label="neuron_input")
self.neuron_output = nengo.Node(size_in=n_neurons * n_ensembles,
label="neuron_output")
self.ea_ensembles = []
for i in range(n_ensembles):
e = nengo.Ensemble(n_neurons,
self.dimensions_per_ensemble,
label=label_prefix + str(i))
nengo.Connection(self.input[i * ens_dimensions:(i + 1) *
ens_dimensions],
e,
synapse=None)
if neuron_nodes and not isinstance(e.neuron_type, nengo.Direct):
nengo.Connection(self.neuron_input[i * n_neurons:(i + 1) *
n_neurons],
e.neurons,
synapse=None)
nengo.Connection(e.neurons,
self.neuron_output[i * n_neurons:(i + 1) *
n_neurons],
synapse=None)
self.ea_ensembles.append(e)
if neuron_nodes and isinstance(e.neuron_type, nengo.Direct):
warnings.warn("Creating neuron nodes in an EnsembleArray"
" with Direct neurons")
self.add_output('output', function=None)
def run_steps(self, n_steps, d=None, dt=None, rng=np.random):
# TODO: allow running with input
d = self.default_size_out if d is None else d
dt = self.default_dt if dt is None else dt
step = self.make_step(0, d, dt, rng)
output = np.zeros((n_steps, d))
for i in range(n_steps):
output[i] = step(i * dt)
return output
def plot(sim, expected, probe, title=""):
simdata = sim.data[probe]
colors = ['b', 'g', 'r', 'c']
for i in range(simdata.shape[1]):
plt.axhline(expected[i], ls='--', color=colors[i % 4])
plt.plot(t, simdata[:, i], color=colors[i % 4])
plt.xticks(np.linspace(0, 0.4, 5))
plt.xlim(right=t[-1])
plt.title(title)
def whitenoise(step, high, rms=0.5, seed=None, dimensions=None):
"""Generate white noise inputs
Parameters
----------
step : float
The step size of different frequencies to generate
high : float
The highest frequency to generate (should be a multiple of step)
rms : float
The RMS power of the signal
seed : int or None
Random number seed
dimensions : int or None
The number of different random signals to generate. The resulting
function will return an array of length `dimensions` for every
point in time. If `dimensions` is None, the resulting function will
just return a float for each point in time.
Returns
-------
function:
A function that takes a variable t and returns the value of the
randomly generated signal. This value is a float if `dimensions` is
None; otherwise it is a list of length `dimensions`.
"""
rng = np.random.RandomState(seed)
if dimensions is not None:
signals = [
whitenoise(step, high, rms=rms, seed=rng.randint(0x7ffffff))
for i in range(dimensions)
]
def whitenoise_function(t, signals=signals):
return [signal(t) for signal in signals]
return whitenoise_function
N = int(float(high) / step) # number of samples
frequencies = np.arange(1, N + 1) * step * 2 * np.pi # frequency of each
amplitude = rng.uniform(0, 1, N) # amplitude for each sample
phase = rng.uniform(0, 2 * np.pi, N) # phase of each sample
# compute the rms of the signal
rawRMS = np.sqrt(np.sum(amplitude**2) / 2)
amplitude = amplitude * rms / rawRMS # rescale
# create a function that computes the bases and weights them by amplitude
def whitenoise_function(t, f=frequencies, a=amplitude, p=phase):
return np.dot(np.sin(f * t[..., np.newaxis] + p), a)
return whitenoise_function
def get_filterbanks(n_filters=20, n_fft=512, fs=16000,
minfreq=0, maxfreq=None):
"""Compute a Mel-filterbank.
The filters are stored in the rows, the columns correspond to fft bins.
The filters are returned as an array of shape (n_filters, n_fft/2 + 1).
Parameters
----------
n_filters : int, optional
The number of filters in the filterbank. Default: 20
n_fft : int, optional
The FFT size. Default: 512
fs : float, optional
The samplerate of the signal we are working with. Affects mel spacing.
Default: 16000
minfreq : int, optional
Lowest band edge of mel filters, in Hz. Default: 0
maxfreq : int, optional
highest band edge of mel filters, in Hz. Default: fs / 2
Returns
-------
Numpy array of shape (n_filters, n_fft/2 + 1) containing the filterbank.
Each row holds 1 filter.
"""
maxfreq = fs // 2 if maxfreq is None else maxfreq
assert maxfreq <= fs // 2, "maxfreq is greater than fs/2"
# compute points evenly spaced in mels
lowmel = hz2mel(minfreq)
highmel = hz2mel(maxfreq)
melpoints = np.linspace(lowmel, highmel, n_filters + 2)
# Our points are in Hz, but we use fft bins, so we have to convert
# from Hz to fft bin number
fbin = np.floor((n_fft + 1) * mel2hz(melpoints) / fs)
fbank = np.zeros([n_filters, int(n_fft/2+1)])
for j in range(n_filters):
for i in range(int(fbin[j]), int(fbin[j + 1])):
fbank[j, i] = (i - fbin[j]) / (fbin[j + 1] - fbin[j])
for i in range(int(fbin[j + 1]), int(fbin[j + 2])):
fbank[j, i] = (fbin[j + 2] - i) / (fbin[j + 2] - fbin[j + 1])
return fbank
def ss2tf(A, B, C, D, input=0):
"""State-space to transfer function.
Parameters
----------
A, B, C, D : ndarray
State-space representation of linear system.
input : int, optional
For multiple-input systems, the input to use.
Returns
-------
num : 2-D ndarray
Numerator(s) of the resulting transfer function(s). `num` has one row
for each of the system's outputs. Each row is a sequence representation
of the numerator polynomial.
den : 1-D ndarray
Denominator of the resulting transfer function(s). `den` is a sequence
representation of the denominator polynomial.
"""
# transfer function is C (sI - A)**(-1) B + D
A, B, C, D = map(asarray, (A, B, C, D))
# Check consistency and make them all rank-2 arrays
A, B, C, D = abcd_normalize(A, B, C, D)
nout, nin = D.shape
if input >= nin:
raise ValueError("System does not have the input specified.")
# make MOSI from possibly MOMI system.
if B.shape[-1] != 0:
B = B[:, input]
B.shape = (B.shape[0], 1)
if D.shape[-1] != 0:
D = D[:, input]
try:
den = poly(A)
except ValueError:
den = 1
if (product(B.shape, axis=0) == 0) and (product(C.shape, axis=0) == 0):
num = np.ravel(D)
if (product(D.shape, axis=0) == 0) and (product(A.shape, axis=0) == 0):
den = []
return num, den
num_states = A.shape[0]
type_test = A[:, 0] + B[:, 0] + C[0, :] + D
num = np.zeros((nout, num_states + 1), type_test.dtype)
for k in range(nout):
Ck = atleast_2d(C[k, :])
num[k] = poly(A - dot(B, Ck)) + (D[k] - 1) * den
return num, den
def plot(sim, a, p, title=""):
a_ref = np.tile(a, (len(t), 1))
a_sim = sim.data[p]
colors = ['b', 'g', 'r', 'c', 'm', 'y']
for i in range(a_sim.shape[1]):
plt.plot(t, a_ref[:, i], '--', color=colors[i % 6])
plt.plot(t, a_sim[:, i], '-', color=colors[i % 6])
plt.xticks(np.linspace(0, 0.4, 5))
plt.xlim(right=t[-1])
plt.title(title)
